{
 "source": "Appendix A, Controlled Glossary",
 "document": "A1",
 "version": "1.1",
 "entries": [
  {
   "term": "A/B test (online controlled experiment)",
   "definition": "A randomized experiment in which units (users, requests, cities) are assigned at random to a control variant and one or more treatment variants, so that differences in outcomes can be attributed to the variants.",
   "chapters": [
    64
   ]
  },
  {
   "term": "Abduction, action, prediction",
   "definition": "The three steps by which a counterfactual is computed for an observed individual: infer the background variables from the evidence, apply the intervention, and recompute the downstream variables (Pearl, 2009).",
   "chapters": [
    18
   ]
  },
  {
   "term": "Ability",
   "definition": "The group of skills, competencies and characteristics that enable a party to have influence within some specific domain; it is domain specific (Mayer et al., 1995).",
   "chapters": [
    1
   ]
  },
  {
   "term": "Abstraction layer (adapter layer)",
   "definition": "Internal code that offers one stable interface to the organization's developers and translates calls into whichever external library performs the computation.",
   "chapters": [
    58
   ]
  },
  {
   "term": "Abstraction trap",
   "definition": "A failure mode that results from not accounting for the interactions between a technical system and the social world in which it operates (Selbst et al., 2019). Chapter 3 states it for a fairness intervention defined on an abstracted technical system, which loses validity once the system is placed in its social context.",
   "chapters": [
    3,
    45
   ]
  },
  {
   "term": "Acceptance criterion",
   "definition": "A verifiable condition that shows a requirement has been met, for example \"selection-rate ratios and intervals are computed for every protected group and every intersection with at least 50 members, and reported\". In agile development (Chapter 103), acceptance criteria are the conditions a story must satisfy to be accepted; they are checked for each story, while the Definition of Done (Chapter 102) applies to every Increment.",
   "chapters": [
    85,
    103
   ]
  },
  {
   "term": "Accessory (subordinate) guarantee",
   "definition": "A non-discrimination clause that applies only to the enjoyment of the other rights in the same instrument, such as Article 2(1) of the International Covenant on Civil and Political Rights (ICCPR) and Article 14 of the European Convention on Human Rights (ECHR).",
   "chapters": [
    84
   ]
  },
  {
   "term": "Accountability",
   "definition": "A relationship between an actor and a forum, in which the actor has an obligation to explain and to justify his or her conduct, the forum can pose questions and pass judgment, and the actor may face consequences (Bovens, 2007). In fairness governance (Chapter 106), accountability is the obligation to answer for a decision and its outcome, including the authority to make it, held by one identifiable role per decision.",
   "chapters": [
    10,
    106
   ]
  },
  {
   "term": "Accreditation",
   "definition": "A statement by an authority that an entity is competent and entitled to perform a specified activity, such as issuing a given type of credential. In EBSI, each accreditation is itself a verifiable credential stored in a registry (EBSI, 2026a).",
   "chapters": [
    100
   ]
  },
  {
   "term": "Actionable, immutable and conditionally immutable features",
   "definition": "A feature is actionable if the person can change it, immutable if no action can change it (such as birthplace), and conditionally immutable if it can change only in one direction or under some conditions, such as a degree that can be obtained but not lost (Ustun et al., 2019).",
   "chapters": [
    74
   ]
  },
  {
   "term": "Actual cause",
   "definition": "A particular event that caused a particular outcome in a specific situation, as opposed to a general causal tendency. Halpern and Pearl (2005a) define it using structural equations and counterfactuals.",
   "chapters": [
    75
   ]
  },
  {
   "term": "Adaptive enrichment",
   "definition": "Changing allocation during the experiment to sample more heavily from subgroups where uncertainty or early signs of disparity are greatest.",
   "chapters": [
    64
   ]
  },
  {
   "term": "Adaptive system",
   "definition": "A decision system whose model, data or environment changes after deployment in response to its own operation, through retraining, online learning or the reactions of the people it affects. The AI Act's definition of an AI system (European Parliament and Council of the European Union, 2024) notes that such systems \"may exhibit adaptiveness after deployment\" (Regulation (EU) 2024/1689, Art. 3(1)).",
   "chapters": [
    114
   ]
  },
  {
   "term": "Adaptive threshold",
   "definition": "An alert threshold that changes with context: group size, historical volatility, time of day or week, or feedback on past alerts.",
   "chapters": [
    62
   ]
  },
  {
   "term": "Advanced AI assistant",
   "definition": "An artificial agent with a natural language interface whose function is to plan and execute sequences of actions on behalf of a user, across one or more domains, in line with the user's expectations (Gabriel et al., 2024).",
   "chapters": [
    113
   ]
  },
  {
   "term": "Adversarial context",
   "definition": "A situation in which the provider and the receiver of an explanation have opposing interests and incentives, so that the provider might manipulate the explanation (Bordt et al., 2022).",
   "chapters": [
    82
   ]
  },
  {
   "term": "Adversarial debiasing",
   "definition": "Training a predictor together with an adversary that tries to predict A from the predictor's representation or output; the predictor is rewarded for defeating the adversary (Edwards and Storkey, 2016; Zhang et al., 2018).",
   "chapters": [
    65
   ]
  },
  {
   "term": "Adversary",
   "definition": "A second model that tries to predict the protected attribute A from the predictor's output, from the output and the true label, or from an internal representation. Its parameters are written U.",
   "chapters": [
    53
   ]
  },
  {
   "term": "Adversary weight (α or λ)",
   "definition": "The coefficient that sets how strongly the predictor is rewarded for increasing the adversary's loss.",
   "chapters": [
    53
   ]
  },
  {
   "term": "Adverse action",
   "definition": "In U.S. credit law, a refusal to grant credit in substantially the amount or on substantially the terms requested (unless the applicant accepts a counteroffer), a termination of an account or an unfavorable change in its terms that does not affect all or substantially all of a class of accounts, or a refusal to increase available credit on application (12 C.F.R. § 1002.2(c)). Adverse action triggers the notice duties of 12 C.F.R. § 1002.9, and ECOA entitles each applicant against whom adverse action is taken to a statement of the specific reasons (15 U.S.C. § 1691(d)).",
   "chapters": [
    88,
    91
   ]
  },
  {
   "term": "Adverse action notice",
   "definition": "In United States credit law, the notice a creditor must give an applicant against whom adverse action is taken, including a statement of reasons (15 U.S.C. § 1691(d)).",
   "chapters": [
    83
   ]
  },
  {
   "term": "Adverse selection in certification",
   "definition": "The situation in which the entities that seek and obtain a certification are, on average, of lower quality than those that do not, because certification is cheap or lax relative to its value to low-quality applicants (Edelman, 2009).",
   "chapters": [
    101
   ]
  },
  {
   "term": "Affected person, deployer, provider",
   "definition": "In the AI Act, the provider develops a high-risk AI system and places it on the market; the deployer uses it under its authority; the affected person is the natural person subject to a decision taken on the basis of its output.",
   "chapters": [
    79
   ]
  },
  {
   "term": "Agentic AI system",
   "definition": "Shavit et al. (2023) characterize agentic AI systems by the ability to take actions that consistently contribute toward achieving goals over an extended period, without their behavior having been specified in advance, and define the degree of agenticness as \"the degree to which a system can adaptably achieve complex goals in complex environments with limited direct supervision\". Chan et al. (2023) treat agency as a matter of degree rather than a binary property and identify four characteristics that, especially in combination, increase it: underspecification, directness of impact, goal-directedness and long-term planning.",
   "chapters": [
    69,
    80
   ]
  },
  {
   "term": "Agentic algorithmic system",
   "definition": "Chan et al. (2023) avoid a binary definition of agency and identify four characteristics that, especially in combination, increase a system's agency: underspecification (the system achieves a goal whose means were not specified), directness of impact (its actions affect the world without a human intermediary), goal-directedness and long-term planning.",
   "chapters": [
    113
   ]
  },
  {
   "term": "Agenticness",
   "definition": "The degree to which a system can adaptably achieve complex goals in complex environments with limited direct supervision (Shavit et al., 2023). Agenticness is a property that comes in degrees, not a category; Shavit et al. (2023) distinguish it from consciousness, moral patienthood and self-motivation, and from anthropomorphism. Shavit et al. (2023) break it into four components: goal complexity, environmental complexity, adaptability and independent execution.",
   "chapters": [
    113
   ]
  },
  {
   "term": "Aggregation paradox",
   "definition": "A pattern that holds in aggregate but disappears or reverses within subgroups, or the reverse; Simpson's paradox is the best-known form.",
   "chapters": [
    63
   ]
  },
  {
   "term": "AI agent",
   "definition": "A system capable of pursuing complex goals with limited supervision (Chan et al., 2024). In current practice an agent typically combines a large language model with memory, planning and access to tools, interleaving reasoning and actions; the ReAct pattern, in which a model alternates reasoning traces with actions that query external sources, is one influential design (Yao et al., 2022).",
   "chapters": [
    113
   ]
  },
  {
   "term": "AI assistant",
   "definition": "Gabriel et al. (2024) define advanced AI assistants as artificial agents with natural language interfaces whose function is to plan and execute sequences of actions on behalf of a user, across one or more domains, in line with the user's expectations.",
   "chapters": [
    69
   ]
  },
  {
   "term": "AI-assisted decision making",
   "definition": "A setting in which an AI system makes a recommendation and a human makes the final decision, accepting or overriding the recommendation (Zhang, Liao and Bellamy, 2020).",
   "chapters": [
    78
   ]
  },
  {
   "term": "AI hazard",
   "definition": "In the OECD vocabulary, an event, circumstance or series of events in which the development, use or malfunction of AI systems could plausibly lead to an AI incident. A hazard is a potential incident; an incident is a hazard that has materialized.",
   "chapters": [
    109
   ]
  },
  {
   "term": "AI incident",
   "definition": "In the definition proposed by the OECD, an event, circumstance or series of events in which the development, use or malfunction of one or more AI systems directly or indirectly leads to harm: injury or harm to the health of a person or groups of people; disruption of the management and operation of critical infrastructure; violations of human rights or a breach of obligations under the applicable law intended to protect fundamental, labor and intellectual property rights; or harm to property, communities or the environment (OECD, 2024; Aranda and Perset, 2024).",
   "chapters": [
    109
   ]
  },
  {
   "term": "AI system",
   "definition": "In the AI Act, a machine-based system designed to operate with varying levels of autonomy, which may exhibit adaptiveness after deployment and which, for explicit or implicit objectives, infers from the input it receives how to generate outputs such as predictions, content, recommendations or decisions that can influence physical or virtual environments (Article 3(1)). The definition covers most machine learning models used for decisions about people.",
   "chapters": [
    86
   ]
  },
  {
   "term": "Alert fatigue",
   "definition": "The state in which so many alerts fire, most of them unimportant, that the people who receive them stop responding. Chapter 59 describes it as the tendency of people to ignore warnings that fire too often or without clear meaning.",
   "chapters": [
    59,
    61
   ]
  },
  {
   "term": "Alert prioritization and routing",
   "definition": "Scoring a confirmed drift event by its context (affected group, legal protection, population size, velocity) and sending it to the channel and team that match its severity.",
   "chapters": [
    62
   ]
  },
  {
   "term": "Algorithmic confounding",
   "definition": "The situation in which recommender systems are evaluated or trained with data from users who were already exposed to algorithmic recommendations (Chaney et al., 2018).",
   "chapters": [
    27
   ]
  },
  {
   "term": "Algorithmic monoculture",
   "definition": "The situation in which many decision makers rely on the same algorithm, or on algorithms that share components such as training data or a foundation model (Kleinberg and Raghavan, 2021; Bommasani et al., 2022).",
   "chapters": [
    114
   ]
  },
  {
   "term": "Algorithmic monoculture and outcome homogenization",
   "definition": "Monoculture is the use of the same algorithm, or systems sharing components such as training data, by many decision-makers (Kleinberg and Raghavan, 2021). Outcome homogenization is \"the extent to which particular individuals or groups experience negative outcomes from all decision-makers\" (Bommasani et al., 2022).",
   "chapters": [
    69
   ]
  },
  {
   "term": "Algorithmic recourse",
   "definition": "The ability of a person to obtain a desired outcome from a fixed model by altering actionable input variables (Ustun et al., 2019). Recourse is a property of the model, the person and the set of feasible actions together.",
   "chapters": [
    74
   ]
  },
  {
   "term": "Alignment",
   "definition": "Post-training methods, such as fine-tuning on human preference judgments, that shape a model's behavior toward stated norms such as helpfulness and harmlessness.",
   "chapters": [
    112
   ]
  },
  {
   "term": "Allocational harm",
   "definition": "A disparate distribution of resources or opportunities between social groups, either through direct discrimination (disparate treatment explicitly due to group membership) or indirect discrimination (disparate treatment through facially neutral factors such as proxies) (Gallegos et al., 2024, Table 1).",
   "chapters": [
    112
   ]
  },
  {
   "term": "Analysis plan",
   "definition": "A written specification, made before outcomes are examined, of the groups, criteria, metrics, tolerances, statistical methods and decision rules of an assessment. Fixing it in advance is called preregistration (Nosek et al., 2018).",
   "chapters": [
    36
   ]
  },
  {
   "term": "Answerability",
   "definition": "The core of accountability in the analysis of Novelli, Taddeo and Floridi (2024), who define accountability in AI in terms of answerability.",
   "chapters": [
    10
   ]
  },
  {
   "term": "Anticipatory compliance",
   "definition": "Designing to meet requirements that are likely but not yet binding, so that formalization does not force a redesign.",
   "chapters": [
    85
   ]
  },
  {
   "term": "Appropriate reliance",
   "definition": "Reliance on AI advice when it is correct and overriding it when it is wrong. Its failure in one direction is automation bias, which AI Act Article 14(4)(b) names: the tendency of automatically relying or over-relying on the output of a high-risk AI system.",
   "chapters": [
    82
   ]
  },
  {
   "term": "Appropriateness of reliance (AoR)",
   "definition": "A two-dimensional measure of reliance on AI advice consisting of relative AI reliance (RAIR) and relative self-reliance (RSR), defined formally in Chapter 2 (Schemmer et al., 2023).",
   "chapters": [
    2
   ]
  },
  {
   "term": "Argument, warrant and backing",
   "definition": "In Toulmin's scheme, as summarized by Bishop, Bloomfield and Guerra (2004), a claim is justified on the basis of grounds (specific facts), the reasoning from grounds to claim is the warrant, and the warrant may rest on backing, such as the validation of the scientific or engineering laws it uses. In CAE, the argument combines Toulmin's warrant and backing.",
   "chapters": [
    94
   ]
  },
  {
   "term": "Artifact",
   "definition": "A document produced by a phase and kept in the audit file: the scoping record, historical context assessment record, selection decision record, bias source register, metric specification record, findings and decision record, audit report and monitoring plan.",
   "chapters": [
    110
   ]
  },
  {
   "term": "Assessment dimension",
   "definition": "One of five properties on which each bias source is scored from 1 to 5: severity, scope, persistence, historical alignment and intervention feasibility.",
   "chapters": [
    35
   ]
  },
  {
   "term": "Assistive and punitive decisions",
   "definition": "An assistive decision gives a benefit to those selected (a loan, an interview, extra care); a punitive decision imposes a burden on those selected (detention, investigation, denial of service). Saleiro et al. (2018) use this distinction to guide metric choice.",
   "chapters": [
    22
   ]
  },
  {
   "term": "Association",
   "definition": "Statistical dependence between variables, for example P(D = 1 | A = a) ≠ P(D = 1 | A = b). It can be estimated from observational data alone.",
   "chapters": [
    30
   ]
  },
  {
   "term": "Assurance",
   "definition": "The process of measuring, evaluating and communicating something about a system or process; for AI, its trustworthiness (DSIT, 2024). The Centre for Data Ethics and Innovation described AI assurance as a way to justify the level of trust that people place in AI systems (Centre for Data Ethics and Innovation [CDEI], 2021). Chapter 81 uses the narrower sense of establishing, communicating and maintaining justified confidence that a system meets stated requirements, typically through an assurance case or audit report.",
   "chapters": [
    81,
    93
   ]
  },
  {
   "term": "Assurance case",
   "definition": "A reasoned and compelling argument, supported by a body of evidence, that a system, service or organization will operate as intended for a defined application in a defined environment (Assurance Case Working Group [ACWG], 2021). In the international standard, it is an auditable artifact that provides a convincing and sound argument for a claim on the basis of tangible evidence under a given context (ISO/IEC/IEEE, 2022).",
   "chapters": [
    93,
    94
   ]
  },
  {
   "term": "Assurance engagement",
   "definition": "An engagement in which a practitioner obtains evidence to express a conclusion designed to enhance the confidence of intended users about the outcome of measuring or evaluating a subject matter against criteria (International Auditing and Assurance Standards Board [IAASB], 2013).",
   "chapters": [
    101
   ]
  },
  {
   "term": "Assurance evidence",
   "definition": "Documented results of tests, measurements and reviews that support a claim about a system's properties, bound to the system version and context to which they apply (Chapters 93 to 99).",
   "chapters": [
    115
   ]
  },
  {
   "term": "AUC and xAUC",
   "definition": "The area under the ROC curve, equal to the probability that a randomly chosen positive is ranked above a randomly chosen negative; the cross-group AUC (xAUC) compares positives of one group with negatives of another (Kallus and Zhou, 2019).",
   "chapters": [
    21
   ]
  },
  {
   "term": "Audit",
   "definition": "A structured examination of a system or of the processes that produced it against criteria. A compliance audit reviews adherence to internal policies, external regulations and legal requirements; a bias audit assesses the inputs and outputs of an algorithmic system to determine whether there is unfair bias in the data or outcomes (DSIT, 2024).",
   "chapters": [
    93
   ]
  },
  {
   "term": "Audit file",
   "definition": "The complete set of artifacts, data references, code versions and sign-offs that documents an audit and allows it to be reproduced.",
   "chapters": [
    110
   ]
  },
  {
   "term": "Audit trail",
   "definition": "A chronological, attributable record of events and decisions relevant to a system (who did what, when, on which version, with what result), kept so that past states and actions can be reconstructed and verified.",
   "chapters": [
    95
   ]
  },
  {
   "term": "Authentication",
   "definition": "Establishing that a claimant controls one or more authenticators bound to an account or identifier.",
   "chapters": [
    98
   ]
  },
  {
   "term": "Automated employment decision tool (AEDT)",
   "definition": "Under New York City law (Administrative Code § 20-870, as clarified by the implementing rule), a tool based on machine learning, statistical modeling, data analytics or artificial intelligence whose simplified output (a score, tag, classification, recommendation or ranking) is used to substantially assist or replace discretionary decision making in employment decisions; the rule explains \"substantially assist or replace\" as relying solely on the output, weighting it more than any other criterion in a set, or using it to overrule conclusions derived from other factors, including human decision making.",
   "chapters": [
    88,
    91
   ]
  },
  {
   "term": "Automated individual decision",
   "definition": "A decision based solely on automated processing, including profiling, which produces legal effects concerning a person or similarly significantly affects that person (GDPR Article 22(1)). Swiss law uses the term for a decision based exclusively on automated processing with a legal consequence for, or a considerable adverse effect on, the person (FADP Article 21(1)).",
   "chapters": [
    9,
    79,
    87
   ]
  },
  {
   "term": "Automation",
   "definition": "Technology that actively selects data, transforms information, makes decisions or controls processes (Lee and See, 2004). AI-based decision support is automation in this sense.",
   "chapters": [
    2
   ]
  },
  {
   "term": "Automation bias",
   "definition": "In the wording of the EU AI Act, the possible tendency of automatically relying or over-relying on the output produced by an AI system (Article 14(4)(b) of Regulation (EU) 2024/1689). The Act names it as a risk that people assigned to human oversight of a high-risk AI system must remain aware of.",
   "chapters": [
    2,
    28,
    78
   ]
  },
  {
   "term": "Autonomous (free-standing) guarantee",
   "definition": "A non-discrimination clause that applies in any field regulated by public authorities, independently of the other rights in the instrument, such as Article 26 ICCPR and Article 1 of Protocol No. 12 to the ECHR.",
   "chapters": [
    84
   ]
  },
  {
   "term": "Average run length (ARL)",
   "definition": "The expected number of observations until a sequential detector raises an alarm: ARL0 when nothing has changed (a measure of false alarm frequency, larger is better) and ARL1 after a change of a given size (a measure of detection delay, smaller is better).",
   "chapters": [
    62
   ]
  },
  {
   "term": "Back-door criterion",
   "definition": "A graphical condition under which adjusting for a set of variables X_C identifies the causal effect of X_S on Y: no variable in X_C is a descendant of X_S, and X_C blocks every path between X_S and Y that contains an arrow into X_S (Zhao and Hastie, 2021).",
   "chapters": [
    75
   ]
  },
  {
   "term": "Balance for the negative class",
   "definition": "The average score of people with Y = 0 is the same in every group: E[R | Y = 0, A = a] = E[R | Y = 0, A = b] (Kleinberg et al., 2017).",
   "chapters": [
    16,
    19
   ]
  },
  {
   "term": "Balance for the positive class",
   "definition": "The average score of people with Y = 1 is the same in every group: E[R | Y = 1, A = a] = E[R | Y = 1, A = b] (Kleinberg et al., 2017).",
   "chapters": [
    16,
    19
   ]
  },
  {
   "term": "Balanced error rate (BER)",
   "definition": "The average of the error rates on the two classes of a binary prediction. Feldman et al. (2015) link disparate impact to how well A can be predicted from the other features, measured by BER.",
   "chapters": [
    48
   ]
  },
  {
   "term": "Base rate (prevalence)",
   "definition": "The proportion of a group for which Y = 1, written p_a = P(Y = 1 | A = a). Also called prevalence. Unequal base rates, p_a ≠ p_b, are the central condition of all impossibility results (Chapter 19).",
   "chapters": [
    12,
    13,
    15,
    19,
    22
   ]
  },
  {
   "term": "Baseline (reference)",
   "definition": "The input, or distribution of inputs, against which an attribution is computed. Attributions are statements about the difference between the prediction and the baseline prediction, so the choice of baseline is part of the explanation (Chapter 70).",
   "chapters": [
    72
   ]
  },
  {
   "term": "Benchmark population",
   "definition": "The reference distribution against which dataset composition is compared, for example census figures for the market served or the applicant pool of a lender. The choice of benchmark is itself a judgment that must be documented.",
   "chapters": [
    25
   ]
  },
  {
   "term": "Benevolence",
   "definition": "The extent to which a trustee is believed to want to do good to the trustor, aside from an egocentric profit motive (Mayer et al., 1995).",
   "chapters": [
    1
   ]
  },
  {
   "term": "Beyond compliance",
   "definition": "Practices that an organization adopts because its ethical commitments require them, although no applicable law does. In this compendium such practices are marked as methodological positions (METHOD), distinguished from legal requirements (LEGAL).",
   "chapters": [
    92
   ]
  },
  {
   "term": "Bias",
   "definition": "In ISO/IEC TR 24027 (clause 3.2.2), a \"systematic difference in treatment of certain objects, people, or groups in comparison to others\", where treatment includes perception, observation, representation, prediction or decision.",
   "chapters": [
    89
   ]
  },
  {
   "term": "Bias amplification",
   "definition": "A model's predictions show a stronger association between a label and a group than the training data do (Zhao et al., 2017; Wang et al., 2019).",
   "chapters": [
    65,
    68
   ]
  },
  {
   "term": "Bias entry point",
   "definition": "The stage of the machine learning life cycle (data collection, model development, deployment) at which a bias source operates.",
   "chapters": [
    35
   ]
  },
  {
   "term": "Bias source",
   "definition": "A specific, nameable mechanism in a particular system through which unwanted disparities in outcomes or errors may arise, for example \"performance ratings used as the target reflect historical gender disparities\".",
   "chapters": [
    35
   ]
  },
  {
   "term": "Bias type",
   "definition": "A class of bias sources sharing a mechanism and a life cycle stage, such as measurement bias. The Bias Source Identification Method uses seven types, following Suresh and Guttag (2021).",
   "chapters": [
    35
   ]
  },
  {
   "term": "Binding",
   "definition": "A verifiable association between two things: a credential and its holder's device key, a DID and an organization's legal name, a hash and the file it identifies.",
   "chapters": [
    98
   ]
  },
  {
   "term": "Biometric data, emotion recognition and biometric categorization",
   "definition": "In EU law, biometric data are personal data resulting from specific technical processing of physical, physiological or behavioral characteristics, such as facial images; an emotion recognition system identifies or infers emotions or intentions from biometric data; a biometric categorization system assigns people to categories on the basis of biometric data (Regulation (EU) 2024/1689, Article 3(34), (39) and (40)).",
   "chapters": [
    68
   ]
  },
  {
   "term": "Black box",
   "definition": "Rudin (2019) distinguishes two kinds: a function too complicated for any human to comprehend, and a function that is proprietary. A model can be a black box of the second kind while being simple, as Rudin argues for the COMPAS recidivism tool.",
   "chapters": [
    70
   ]
  },
  {
   "term": "Block design",
   "definition": "A design that groups experimental units by a nuisance factor (such as day of week) and compares variants within blocks.",
   "chapters": [
    64
   ]
  },
  {
   "term": "Bootstrap",
   "definition": "A resampling method that approximates the sampling distribution of an estimate by recomputing it on many samples drawn with replacement from the observed data (Efron, 1979).",
   "chapters": [
    41
   ]
  },
  {
   "term": "Breakdown point",
   "definition": "The largest proportion of arbitrarily corrupted observations an estimator can tolerate before it can be made arbitrarily wrong (Hampel, 1971).",
   "chapters": [
    42
   ]
  },
  {
   "term": "Brussels effect",
   "definition": "The process by which European Union rules come to govern conduct outside the Union because firms operating in many markets find it efficient to adopt the strictest standard globally (Bradford, 2020).",
   "chapters": [
    86
   ]
  },
  {
   "term": "Business necessity and job-relatedness",
   "definition": "The United States employment defense to disparate impact: the respondent shows that the challenged practice is job related for the position in question and consistent with business necessity (42 U.S.C. § 2000e-2(k)).",
   "chapters": [
    83,
    88
   ]
  },
  {
   "term": "Calibration",
   "definition": "A score R is calibrated if, among all individuals who receive the score r, the share with Y = 1 equals r, for every r.",
   "chapters": [
    16,
    56
   ]
  },
  {
   "term": "Calibration disparity",
   "definition": "A summary of how much calibration differs between groups, for example the difference between the largest and the smallest group ECE.",
   "chapters": [
    56
   ]
  },
  {
   "term": "Calibration (of recommendations)",
   "definition": "The property that the proportions of the categories in a user's recommendations reflect the proportions in the user's past interests (Steck, 2018). It differs from the calibration of a risk score (Chapter 16).",
   "chapters": [
    66
   ]
  },
  {
   "term": "Calibration, resolution and specificity",
   "definition": "Calibration is the correspondence between a person's trust in the automation and the automation's capabilities; resolution is how precisely a judgment of trust differentiates levels of automation capability; specificity is the degree to which trust is associated with a particular component or aspect of the trustee, as functional specificity (sub-functions and modes) or temporal specificity (changes over time and context) (Lee and See, 2004).",
   "chapters": [
    2
   ]
  },
  {
   "term": "Calibration set",
   "definition": "A labeled data set, disjoint from the training and test sets, used only to fit recalibration maps.",
   "chapters": [
    56
   ]
  },
  {
   "term": "Calibration within groups",
   "definition": "The score is calibrated separately in each group defined by the protected attribute A. It is the score-level form of sufficiency (Chapter 16).",
   "chapters": [
    16,
    19,
    56
   ]
  },
  {
   "term": "Capture conditions",
   "definition": "The physical circumstances in which an image or recording is made: illumination, camera and microphone quality, resolution, background, pose and occlusion.",
   "chapters": [
    68
   ]
  },
  {
   "term": "Causal discovery",
   "definition": "The inference of causal structure from data, usually up to a class of graphs that imply the same observable constraints.",
   "chapters": [
    33
   ]
  },
  {
   "term": "Causal graph",
   "definition": "The directed graph with an edge from each parent to each variable it determines.",
   "chapters": [
    18
   ]
  },
  {
   "term": "Causal hierarchy",
   "definition": "The classification of questions by Pearl (2019) into three levels: association (what does observing X tell me about Y?), intervention (what happens to Y if I set X?) and counterfactuals (what would have happened to Y had X been different, given what actually happened?). A question at a level can be answered only with information from that level or a higher one.",
   "chapters": [
    75
   ]
  },
  {
   "term": "Causal path",
   "definition": "A directed route from A to the prediction through intermediate variables, for example A → neighborhood → property value → score.",
   "chapters": [
    32
   ]
  },
  {
   "term": "Censored representation",
   "definition": "A representation trained to be informative for the task and uninformative about a sensitive variable, by minimizing an adversary's ability to recover that variable (Edwards and Storkey, 2016).",
   "chapters": [
    53
   ]
  },
  {
   "term": "Certification",
   "definition": "A third-party attestation that a product, process, system or organization meets specified requirements, typically documented by a certificate with a scope and a validity period. Under the EU AI Act, certificates issued by notified bodies are one output of conformity assessment (Regulation (EU) 2024/1689, Art. 44).",
   "chapters": [
    101
   ]
  },
  {
   "term": "Chain of thought (CoT)",
   "definition": "A series of intermediate reasoning steps that a model generates before its final answer, either because the prompt contains such examples (Wei et al., 2022) or because the model has been trained to reason before answering.",
   "chapters": [
    80
   ]
  },
  {
   "term": "Changepoint",
   "definition": "A point in time at which the parameters of the process generating a series change abruptly.",
   "chapters": [
    62
   ]
  },
  {
   "term": "Claim",
   "definition": "An assertion put forward for general acceptance, typically a statement about a property of a system or subsystem; claims supporting an argument are sub-claims, and claims asserted as true without justification become assumptions (Bloomfield and Netkachova, 2014). In this compendium a fairness claim always names the system, the population, the period, the criterion and the tolerance. In verifiable credentials (Chapter 97), a claim is an assertion made about a subject, expressed as a subject, property and value.",
   "chapters": [
    93,
    94,
    97
   ]
  },
  {
   "term": "Claim, argument, evidence",
   "definition": "A claim is a statement about the system to be justified (\"the reason codes identify the principal factors of each decision\"), and evidence is the material that supports it (test results, records). The argument explains why the evidence supports the claim, including the assumptions on which it rests; Chapter 94 treats assurance arguments in full.",
   "chapters": [
    81
   ]
  },
  {
   "term": "Classification consistency",
   "definition": "The extent to which similar systems receive the same classification, regardless of who carries out the assessment. It can be measured as percentage agreement or as a chance-corrected agreement coefficient such as Cohen's kappa (Cohen, 1960).",
   "chapters": [
    90
   ]
  },
  {
   "term": "Classification system",
   "definition": "A scheme that divides a domain into discrete categories (diagnoses, races, genders, occupations, credit grades). In machine learning, classification systems define both input features and prediction targets.",
   "chapters": [
    7
   ]
  },
  {
   "term": "Closed and open lists",
   "definition": "A closed list names the protected grounds exhaustively (for example the EU equal treatment directives). An open list names examples and adds a residual category (\"any ground such as\", \"other status\"), as in Article 21 of the EU Charter of Fundamental Rights and Article 14 of the European Convention on Human Rights.",
   "chapters": [
    24
   ]
  },
  {
   "term": "Cluster randomization",
   "definition": "Randomization of whole clusters (cities, regions, stores) rather than individuals, used when individual randomization would let treatment spill over to control units.",
   "chapters": [
    64
   ]
  },
  {
   "term": "Codification",
   "definition": "The translation of social categories into computational representations, such as one-hot vectors, integer codes or numeric scales.",
   "chapters": [
    7
   ]
  },
  {
   "term": "Cognitive chunk",
   "definition": "The basic unit of an explanation, such as a raw feature, a derived concept or a prototype (Doshi-Velez and Kim, 2017). The form, number and composition of chunks shape how understandable an explanation is.",
   "chapters": [
    70
   ]
  },
  {
   "term": "Cognitive forcing function",
   "definition": "An intervention that compels people to engage analytically with a decision, for example by asking for their own judgment before showing the AI's recommendation (Buçinca, Malaya and Gajos, 2021).",
   "chapters": [
    78
   ]
  },
  {
   "term": "Collider",
   "definition": "A variable that is a common effect of two other variables. Conditioning on a collider, including selecting data by it, creates an association between its causes (Elwert and Winship, 2014).",
   "chapters": [
    30
   ]
  },
  {
   "term": "Collider on a path",
   "definition": "A node where two arrows on the path meet head to head (→ V ←).",
   "chapters": [
    31
   ]
  },
  {
   "term": "Comparator",
   "definition": "The person or group, real or hypothetical, against whom the treatment of the claimant is compared. The words \"would be treated\" allow a hypothetical comparator.",
   "chapters": [
    23
   ]
  },
  {
   "term": "Complementary performance",
   "definition": "Team performance that is higher than that of either the human or the AI working alone (Bansal et al., 2021).",
   "chapters": [
    78
   ]
  },
  {
   "term": "Completeness",
   "definition": "The extent to which an explanation covers the model's behavior, either its internal reasoning or its output (Nauta et al., 2023).",
   "chapters": [
    77
   ]
  },
  {
   "term": "Compliance risk",
   "definition": "The risk that a requirement is not met, assessed by the severity of the consequences, the likelihood of failure, the difficulty of detecting a gap, the complexity of remediation and the level of regulatory scrutiny.",
   "chapters": [
    85
   ]
  },
  {
   "term": "Comprehensive ecosystem",
   "definition": "A library that aims to cover most of the fairness lifecycle in one package: data analysis, bias detection, mitigation across pre-, in- and post-processing, and often explanation.",
   "chapters": [
    58
   ]
  },
  {
   "term": "Computational reproducibility",
   "definition": "Obtaining consistent computational results using the same input data, computational steps, methods, code and conditions of analysis (National Academies of Sciences, Engineering, and Medicine, 2019).",
   "chapters": [
    96
   ]
  },
  {
   "term": "Concept drift (concept shift)",
   "definition": "The relation between the input data and the target variable changes over time (Gama, Žliobaitė, Bifet, Pechenizkiy and Bouchachia, 2014): P_T(Y | X) ≠ P_S(Y | X).",
   "chapters": [
    34
   ]
  },
  {
   "term": "Conceptual soundness, ongoing monitoring, outcomes analysis",
   "definition": "The three core elements of model validation in the US supervisory guidance on model risk management: assessing the quality of model design and construction; confirming that the model is appropriately implemented and performing as intended over time; and comparing model outputs with actual outcomes (Board of Governors of the Federal Reserve System and Office of the Comptroller of the Currency, 2011).",
   "chapters": [
    81
   ]
  },
  {
   "term": "Conditional generation",
   "definition": "Generation in which the model receives conditioning variables, for example a value of A or of the outcome Y, and produces samples from the corresponding conditional distribution. It allows the analyst to specify how many records of each group are produced.",
   "chapters": [
    50
   ]
  },
  {
   "term": "Conditional statistical parity",
   "definition": "Independence within strata defined by a set of factors considered legitimate: Ŷ ⊥ A | L (Corbett-Davies et al., 2017).",
   "chapters": [
    14
   ]
  },
  {
   "term": "Conditional use accuracy equality",
   "definition": "Equal PPV and equal NPV across groups (Berk et al., 2021); this is sufficiency for a binary prediction.",
   "chapters": [
    16
   ]
  },
  {
   "term": "Confidence argument",
   "definition": "A separate argument about how much confidence an assertion in the main argument deserves, attached in GSN at an assurance claim point (ACWG, 2021).",
   "chapters": [
    94
   ]
  },
  {
   "term": "Confidence interval",
   "definition": "An interval computed by a procedure that, over repeated sampling, contains the true value with a stated probability, the confidence level (for example 95 percent).",
   "chapters": [
    13,
    41
   ]
  },
  {
   "term": "Confirmatory and exploratory evaluation",
   "definition": "A confirmatory evaluation is designed to provide conclusive evidence about a stated hypothesis; an exploratory evaluation searches for patterns that may motivate later confirmatory work (Barocas et al., 2021).",
   "chapters": [
    36
   ]
  },
  {
   "term": "Conformity assessment",
   "definition": "The demonstration that specified requirements are fulfilled (International Organization for Standardization and International Electrotechnical Commission [ISO/IEC], 2020). Under the AI Act (Chapter 86), it is the procedure by which a provider demonstrates that a high-risk AI system meets the requirements of Chapter III, Section 2 of the Act before the system is placed on the market or put into service.",
   "chapters": [
    86,
    93
   ]
  },
  {
   "term": "Confounder",
   "definition": "A variable that influences both the exposure of interest (here often the protected attribute or a decision) and the outcome, so that their association mixes causal effect and common cause.",
   "chapters": [
    30,
    33
   ]
  },
  {
   "term": "Consequential decision",
   "definition": "A decision with a material effect on access to, eligibility for or the terms of an important opportunity or service. Colorado's 2026 automated decision-making law, for example, covers decisions concerning education, employment, housing, lending, insurance, health care and essential government services (SB 26-189; Colorado General Assembly, 2026).",
   "chapters": [
    91
   ]
  },
  {
   "term": "Consequentialism",
   "definition": "The family of theories holding that choices are to be assessed by the states of affairs they bring about; the right choice is the one that brings about the most good, however \"the good\" is specified (Alexander and Moore, 2024; Sinnott-Armstrong, 2023).",
   "chapters": [
    4
   ]
  },
  {
   "term": "Consistency",
   "definition": "The property that identical inputs receive identical explanations, and, for methods that only observe inputs and outputs, implementation invariance: two models with the same outputs for all inputs receive the same explanations (Nauta et al., 2023). In individual fairness (Chapter 17), consistency is a measure of how often a model gives an individual the same prediction as the individual's nearest neighbors in feature space (Zemel et al., 2013).",
   "chapters": [
    17,
    77
   ]
  },
  {
   "term": "Constraint",
   "definition": "A condition C(θ) ≤ ε (inequality) or C(θ) = 0 (equality) that the model parameters θ must satisfy. In fairness, C measures a disparity between groups and ε is the tolerance.",
   "chapters": [
    51
   ]
  },
  {
   "term": "Construct",
   "definition": "The concept a decision actually cares about, such as creditworthiness, job performance, health need or criminal behavior. The construct is distinct from the measured target Y that stands for it (Chapter 12).",
   "chapters": [
    25
   ]
  },
  {
   "term": "Construct and operationalization",
   "definition": "A construct is an unobservable theoretical quantity, such as creditworthiness or need for care; an operationalization is the measurement model that infers it from observable properties (Jacobs and Wallach, 2021).",
   "chapters": [
    36
   ]
  },
  {
   "term": "Construct space, observed space and decision space",
   "definition": "Friedler, Scheidegger and Venkatasubramanian (2016) distinguish the space of unobservable but meaningful variables for a decision (construct space), the space of the features actually observed (observed space) and the space of decisions (decision space).",
   "chapters": [
    45
   ]
  },
  {
   "term": "Construct validity",
   "definition": "The degree to which an operationalization measures the theoretical construct it is meant to measure. Jacobs and Wallach (2021) argue that many fairness harms result from mismatches between constructs and their operationalizations, and that fairness itself is an essentially contested construct.",
   "chapters": [
    115
   ]
  },
  {
   "term": "Constructed scenario",
   "definition": "A hypothetical but realistic case written to illustrate a method. Its context and mechanisms are plausible; its numbers are chosen for illustration.",
   "chapters": [
    111
   ]
  },
  {
   "term": "Containment",
   "definition": "Measures that stop or limit further harm while the cause is investigated: suspension, rollback, a manual review queue, a narrower scope of use.",
   "chapters": [
    109
   ]
  },
  {
   "term": "Contestability",
   "definition": "The ability to contest decisions made or supported by algorithmic systems (Lyons et al., 2021); in design terms, the property of a system that is responsive to human intervention throughout its life cycle (Alfrink et al., 2023).",
   "chapters": [
    9
   ]
  },
  {
   "term": "Context, assumption and justification",
   "definition": "In GSN a context presents contextual information or a reference to it; an assumption presents an intentionally unsubstantiated statement; a justification presents a statement of rationale (ACWG, 2021).",
   "chapters": [
    94
   ]
  },
  {
   "term": "Contingency table",
   "definition": "A cross-tabulation of counts, here of outcomes (for example positive and negative predictions) by group.",
   "chapters": [
    42
   ]
  },
  {
   "term": "Continuous assurance",
   "definition": "An assurance arrangement in which the evidence behind a claim is renewed on a declared schedule or by monitoring with alert thresholds, so that the claim describes the present state of the system rather than a past snapshot.",
   "chapters": [
    99
   ]
  },
  {
   "term": "Continuous deployment or delivery (CD)",
   "definition": "The automation of the path from a tested change to a release. Humble and Farley (2010) describe the deployment pipeline as an automated process that takes every change from check-in to release.",
   "chapters": [
    60
   ]
  },
  {
   "term": "Continuous fairness validation",
   "definition": "The practice of testing an AI system for bias repeatedly and automatically throughout its lifecycle, so that fairness regressions are caught when they occur.",
   "chapters": [
    60
   ]
  },
  {
   "term": "Continuous integration (CI)",
   "definition": "The practice of merging changes frequently into a shared branch, with each change automatically built and tested.",
   "chapters": [
    60
   ]
  },
  {
   "term": "Contractualism",
   "definition": "A family of theories, including Rawls's justice as fairness, that derive principles from what people could agree to under fair conditions, such as behind a veil of ignorance about their own position (Freeman, 2023).",
   "chapters": [
    4
   ]
  },
  {
   "term": "Contrastive explanation",
   "definition": "An explanation of why P happened instead of Q. Miller (2019) identifies contrastiveness as a central feature of how people seek and give explanations; a counterfactual explanation is contrastive by construction.",
   "chapters": [
    74
   ]
  },
  {
   "term": "Control effectiveness",
   "definition": "The degree to which a safeguard, such as human review, an appeal mechanism or bias testing, demonstrably reduces the probability or severity of harm.",
   "chapters": [
    90
   ]
  },
  {
   "term": "Controller",
   "definition": "The private person or federal body that decides on the purpose and means of processing. Most FADP duties discussed in Chapter 87 fall on the controller.",
   "chapters": [
    87
   ]
  },
  {
   "term": "Cooperative and non-cooperative (outside-in) audit",
   "definition": "A cooperative audit has access granted by the audited organization (data, code, documentation, staff). An outside-in audit studies a system through its observable behavior, for example by repeatedly querying it and observing outputs (Metaxa et al., 2021).",
   "chapters": [
    96
   ]
  },
  {
   "term": "Cost of fairness",
   "definition": "The reduction in utility caused by imposing a fairness constraint, relative to the unconstrained optimum (Corbett-Davies et al., 2017).",
   "chapters": [
    20
   ]
  },
  {
   "term": "Cost of recourse",
   "definition": "A measure of how difficult the cheapest action that flips the decision is for a person, for example the largest percentile shift in any feature (Ustun et al., 2019).",
   "chapters": [
    74
   ]
  },
  {
   "term": "Council of Europe AI Convention",
   "definition": "The Council of Europe Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law (CETS No. 225), opened for signature on 5 September 2024 and signed by Switzerland on 27 March 2025.",
   "chapters": [
    87
   ]
  },
  {
   "term": "Counterfactual",
   "definition": "A statement about what would have happened to a specific unit under a condition other than the one that actually held, given what was observed about that unit (Pearl, 2019). In a structural causal model (Chapter 18), the counterfactual Ŷ_{A←a}(U) is the value Ŷ would take, for background variables U, in the model modified by the intervention A = a.",
   "chapters": [
    18,
    30
   ]
  },
  {
   "term": "Counterfactual consistency",
   "definition": "The share of cases whose prediction does not change when only the protected attribute, or a direct reference to it in the input, is changed (Garg et al., 2019; Chapter 32 for the causal concept).",
   "chapters": [
    65
   ]
  },
  {
   "term": "Counterfactual data augmentation",
   "definition": "The creation of paired records that differ in the protected attribute, and in what the protected attribute causally affects, while characteristics that should not depend on it are held fixed (Lu et al., 2020).",
   "chapters": [
    50
   ]
  },
  {
   "term": "Counterfactual demonstration",
   "definition": "A demonstration of how the system responds to inputs that differ only in a protected attribute (Chapter 32 treats counterfactual fairness formally).",
   "chapters": [
    105
   ]
  },
  {
   "term": "Counterfactual effect size",
   "definition": "The average difference between the counterfactual and the factual prediction over a population or group (a measure defined in Chapter 32 as a validant.ai method).",
   "chapters": [
    32
   ]
  },
  {
   "term": "Counterfactual evaluation",
   "definition": "Testing whether a model's output changes when only a group-identifying element of the input (a name, a pronoun, an identity term, a dialect feature) is changed.",
   "chapters": [
    67
   ]
  },
  {
   "term": "Counterfactual explanation (CFE)",
   "definition": "A statement of how the inputs to a decision would have had to differ for the decision to be different, typically the closest input x′ to the actual input x that the model assigns the desired outcome. Wachter et al. (2018) describe it as identifying the \"closest possible world\" in which the desired outcome occurs.",
   "chapters": [
    10,
    74
   ]
  },
  {
   "term": "Counterfactual fairness",
   "definition": "A predictor is counterfactually fair if, for every individual, the distribution of its prediction is the same in the actual world and in the counterfactual world in which the protected attribute takes any other attainable value, with everything not caused by the attribute held fixed (Kusner et al., 2017). Chapter 12 uses the term for the family of causal criteria that ask whether a prediction for an individual would change if the protected attribute had been different.",
   "chapters": [
    12,
    32
   ]
  },
  {
   "term": "Counterfactual input (social group substitution)",
   "definition": "A pair of inputs that are identical except for the social group mentioned, for example a résumé with a name typically associated with one group and the same résumé with a name associated with another. Comparing the outputs is the generative analogue of the counterfactual tests of Chapters 18 and 32.",
   "chapters": [
    112
   ]
  },
  {
   "term": "Counterfactual prediction",
   "definition": "Ŷ_{A←a'}(U): the prediction the model would produce for an individual with background variables U in the modified model in which A is set to a'.",
   "chapters": [
    32
   ]
  },
  {
   "term": "Counterfactual statement",
   "definition": "A statement of how a variation in the person's data would have changed the outcome (Chapter 74 gives the formal treatment). The Court of Justice named this form as one that may be sufficiently transparent and intelligible for profiling.",
   "chapters": [
    79
   ]
  },
  {
   "term": "Counterfactual violation rate",
   "definition": "The share of individuals whose counterfactual prediction differs from the factual one by more than a stated tolerance, or whose decision would flip (defined in Chapter 32 as a validant.ai method).",
   "chapters": [
    32
   ]
  },
  {
   "term": "Covariate shift",
   "definition": "The inputs change while the conditional label distribution remains the same (Rezaei, Liu, Memarrast and Ziebart, 2021): P_T(X) ≠ P_S(X) but P_T(Y | X) = P_S(Y | X). In fairness analysis the relevant version conditions on the group: P_T(X | A) may change while P_T(Y | X, A) = P_S(Y | X, A).",
   "chapters": [
    34
   ]
  },
  {
   "term": "Coverage",
   "definition": "The actual probability with which an interval procedure contains the true value. A procedure with nominal coverage 0.95 may have actual coverage well below or above it.",
   "chapters": [
    41
   ]
  },
  {
   "term": "Coverage probability",
   "definition": "The probability, over repeated sampling, that an interval procedure contains the true value. A 95 percent procedure has nominal coverage 0.95; its actual coverage may differ.",
   "chapters": [
    43
   ]
  },
  {
   "term": "Credential",
   "definition": "A set of one or more claims made by an issuer. The claims can concern different subjects.",
   "chapters": [
    97
   ]
  },
  {
   "term": "Credible interval",
   "definition": "In Bayesian analysis, an interval that contains the unknown quantity with a stated posterior probability, given the data and a prior distribution.",
   "chapters": [
    13,
    41
   ]
  },
  {
   "term": "Criteria",
   "definition": "The benchmarks against which the subject matter is measured or evaluated. For fairness, the criteria include the chosen fairness definitions, thresholds, groups and data (Chapters 22 and 44).",
   "chapters": [
    101
   ]
  },
  {
   "term": "Cross-functional responsibility",
   "definition": "A fairness duty assigned to a function other than data science, such as product management, engineering, legal, marketing, user research or executive leadership.",
   "chapters": [
    106
   ]
  },
  {
   "term": "d-separation",
   "definition": "The graphical criterion that determines which conditional independencies a DAG implies (Pearl, 1995), stated formally in Chapter 31.",
   "chapters": [
    31
   ]
  },
  {
   "term": "Daily Scrum",
   "definition": "A 15-minute event for the Developers to inspect progress toward the Sprint Goal and adapt the Sprint Backlog (Schwaber and Sutherland, 2020).",
   "chapters": [
    104
   ]
  },
  {
   "term": "Data audit for fairness",
   "definition": "A structured examination of a dataset intended to identify representation disparities, problematic associations, label problems and baseline violations of fairness criteria, documented so that it can guide and later evaluate interventions.",
   "chapters": [
    37
   ]
  },
  {
   "term": "Data cascades",
   "definition": "Compounding events causing negative downstream effects from data issues, triggered by practices that undervalue data quality (Sambasivan, Kapania, et al., 2021).",
   "chapters": [
    28
   ]
  },
  {
   "term": "Data fairness and classification fairness",
   "definition": "Xu et al. (2018) distinguish fairness of the generated dataset itself from fairness of a classifier trained on the generated data and applied to real data. The second does not follow from the first.",
   "chapters": [
    50
   ]
  },
  {
   "term": "Data production",
   "definition": "The full process by which data come to exist: schema design, collection, annotation or labeling, cleaning and validation. Data production involves actors with unequal power over its terms.",
   "chapters": [
    7
   ]
  },
  {
   "term": "Data representation",
   "definition": "The way information about people and the world is collected, categorized, encoded and documented for use by a learning system.",
   "chapters": [
    7
   ]
  },
  {
   "term": "Datasheet for a dataset",
   "definition": "A document that records a dataset's motivation, composition, collection process, preprocessing, uses, distribution and maintenance (Gebru et al., 2021).",
   "chapters": [
    107
   ]
  },
  {
   "term": "Decentralized identifier (DID)",
   "definition": "An identifier, standardized by the W3C, of the form did:<method>:<method-specific-id>, that resolves to a DID document containing verification material (such as public keys) and service endpoints, and that is designed to be decoupled from centralized registries, identity providers and certificate authorities, so that its controller can prove control over it without permission from any other party (Sporny et al., 2022).",
   "chapters": [
    97,
    100
   ]
  },
  {
   "term": "Decision based solely on automated processing",
   "definition": "A decision in which no human with the authority and competence to change it meaningfully takes part. Article 22(1) GDPR applies only to such decisions when they produce legal effects or similarly significant effects.",
   "chapters": [
    86
   ]
  },
  {
   "term": "Decision boundary covariance",
   "definition": "The covariance between the protected attribute and the signed distance of each example from a classifier's decision boundary, used as a convex proxy for demographic parity (Zafar et al., 2017).",
   "chapters": [
    52
   ]
  },
  {
   "term": "Decision (D)",
   "definition": "The action actually taken about the subject, such as approve or reject, interview or not, detain or release. D may equal Ŷ when decisions are automated, or may differ when a human or a policy intervenes.",
   "chapters": [
    12
   ]
  },
  {
   "term": "Decision rights",
   "definition": "The explicit statement of which decisions a role may take alone, which require consultation, and which lie outside its authority.",
   "chapters": [
    106
   ]
  },
  {
   "term": "Decision rule",
   "definition": "The rule that turns a measured value into pass, fail or \"could not verify\", for example \"pass if the upper bound of the 95% confidence interval of the gap is below the tolerance\".",
   "chapters": [
    103
   ]
  },
  {
   "term": "Decision subject",
   "definition": "A person (or other entity, such as a small business) about whom a decision is made. The population of decision subjects is modeled as a probability distribution; one subject is one draw from it.",
   "chapters": [
    12
   ]
  },
  {
   "term": "Decision threshold",
   "definition": "A value t such that the prediction is positive when the score reaches it: Ŷ = 1 if R ≥ t. In the compendium notation Y = 1 is the outcome the decision process looks for (for example, a loan is repaid), so a positive prediction is the favorable decision.",
   "chapters": [
    55
   ]
  },
  {
   "term": "Decision tier",
   "definition": "A class of fairness decisions with a common authority level and common input requirements, typically strategic, tactical and operational.",
   "chapters": [
    108
   ]
  },
  {
   "term": "Decision trigger",
   "definition": "A predefined rule that connects a metric value or alert to a governance action (for example, a deployment pause when a critical threshold is crossed).",
   "chapters": [
    63
   ]
  },
  {
   "term": "Declared ceiling",
   "definition": "A statement, attached to a finding, of what the assessment configuration could not have detected by construction (Glinz, 2026).",
   "chapters": [
    101
   ]
  },
  {
   "term": "Deep fake",
   "definition": "In EU law, \"AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful\" (Regulation (EU) 2024/1689, Article 3(60)).",
   "chapters": [
    69
   ]
  },
  {
   "term": "Defeater",
   "definition": "A reason for doubting a claim or an argument step. Goodenough, Weinstock and Klein (2013) distinguish rebutting defeaters (counterexamples to a claim), undermining defeaters (doubts about the validity of evidence) and undercutting defeaters (circumstances in which a conclusion is in doubt even though the premises are true).",
   "chapters": [
    94
   ]
  },
  {
   "term": "Definition of Done (DoD)",
   "definition": "A formal description of the state of the Increment when it meets the quality measures required for the product (Schwaber and Sutherland, 2020).",
   "chapters": [
    102
   ]
  },
  {
   "term": "Definition of Ready",
   "definition": "A team's own agreement on when a backlog item is ready to be selected; it is a common practice but not a Scrum Guide element.",
   "chapters": [
    104
   ]
  },
  {
   "term": "Delayed impact",
   "definition": "The effect of a decision policy on a variable of interest, such as a group's credit score, after the decisions have been acted on (Liu et al., 2018).",
   "chapters": [
    20,
    27,
    114
   ]
  },
  {
   "term": "Demographic parity difference",
   "definition": "The difference in selection rates between two groups, P(Ŷ = 1 | A = a) − P(Ŷ = 1 | A = b) (Chapter 14).",
   "chapters": [
    103
   ]
  },
  {
   "term": "Demographic parity (statistical parity)",
   "definition": "Independence for a binary prediction: the probability of a positive prediction is the same in every group. Other names in the literature include group fairness, equal acceptance rate and benchmarking (Verma and Rubin, 2018).",
   "chapters": [
    14
   ]
  },
  {
   "term": "Denominator",
   "definition": "The subset of a group on which a rate is computed: the whole group for a selection rate, the members with Y = 1 for a true positive rate, the members with Ŷ = 1 for a positive predictive value. The denominator sets the precision of the estimate (Chapter 43).",
   "chapters": [
    39,
    43
   ]
  },
  {
   "term": "Deontology",
   "definition": "A duty-based approach to ethics that assesses choices by their conformity to moral norms (duties, rights, constraints), which can forbid some actions even when they would produce better consequences (Alexander and Moore, 2024).",
   "chapters": [
    4
   ]
  },
  {
   "term": "Deployer",
   "definition": "A natural or legal person, public authority, agency or other body using an AI system under its authority, except in a personal non-professional activity (Article 3(4) AI Act). A bank that buys a credit scoring model is a deployer; a university that builds its own admissions model is both provider and deployer.",
   "chapters": [
    86
   ]
  },
  {
   "term": "Deployment bias",
   "definition": "Bias that arises when there is a mismatch between the problem a model is intended to solve and the way it is actually used, often because a system is built and evaluated as if it were fully autonomous while in reality it operates in a sociotechnical system moderated by institutions and human decision makers (Suresh and Guttag, 2021).",
   "chapters": [
    28
   ]
  },
  {
   "term": "Derived predictor",
   "definition": "A possibly randomized function of the score R and the protected attribute A alone (Hardt et al., 2016). Every threshold rule, with or without randomization, is a derived predictor. The features X play no further role once the score has been computed.",
   "chapters": [
    15,
    46,
    55
   ]
  },
  {
   "term": "Descendant of A",
   "definition": "A variable reachable from A along directed edges; a non-descendant is not causally affected by A.",
   "chapters": [
    18
   ]
  },
  {
   "term": "Design decision record",
   "definition": "A structured record of a fairness-relevant choice, its alternatives, its rationale, the trade-offs accepted and the expected effect on different groups.",
   "chapters": [
    95
   ]
  },
  {
   "term": "Detection latency",
   "definition": "The time from the moment a fairness problem emerges to the moment the monitoring system raises an alert about it.",
   "chapters": [
    61
   ]
  },
  {
   "term": "Detection technique",
   "definition": "An analysis that tests whether a suspected bias source is actually present and estimates its magnitude.",
   "chapters": [
    35
   ]
  },
  {
   "term": "Differential privacy",
   "definition": "A mathematical guarantee that the output of an analysis changes only slightly, in a precisely bounded way, when any one person's data are added or removed (Dwork and Roth, 2014).",
   "chapters": [
    63
   ]
  },
  {
   "term": "Diffusion of responsibility",
   "definition": "The situation in which every team member assumes that someone else owns an outcome, so that nobody acts on it.",
   "chapters": [
    102
   ]
  },
  {
   "term": "Digital attestation",
   "definition": "In this compendium, any signed statement by an identifiable party about a subject that a third party can verify. A verifiable credential is the standardized form of a digital attestation; the European Digital Identity Framework uses the term \"electronic attestation of attributes\" (Regulation (EU) 2024/1183).",
   "chapters": [
    97
   ]
  },
  {
   "term": "Digital trust",
   "definition": "Individuals' expectation that digital technologies and services, and the organizations providing them, will protect all stakeholders' interests and uphold societal expectations and values (World Economic Forum, 2022).",
   "chapters": [
    1
   ]
  },
  {
   "term": "Direct discrimination",
   "definition": "In EU law, the situation \"where one person is treated less favourably than another is, has been or would be treated in a comparable situation\" on a protected ground (Article 2(2)(a) of Council Directive 2000/43/EC; equivalent wording appears in Article 2(2)(a) of Directive 2000/78/EC, Article 2(a) of Directive 2004/113/EC and Article 2(1)(a) of Directive 2006/54/EC). In causal fairness analysis (Chapter 18), direct discrimination is the influence of A on the decision along the direct edge A → decision.",
   "chapters": [
    18,
    23
   ]
  },
  {
   "term": "Direct, indirect and spurious effect",
   "definition": "The direct effect of A on Y is transmitted along the edge A → Y; an indirect effect is transmitted through mediators (A → W → Y); a spurious association arises from paths that are not causal from A, for example through a common cause Z of A and Y (A ← Z → Y) (Zhang and Bareinboim, 2018).",
   "chapters": [
    75
   ]
  },
  {
   "term": "Direct mechanism",
   "definition": "Influence of A on the decision along the edge A → D.",
   "chapters": [
    30
   ]
  },
  {
   "term": "Directed acyclic graph (DAG)",
   "definition": "A set of nodes (variables) joined by directed edges with no directed cycles. In a causal DAG an edge V → V' asserts that V is a direct cause of V' relative to the variables in the graph.",
   "chapters": [
    31
   ]
  },
  {
   "term": "Disaggregated evaluation",
   "definition": "An evaluation in which system performance is assessed and reported separately for different groups of people (Barocas et al., 2021). Most group fairness assessments are disaggregated evaluations; in toolkit practice (Chapter 58) it means computing a performance or fairness metric separately for each group defined by one or more sensitive features and comparing the group values.",
   "chapters": [
    26,
    36,
    58
   ]
  },
  {
   "term": "Disaggregation",
   "definition": "Breaking an overall metric into values for demographic groups, intersections, segments (geography, channel, product) and time periods.",
   "chapters": [
    63
   ]
  },
  {
   "term": "Disagreement",
   "definition": "The extent to which explanations produced by different methods for the same model and prediction differ, for example in their top features, ranks or signs (Krishna et al., 2024).",
   "chapters": [
    77
   ]
  },
  {
   "term": "Disagreement problem",
   "definition": "The phenomenon that different explanation methods produce different explanations for the same prediction of the same model (Krishna et al., 2024).",
   "chapters": [
    82
   ]
  },
  {
   "term": "Discovered and reported incidents",
   "definition": "In predictive policing, discovered incidents are those observed by officers sent by the algorithm; reported incidents are those reported by residents, for example through emergency calls. Only discovered incidents depend on where the algorithm sends police (Ensign et al., 2018).",
   "chapters": [
    27
   ]
  },
  {
   "term": "Discrimination (international definition)",
   "definition": "Any distinction, exclusion, restriction or preference based on a protected ground which has the purpose or effect of nullifying or impairing the recognition, enjoyment or exercise, on an equal footing, of human rights and fundamental freedoms. The formula appears in Article 1 of the race convention (CERD), in Article 1 of the women's convention (CEDAW, \"effect or purpose\") and in the Human Rights Committee's General Comment 18.",
   "chapters": [
    84
   ]
  },
  {
   "term": "Disparate impact",
   "definition": "In United States law, liability for a facially neutral practice that causes a disproportionate adverse effect on a protected group, whatever the intent, unless the respondent shows that it is \"job related for the position in question and consistent with business necessity\", or where a less discriminatory alternative practice exists and the respondent refuses to adopt it (42 U.S.C. § 2000e-2(k)(1)(A); Griggs v. Duke Power Co., 1971). The corresponding European concept is indirect discrimination.",
   "chapters": [
    14,
    23,
    83,
    88
   ]
  },
  {
   "term": "Disparate impact ratio (DIR)",
   "definition": "The ratio of a group's selection rate to that of a reference group, usually the most favored group.",
   "chapters": [
    14
   ]
  },
  {
   "term": "Disparate impact removal (repair)",
   "definition": "A pre-processing transformation that changes the values of features so that their distribution no longer differs by group, while keeping the order of individuals within each group (Feldman et al., 2015).",
   "chapters": [
    49
   ]
  },
  {
   "term": "Disparate interactions",
   "definition": "Patterns in which the use of an algorithmic aid by human decision makers shifts their judgments differently for different groups (Green and Chen, 2019).",
   "chapters": [
    28
   ]
  },
  {
   "term": "Disparate treatment",
   "definition": "In United States law, intentional differential treatment of a person because of a protected characteristic. In employment, Title VII makes it unlawful to discriminate against an individual \"because of\" race, color, religion, sex or national origin (42 U.S.C. § 2000e-2(a)(1)), and also when such a characteristic \"was a motivating factor\" even though other factors also motivated the practice (§ 2000e-2(m)). The European counterpart is direct discrimination.",
   "chapters": [
    23,
    83,
    88
   ]
  },
  {
   "term": "Disparity",
   "definition": "The comparison of a metric between a group and the reference, expressed as a difference (in percentage points) or a ratio.",
   "chapters": [
    44
   ]
  },
  {
   "term": "Disparity attribution",
   "definition": "An attribution of a group fairness metric, such as the demographic parity difference, to input features (Begley et al., 2020). In the additive form of Chapter 76, the contributions sum to the metric.",
   "chapters": [
    71,
    76
   ]
  },
  {
   "term": "Disparity growth rate",
   "definition": "The relative change of a disparity measure between successive system iterations (defined formally in Chapter 27).",
   "chapters": [
    27
   ]
  },
  {
   "term": "Distortion",
   "definition": "A measure of how far the transformed output departs from the original, for example the mean squared difference or the share of changed decisions.",
   "chapters": [
    57
   ]
  },
  {
   "term": "Distribution alignment",
   "definition": "A transformation that makes the distribution of outputs the same, or closer, across groups. Applied to scores, full alignment yields demographic parity at every threshold.",
   "chapters": [
    57
   ]
  },
  {
   "term": "Distribution shift (dataset shift)",
   "definition": "A difference between the distribution of (X, A, Y) in production or testing, P_T, and the distribution the model was trained or validated on, P_S. Distribution shifts can substantially degrade the accuracy of machine learning systems deployed in the wild (Koh et al., 2021); the main forms are covariate shift, label shift and concept drift.",
   "chapters": [
    34,
    61
   ]
  },
  {
   "term": "Distribution transformation",
   "definition": "A mapping T applied to the features X, possibly depending on the protected attribute A during fitting, that produces new features X̃ whose distribution depends less, or not at all, on A.",
   "chapters": [
    48
   ]
  },
  {
   "term": "Distributionally robust optimization (DRO)",
   "definition": "Optimization or evaluation of the worst-case expected loss over a set of plausible distributions, the uncertainty set.",
   "chapters": [
    42
   ]
  },
  {
   "term": "Distributive justice",
   "definition": "The part of political philosophy concerned with how benefits and burdens (income, opportunities, positions, risks) should be allocated among people (Lamont and Favor, 2017).",
   "chapters": [
    3
   ]
  },
  {
   "term": "Distrust",
   "definition": "Not merely the absence of trust. Philosophical accounts treat it as an attitude in its own right, for example a tendency to withdraw from reliance or vulnerability in contexts of normative expectation (McLeod, 2020).",
   "chapters": [
    1
   ]
  },
  {
   "term": "Diversity, novelty, serendipity, coverage",
   "definition": "Beyond-accuracy properties of recommendations: how varied a list is, how unfamiliar its items are to the user, how valuable and unexpected they are, and how much of the catalog the system recommends at all (Kaminskas and Bridge, 2016).",
   "chapters": [
    66
   ]
  },
  {
   "term": "Documented case",
   "definition": "A case concerning a real system whose facts are established in a verifiable source: a court or regulator record, a peer-reviewed study, an official statement or an investigative report with published data.",
   "chapters": [
    111
   ]
  },
  {
   "term": "Domain",
   "definition": "The field of application in which a system makes or supports decisions: criminal justice, health care, employment, credit, education, public services, digital platforms, or perception systems.",
   "chapters": [
    111
   ]
  },
  {
   "term": "Domain generalization (domain shift)",
   "definition": "Training and test data come from related but distinct domains, such as different hospitals or regions (Koh et al., 2021).",
   "chapters": [
    34
   ]
  },
  {
   "term": "E-value",
   "definition": "The minimum strength of association, on the risk ratio scale, that an unmeasured confounder would need to have with both the exposure and the outcome to fully explain away an observed association, conditional on the measured covariates (VanderWeele and Ding, 2017).",
   "chapters": [
    33
   ]
  },
  {
   "term": "Ecosystem-level analysis",
   "definition": "The analysis of the collection of models deployed in a context rather than of a single model, so that outcomes for individuals across all systems they encounter become visible (Toups et al., 2023; Chapter 114).",
   "chapters": [
    115
   ]
  },
  {
   "term": "Effect size",
   "definition": "A measure of the magnitude of a difference or disparity, on an interpretable scale, that does not grow with sample size, such as a difference or ratio of rates, an odds ratio or a standardized mean difference.",
   "chapters": [
    38,
    41
   ]
  },
  {
   "term": "Effective challenge",
   "definition": "Critical analysis of models by objective, informed parties who can identify limitations and assumptions and produce appropriate changes; the same guidance says it depends on a combination of incentives, competence and influence, and that incentives are stronger when the challenge is separated from model development (Board of Governors of the Federal Reserve System and Office of the Comptroller of the Currency, 2011).",
   "chapters": [
    81
   ]
  },
  {
   "term": "Effective sample size (ESS)",
   "definition": "The number of equally weighted observations that would carry about as much information as a weighted sample. For weights w_1, ..., w_n the common approximation is (∑ w_i)² / ∑ w_i².",
   "chapters": [
    47
   ]
  },
  {
   "term": "Effects test",
   "definition": "The term Regulation B used for disparate impact liability under ECOA, removed by the CFPB's final rule of April 2026.",
   "chapters": [
    88
   ]
  },
  {
   "term": "Egalitarianism",
   "definition": "The family of views holding that people are in a fundamental sense equal and that certain inequalities should be avoided or corrected; what exactly should be equal (welfare, resources, capabilities, opportunities, or standing) is itself contested (Binns, 2018; Gosepath, 2021). Outcome egalitarianism (Chapter 4) holds that it is in itself bad if some people are worse off than others, and that outcomes should therefore be equalized, in some cases even at a cost in total well-being (Adler and Holtug, 2025).",
   "chapters": [
    3,
    4
   ]
  },
  {
   "term": "Emergent bias",
   "definition": "Bias that arises in a context of use rather than in design, for example when a system built for one population is deployed in another. It is one of three categories identified by Friedman and Nissenbaum (1996), alongside preexisting bias (rooted in social institutions, practices and attitudes) and technical bias (arising from technical constraints or considerations).",
   "chapters": [
    8,
    28
   ]
  },
  {
   "term": "Emerging discrimination",
   "definition": "In this compendium, a discriminatory effect of an AI system that was absent or undetectable at assessment and arises after deployment from changes in the population, the context of use, the system's own influence, or the groups the system implicitly creates.",
   "chapters": [
    34
   ]
  },
  {
   "term": "Empirical risk minimization (ERM)",
   "definition": "The standard learning principle: choose model parameters θ that minimize the average loss over the training examples. It is the default objective of most machine learning methods.",
   "chapters": [
    26,
    52
   ]
  },
  {
   "term": "Encoded social category",
   "definition": "A classification of people (sex, race, disease, creditworthiness, risk) that is built into a technology's data structures and that reflects the historical and political context in which it was created rather than a natural or fixed distinction.",
   "chapters": [
    6
   ]
  },
  {
   "term": "Encoder and fusion",
   "definition": "An encoder maps one modality into a representation. Fusion combines representations, early (by concatenating inputs or features), late (by combining per-modality predictions) or through attention mechanisms that let one modality condition on another.",
   "chapters": [
    68
   ]
  },
  {
   "term": "Endowment and coefficient effects",
   "definition": "In the Oaxaca-Blinder decomposition, the part of a gap between group means explained by differences in the groups' average characteristics (endowments) and the part due to differences in how characteristics are rewarded (coefficients), the latter often called the unexplained part.",
   "chapters": [
    76
   ]
  },
  {
   "term": "Epistemic and substantial facets",
   "definition": "Deck et al. (2024), following earlier work, separate the use of explanation to learn about unfairness (detecting it and analyzing its sources: the epistemic facet) from its use to change outcomes (mitigation: the substantial facet).",
   "chapters": [
    71
   ]
  },
  {
   "term": "Equal opportunity",
   "definition": "Equal TPR across groups (a relaxation of equalized odds).",
   "chapters": [
    15
   ]
  },
  {
   "term": "Equal opportunity difference",
   "definition": "The difference in true positive rates, TPR_a − TPR_b (Chapter 15).",
   "chapters": [
    103
   ]
  },
  {
   "term": "Equality of opportunity",
   "definition": "The ideal that access to desirable positions should not depend on arbitrary factors. Its formal version requires that positions be open to all with the relevant qualifications; its substantive version also requires fair access to the qualifications themselves (Heidari et al., 2019).",
   "chapters": [
    5
   ]
  },
  {
   "term": "Equalized odds",
   "definition": "Equal TPR and equal FPR across groups (separation for a binary prediction).",
   "chapters": [
    15
   ]
  },
  {
   "term": "Equity",
   "definition": "In health, the absence of systematic disparities between groups with different levels of underlying social advantage (Braveman and Gruskin, 2003); more generally, attention to differences that are unnecessary, avoidable and unfair (Whitehead, 1992).",
   "chapters": [
    5
   ]
  },
  {
   "term": "Error rate balance",
   "definition": "FPR_a = FPR_b and FNR_a = FNR_b. For a binary classifier this is the same as equalized odds (Chapter 15).",
   "chapters": [
    19
   ]
  },
  {
   "term": "Escalation procedure",
   "definition": "A defined process for raising a fairness concern, classifying it by severity, routing it to the right authority, and resolving it within a stated time.",
   "chapters": [
    108
   ]
  },
  {
   "term": "Essentially contested concept",
   "definition": "A concept with multiple context-dependent, and sometimes conflicting, theoretical understandings. Jacobs and Wallach (2021) argue that fairness is one.",
   "chapters": [
    45
   ]
  },
  {
   "term": "Estimand and estimate",
   "definition": "The estimand is the population quantity a metric is meant to describe, for example the difference in approval rates between groups a and b among all applicants the model will score. The estimate is the value computed from the sample. The hat notation distinguishes them: p_a is a group's population rate and p̂_a its sample estimate.",
   "chapters": [
    41
   ]
  },
  {
   "term": "Estimand, estimator, estimate",
   "definition": "The estimand is the population quantity of interest (for example the true positive rate of group b in the population scored by the system); the estimator is the rule used to compute it from a sample (for example the sample proportion); the estimate is the number that results.",
   "chapters": [
    13
   ]
  },
  {
   "term": "Ethical framework for AI",
   "definition": "A set of principles or requirements intended to guide the development and use of AI, such as the Ethics Guidelines for Trustworthy AI presented in 2019 by the European Commission's High-Level Expert Group (European Commission, n.d.) or the UNESCO Recommendation (UNESCO, 2021; Chapter 89).",
   "chapters": [
    92
   ]
  },
  {
   "term": "Ethics owners",
   "definition": "Staff whose role is to hold responsibility for ethics across an organization's divisions; Metcalf, Moss and boyd (2019) use the term for the ethics roles that technology companies created in the late 2010s.",
   "chapters": [
    92
   ]
  },
  {
   "term": "Ethics washing",
   "definition": "The use of ethical language, principles or bodies to give the appearance of responsible AI without changing practice or accepting accountability. Yeung, Howes and Pogrebna (2020) propose human rights-centered design, deliberation and oversight as a way to end it.",
   "chapters": [
    92
   ]
  },
  {
   "term": "Evaluation",
   "definition": "The assessment of measured results against agreed benchmarks, for example a tolerance on a metric or a legal threshold (DSIT, 2024).",
   "chapters": [
    93
   ]
  },
  {
   "term": "Evaluation bias",
   "definition": "Bias that arises when the benchmark data used for a task do not represent the use population, so that models are optimized and selected for the represented subset (Suresh and Guttag, 2021).",
   "chapters": [
    26
   ]
  },
  {
   "term": "Evaluation protocol",
   "definition": "The design of model testing: test set composition, chosen metrics, significance tests, and whether results are reported in aggregate or by slice.",
   "chapters": [
    26
   ]
  },
  {
   "term": "Evaluation question",
   "definition": "The precise question the assessment answers, including the object, the population, the period, the groups compared and the criterion applied.",
   "chapters": [
    36
   ]
  },
  {
   "term": "Evidence",
   "definition": "Material used as the basis of the justification of a claim; its sources include the design, the development process, prior field experience, testing, source code analysis and formal analysis (Bloomfield and Netkachova, 2014). In compliance (Chapter 95), it is any artifact that shows that a requirement has been met or that a claim holds, such as test reports, data audits, logs, monitoring records and approvals. In GSN, evidence is cited by a solution, which makes no claim of its own but is a reference to an evidence item (ACWG, 2021).",
   "chapters": [
    93,
    94,
    95
   ]
  },
  {
   "term": "Evidence lifecycle management",
   "definition": "The systematic management of evidence across planning, generation, validation, storage, updating and retention, so that evidence remains complete and current for the whole life of an AI system.",
   "chapters": [
    95
   ]
  },
  {
   "term": "Evidence of record",
   "definition": "The version of an explanation or test result that is retained, versioned and can be produced to an auditor, authority or court, as opposed to an interactive visualization that cannot be reproduced.",
   "chapters": [
    81
   ]
  },
  {
   "term": "Evidence tagging",
   "definition": "Metadata attached to an artifact that records the provisions it relates to, its owner, version, date, validation status and retention period.",
   "chapters": [
    95
   ]
  },
  {
   "term": "Ex ante and ex post information",
   "definition": "Information given before processing or a decision (for example at data collection under Articles 13 and 14) and information given after a decision has been taken (for example on an access request under Article 15).",
   "chapters": [
    79
   ]
  },
  {
   "term": "Exchangeability",
   "definition": "The modeling assumption, behind partial pooling, that before seeing the data the subgroups' true rates are regarded as draws from a common distribution.",
   "chapters": [
    43
   ]
  },
  {
   "term": "Expected calibration error (ECE)",
   "definition": "The average absolute gap between mean score and observed rate over bins, weighted by bin size. Maximum calibration error (MCE) is the largest such gap. Both are attributed to Naeini, Cooper and Hauskrecht (2015) by Guo et al. (2017).",
   "chapters": [
    16,
    56
   ]
  },
  {
   "term": "Experiment tracking",
   "definition": "The systematic recording of each model training run: its code version, data, parameters, metrics and artifacts, in a platform such as MLflow or Weights & Biases.",
   "chapters": [
    59
   ]
  },
  {
   "term": "Expiry",
   "definition": "The end of a validity period. Expiry is decided at issuance and needs no further action by the issuer.",
   "chapters": [
    99
   ]
  },
  {
   "term": "Explainability",
   "definition": "A representation of the mechanisms underlying an AI system's operation (National Institute of Standards and Technology, 2023). Barredo Arrieta et al. (2020) define explainable AI relative to an audience, as one that produces details or reasons to make its functioning clear or easy to understand. The compendium uses \"explainability\" for the capacity of a system, together with its explanation methods, to give such reasons, and \"interpretability\" for the degree to which a given audience can understand them.",
   "chapters": [
    10,
    70
   ]
  },
  {
   "term": "Explainability need",
   "definition": "A question a user wants answered about an AI system, its data, its outputs or a specific prediction, together with the goal behind it (Liao, Gruen and Miller, 2020).",
   "chapters": [
    78
   ]
  },
  {
   "term": "Explanation",
   "definition": "In the sense used in Part VIII, an answer to a question about why a system produced an output or how it works. Miller (2019) adopts the definition of an explanation as an answer to a why-question and distinguishes three aspects: a cognitive process (identifying causes of an event and selecting some of them), a product (the explanation that results) and a social process (transferring that knowledge from an explainer to an explainee). The event to be explained is the explanandum; the material that explains it is the explanans.",
   "chapters": [
    70
   ]
  },
  {
   "term": "Explanation accuracy",
   "definition": "The principle that an explanation correctly reflects the reason for generating the output or accurately reflects the system's process (Phillips et al., 2021). The machine learning literature calls this fidelity or faithfulness (Chapter 77). It is distinct from decision accuracy, which concerns whether the system's output is correct.",
   "chapters": [
    81
   ]
  },
  {
   "term": "Explanation accuracy and decision accuracy",
   "definition": "Phillips et al. (2021) separate the accuracy of an explanation (whether it correctly reflects the reason for the output or the system's process) from decision accuracy (whether the system's judgment is correct). A correct decision can come with an inaccurate explanation, and an incorrect decision with an accurate one.",
   "chapters": [
    70
   ]
  },
  {
   "term": "Explanation attack",
   "definition": "A modification of a model, or of how it responds to the inputs an explanation method generates, that changes its explanations while leaving its behavior on real data largely unchanged (Slack et al., 2020; Dimanov et al., 2020; Anders et al., 2020).",
   "chapters": [
    71
   ]
  },
  {
   "term": "Explanation fidelity",
   "definition": "The degree to which an explanation model approximates the behavior of the model it explains; Balagopalan et al. (2022) measure it as the approximation quality of an explanation model.",
   "chapters": [
    10
   ]
  },
  {
   "term": "Explanation quality disparity",
   "definition": "A difference between groups in any explanation quality property, such as fidelity, stability, consistency or sparsity (Dai et al., 2022).",
   "chapters": [
    82
   ]
  },
  {
   "term": "Explanatory evidence",
   "definition": "Evidence produced by an explanation method (attributions, counterfactuals, surrogate models, rule extractions, reason codes) or about explanations (tests of their accuracy, user studies, logs of explanations given).",
   "chapters": [
    81
   ]
  },
  {
   "term": "Explicability",
   "definition": "In the High-Level Expert Group guidelines, the principle that processes are transparent, capabilities and purpose openly communicated and decisions, to the extent possible, explainable to those directly and indirectly affected (High-Level Expert Group on Artificial Intelligence, 2019). Floridi and Cowls (2019) give it an epistemological sense (intelligibility: how does it work?) and an ethical sense (accountability: who is responsible for the way it works?).",
   "chapters": [
    10
   ]
  },
  {
   "term": "Explicand",
   "definition": "The input x whose prediction is explained (Sundararajan and Najmi, 2020). In local explanation it is one decision subject; in global explanation the explicands are a whole data set.",
   "chapters": [
    73
   ]
  },
  {
   "term": "Exposure",
   "definition": "The expected attention an item, person or provider receives from its positions in rankings, under a model of how attention declines with rank, measured by a position weight that decreases with rank (Singh and Joachims, 2018).",
   "chapters": [
    21,
    40,
    66
   ]
  },
  {
   "term": "Factorial design",
   "definition": "A design that crosses two or more factors (for example, intervention on or off and user engagement high or low) to estimate their interaction.",
   "chapters": [
    64
   ]
  },
  {
   "term": "Factual and counterfactual instance",
   "definition": "The factual instance x is the person's actual feature vector; the counterfactual instance x′ is the hypothetical vector that receives the desired decision.",
   "chapters": [
    74
   ]
  },
  {
   "term": "FADP",
   "definition": "The Federal Act on Data Protection of 25 September 2020 (SR 235.1), in force since 1 September 2023. It replaced the act of 1992 and aims to protect the personality and fundamental rights of natural persons whose personal data are processed (Article 1).",
   "chapters": [
    87
   ]
  },
  {
   "term": "FAIR",
   "definition": "The validant.ai structure for fairness acceptance criteria (Fairness metric thresholds, Auditing requirements, Intersectional analysis, Reporting guidelines). It is unrelated to the FAIR principles for scientific data management (Findable, Accessible, Interoperable, Reusable) of Wilkinson et al. (2016); the two should not be confused.",
   "chapters": [
    103
   ]
  },
  {
   "term": "Fair AI Scrum",
   "definition": "The approach of Chapter 102, in which the standard Scrum roles, events and artifacts are extended with fairness dimensions rather than supplemented by a parallel fairness process.",
   "chapters": [
    102
   ]
  },
  {
   "term": "Fair re-ranking",
   "definition": "Post-processing of a ranked list so that it satisfies a group fairness constraint on positions or exposure.",
   "chapters": [
    57
   ]
  },
  {
   "term": "Fair representation",
   "definition": "A learned transformation Z = g(X) of the features into a new space, possibly using A during training, that keeps the information needed for the task while making the protected attribute hard to recover from Z (Zemel et al., 2013).",
   "chapters": [
    48,
    49
   ]
  },
  {
   "term": "Fair representation learning",
   "definition": "Learning an encoder whose representations retain information useful for the task while removing, or limiting, information about A (Zemel et al., 2013).",
   "chapters": [
    65
   ]
  },
  {
   "term": "Fair, unfair and contested paths",
   "definition": "A classification of causal paths, made with domain experts, affected stakeholders and legal advice, into routes whose influence on the prediction is accepted, routes whose influence is to be removed, and routes on which stakeholders disagree.",
   "chapters": [
    32
   ]
  },
  {
   "term": "Fairness",
   "definition": "In algorithmic decision-making, the property of a decision system, its outputs or its effects of treating people in a way that can be justified to them. The term functions as a placeholder for several distinct normative ideals, most of them egalitarian in character (Binns, 2018).",
   "chapters": [
    3
   ]
  },
  {
   "term": "Fairness acceptance criterion",
   "definition": "A measurable condition on fairness that a feature must satisfy before acceptance.",
   "chapters": [
    103
   ]
  },
  {
   "term": "Fairness-accuracy frontier (Pareto frontier)",
   "definition": "The set of models for which fairness cannot be improved without losing accuracy, and accuracy cannot be improved without losing fairness. Varying λ traces an approximation of it (Chapter 54).",
   "chapters": [
    52
   ]
  },
  {
   "term": "Fairness assessment",
   "definition": "A structured empirical study that establishes, with stated uncertainty, whether a specified system treats specified groups of people in a way that meets specified fairness criteria, and documents the reasoning behind each choice.",
   "chapters": [
    36
   ]
  },
  {
   "term": "Fairness audit",
   "definition": "A structured, documented examination of an AI system, its data and its use context, which determines whether the system's outcomes meet fairness criteria selected for that context and identifies the actions needed where they do not.",
   "chapters": [
    110
   ]
  },
  {
   "term": "Fairness Audit Playbook",
   "definition": "The standardized workflow of Chapter 110, consisting of eight phases with defined inputs, outputs and gate criteria, which integrates the methods of Chapters 11, 22, 35 and 44.",
   "chapters": [
    110
   ]
  },
  {
   "term": "Fairness-aware backlog ordering",
   "definition": "Ordering of Product Backlog items that adds fairness dimensions (impact, bias risk, harm severity, regulatory exposure) to business value, dependency and risk.",
   "chapters": [
    104
   ]
  },
  {
   "term": "Fairness-aware generation",
   "definition": "Generation in which the objective or the sampling process includes a fairness requirement, such as independence between generated features and the protected attribute A.",
   "chapters": [
    50
   ]
  },
  {
   "term": "Fairness-aware loss function",
   "definition": "A training objective that combines a prediction loss with one or more terms measuring unfairness, so that minimizing it rewards both accuracy and reduced disparity.",
   "chapters": [
    52
   ]
  },
  {
   "term": "Fairness baseline",
   "definition": "The values of the chosen fairness metrics, with their uncertainty, computed on the data (or on a reference model trained on the unmodified data) before any intervention.",
   "chapters": [
    37
   ]
  },
  {
   "term": "Fairness blocker",
   "definition": "An impediment that prevents a fairness task from progressing, for example missing group data or an unresolved choice of fairness definition.",
   "chapters": [
    104
   ]
  },
  {
   "term": "Fairness boundary statement",
   "definition": "A statement of the conditions under which a system's fairness properties have been validated and of those under which they have not.",
   "chapters": [
    107
   ]
  },
  {
   "term": "Fairness budget",
   "definition": "A pre-agreed tolerance that allows a metric to fall below its target for a short, bounded period (for example, during a surge in demand) provided that the deviation is recorded and compensated within a defined horizon.",
   "chapters": [
    61
   ]
  },
  {
   "term": "Fairness capacity allocation",
   "definition": "The share of Sprint capacity explicitly reserved for fairness analysis, implementation and validation.",
   "chapters": [
    104
   ]
  },
  {
   "term": "Fairness champion",
   "definition": "A member of a product or business team designated to carry fairness practice into that team and to connect it to central expertise.",
   "chapters": [
    106
   ]
  },
  {
   "term": "Fairness checkpoint",
   "definition": "A fixed point in a Scrum event at which fairness risks, blockers or results are explicitly discussed and recorded.",
   "chapters": [
    102
   ]
  },
  {
   "term": "Fairness circuit breaker",
   "definition": "An automatic safeguard that, on a severe fairness violation, switches the system to a predefined safer state, such as routing traffic to a baseline model or sending decisions to human review.",
   "chapters": [
    61
   ]
  },
  {
   "term": "Fairness constraint",
   "definition": "A condition that a solution must satisfy, for example that the difference in selection rates between groups is at most ε. Solutions that violate it are excluded (Chapter 51).",
   "chapters": [
    52
   ]
  },
  {
   "term": "Fairness council",
   "definition": "A dedicated governance body with an explicit fairness mandate, defined membership, operating model, decision authority, documentation requirements and performance measures.",
   "chapters": [
    108
   ]
  },
  {
   "term": "Fairness criterion (fairness definition)",
   "definition": "A formally stated condition on the joint distribution of predictions, outcomes and protected attributes, such as demographic parity or equal opportunity.",
   "chapters": [
    22
   ]
  },
  {
   "term": "Fairness dashboard",
   "definition": "A visual interface that presents fairness metrics for one or more AI systems, disaggregated by group and over time, with the context needed to interpret them and links to the actions they should trigger.",
   "chapters": [
    63
   ]
  },
  {
   "term": "Fairness debt",
   "definition": "Known fairness work that has been deferred, recorded explicitly so that it is not lost.",
   "chapters": [
    104
   ]
  },
  {
   "term": "Fairness decision framework",
   "definition": "A structured approach that determines who makes which fairness decisions under which conditions, with categories, authority mapping, required inputs, review criteria and approval thresholds.",
   "chapters": [
    108
   ]
  },
  {
   "term": "Fairness Decision Record (FDR)",
   "definition": "A structured record of one fairness decision, with its context, the decision, the alternatives considered, the rationale, the stakeholders, the trade-offs, the success metrics, the limitations and the supporting references.",
   "chapters": [
    107
   ]
  },
  {
   "term": "Fairness desideratum",
   "definition": "A specific fairness goal that an explanation is meant to serve, for example reporting a formal criterion such as demographic parity, identifying the source of a disparity, or enabling an affected person to contest a decision (Deck et al., 2024).",
   "chapters": [
    71
   ]
  },
  {
   "term": "Fairness dimension",
   "definition": "The part of a requirement, task or check that names the affected groups, the fairness goal and the way it is verified.",
   "chapters": [
    102
   ]
  },
  {
   "term": "Fairness documentation framework",
   "definition": "The set of templates, processes and responsibilities through which an organization records fairness information across the lifecycle.",
   "chapters": [
    107
   ]
  },
  {
   "term": "Fairness drift",
   "definition": "A persistent change in a fairness metric, or in the group-conditional distributions that determine it, beyond the range explained by sampling variation and known cycles.",
   "chapters": [
    62
   ]
  },
  {
   "term": "Fairness-enhanced Sprint Review",
   "definition": "A Sprint Review that presents fairness metrics, disaggregated performance, mitigation results and remaining fairness debt alongside functional results.",
   "chapters": [
    105
   ]
  },
  {
   "term": "Fairness-enhanced user story",
   "definition": "A story extended with a fairness goal, the groups it applies to and, where possible, the fairness definition.",
   "chapters": [
    103
   ]
  },
  {
   "term": "Fairness gap",
   "definition": "A measure of the violation of a fairness criterion, for example the absolute difference in selection rates between groups.",
   "chapters": [
    20
   ]
  },
  {
   "term": "Fairness gerrymandering",
   "definition": "The situation, named by Kearns, Neel, Roth and Wu (2018), in which a classifier appears fair on each of a few predefined groups but badly violates the fairness criterion on structured subgroups, such as combinations of protected attribute values.",
   "chapters": [
    26,
    29,
    42,
    52,
    76
   ]
  },
  {
   "term": "Fairness governance body",
   "definition": "A standing forum with a written fairness mandate, such as a steering committee, a review board, a working group, a community advisory council or a technical committee.",
   "chapters": [
    106
   ]
  },
  {
   "term": "Fairness health score",
   "definition": "A composite indicator that rolls several fairness metrics into one number or traffic-light status for executive views. It is a summary device, not a fairness criterion.",
   "chapters": [
    63
   ]
  },
  {
   "term": "Fairness Impact section",
   "definition": "A part of a pull request description in which the author states how the change may affect different groups and links the automated fairness report.",
   "chapters": [
    59
   ]
  },
  {
   "term": "Fairness improvement cycle",
   "definition": "A deliberate learning loop that assesses fairness capability, sets improvement goals and tests new practices.",
   "chapters": [
    105
   ]
  },
  {
   "term": "Fairness incident",
   "definition": "In this compendium, an AI incident whose harm consists of, or includes, unjustified differential treatment or outcomes for people defined by a protected attribute or another socially salient characteristic (Chapter 24). It includes allocative harms (unequal access to resources or opportunities), quality-of-service harms (a system that works worse for some groups) and representational harms (stereotyping or demeaning portrayals), three of the five categories of the sociotechnical harm taxonomy of Shelby et al. (2023).",
   "chapters": [
    109
   ]
  },
  {
   "term": "Fairness intervention",
   "definition": "A deliberate change to a decision system made to reduce a measured violation of a chosen fairness criterion. Also called bias mitigation.",
   "chapters": [
    46
   ]
  },
  {
   "term": "Fairness leadership role",
   "definition": "A position with an explicit fairness mandate, decision rights and resources, as opposed to fairness as an informal concern added to another job.",
   "chapters": [
    106
   ]
  },
  {
   "term": "Fairness library (fairness toolkit)",
   "definition": "A software package that computes fairness metrics, implements bias mitigation methods, or both, for machine learning models and data. The literature uses \"toolkit\" and \"library\" interchangeably (Lee and Singh, 2021).",
   "chapters": [
    58
   ]
  },
  {
   "term": "Fairness maturity assessment",
   "definition": "A periodic structured self-assessment of a team's fairness capabilities.",
   "chapters": [
    105
   ]
  },
  {
   "term": "Fairness measurement",
   "definition": "The computation of a fairness metric (for example a selection rate ratio, a true positive rate difference or a calibration gap) from decision records or test data, together with its statistical uncertainty (Chapters 39 to 43).",
   "chapters": [
    93
   ]
  },
  {
   "term": "Fairness metric",
   "definition": "A function of a system's outputs, the true outcomes and group membership that quantifies the degree to which a fairness criterion holds, for example the difference in true positive rates between two groups.",
   "chapters": [
    44
   ]
  },
  {
   "term": "Fairness regression",
   "definition": "A change to code, data or configuration that makes a fairness metric worse than it was on the reference version (usually the main branch), beyond an agreed tolerance.",
   "chapters": [
    59
   ]
  },
  {
   "term": "Fairness regularizer (fairness penalty)",
   "definition": "A non-negative function Ω(θ) that is zero, or small, when the model meets a fairness criterion and grows with the size of the violation. It is added to the prediction loss rather than imposed as a constraint.",
   "chapters": [
    52
   ]
  },
  {
   "term": "Fairness requirement",
   "definition": "A specific and testable statement of a fairness objective: a criterion, the protected attributes it applies to, a metric, a threshold and a test.",
   "chapters": [
    107
   ]
  },
  {
   "term": "Fairness responsibility",
   "definition": "A defined task in the fairness lifecycle (for example selecting a fairness criterion, setting a metric threshold, approving a mitigation) assigned to a specific role.",
   "chapters": [
    106
   ]
  },
  {
   "term": "Fairness retrospective technique",
   "definition": "A prompt, exercise or framework that extracts fairness lessons from a retrospective.",
   "chapters": [
    105
   ]
  },
  {
   "term": "Fairness task taxonomy",
   "definition": "A catalog of recurring fairness task types with completion criteria and reference estimates.",
   "chapters": [
    104
   ]
  },
  {
   "term": "Fairness test",
   "definition": "An automated test, written in a standard testing framework, that computes a fairness metric for a model or pipeline on reference data and asserts that it lies within an agreed threshold.",
   "chapters": [
    59
   ]
  },
  {
   "term": "Fairness theater",
   "definition": "Performing fairness activities mechanically, so that the process is followed but no fairness problem is found or fixed.",
   "chapters": [
    102
   ]
  },
  {
   "term": "Fairness through awareness",
   "definition": "The approach of Dwork et al. (2012), in which fairness is defined relative to an explicit, publicly discussable similarity metric that may take protected attributes and their social context into account.",
   "chapters": [
    17
   ]
  },
  {
   "term": "Fairness through unawareness",
   "definition": "The approach of excluding protected attributes from a model's inputs and treating this as sufficient for fairness (Kusner et al., 2017). Proxy variables are the main reason it fails.",
   "chapters": [
    17,
    24,
    38
   ]
  },
  {
   "term": "Fairness timeline",
   "definition": "A retrospective exercise that maps when each bias issue was introduced, when it was detected and why.",
   "chapters": [
    105
   ]
  },
  {
   "term": "Fairness transfer",
   "definition": "The extent to which a fairness property established on P_S continues to hold on P_T (Schrouff et al., 2022).",
   "chapters": [
    34
   ]
  },
  {
   "term": "Fairness-utility frontier",
   "definition": "The set of decision rules for which neither fairness (measured by a chosen disparity) nor utility can be improved without worsening the other. It is the Pareto frontier of the two objectives (Chapter 54).",
   "chapters": [
    55
   ]
  },
  {
   "term": "Fairness velocity and acceleration",
   "definition": "The first and second differences (rate of change, and change of that rate) of a fairness metric tracked over time.",
   "chapters": [
    61
   ]
  },
  {
   "term": "Fairwashing",
   "definition": "Promoting the false perception that a machine learning model respects some ethical values, for example by producing a fair-looking explanation of an unfair model (Aïvodji et al., 2019).",
   "chapters": [
    10,
    71,
    82
   ]
  },
  {
   "term": "Faithfulness",
   "definition": "The degree to which an explanation accurately represents the reasoning process behind a model's prediction (Jacovi and Goldberg, 2020). Turpin et al. (2023) call an explanation unfaithful when it misrepresents the true reason for a prediction, for example when a feature of the input changes the prediction but the explanation does not mention it.",
   "chapters": [
    80
   ]
  },
  {
   "term": "Faithfulness (fidelity, correctness)",
   "definition": "The degree to which an explanation accurately reflects the reasoning process of the model it explains (Jacovi and Goldberg, 2020). For a surrogate model, fidelity is usually measured as agreement with the model's predictions (Chapter 70).",
   "chapters": [
    77
   ]
  },
  {
   "term": "False discovery rate (FDR)",
   "definition": "The expected proportion of false positives among the comparisons declared significant (Benjamini and Hochberg, 1995). For a classifier (Chapter 16), the false discovery rate is 1 − PPV and the false omission rate (FOR) is 1 − NPV.",
   "chapters": [
    16,
    41,
    43
   ]
  },
  {
   "term": "False negative rate (FNR)",
   "definition": "Of group a: FNR_a = P(Ŷ = 0 | Y = 1, A = a) = 1 − TPR_a.",
   "chapters": [
    15,
    19
   ]
  },
  {
   "term": "False positive rate (FPR)",
   "definition": "Of group a: FPR_a = P(Ŷ = 1 | Y = 0, A = a).",
   "chapters": [
    15,
    19
   ]
  },
  {
   "term": "Family-wise error rate (FWER)",
   "definition": "The probability of at least one Type I error among all tests in an analysis.",
   "chapters": [
    41,
    43
   ]
  },
  {
   "term": "FDPIC",
   "definition": "The Federal Data Protection and Information Commissioner, the federal supervisory authority for data protection, which investigates possible violations and can order administrative measures (Articles 49 and 51 FADP).",
   "chapters": [
    87
   ]
  },
  {
   "term": "Feasible region",
   "definition": "The set of parameter values that satisfy all constraints. A problem is infeasible when this set is empty.",
   "chapters": [
    51
   ]
  },
  {
   "term": "Feature attribution",
   "definition": "An explanation that assigns to each input feature a number expressing its contribution to a particular prediction, typically so that the numbers sum to the prediction minus a baseline. More generally (Chapter 73), a vector of numbers φ_1, ..., φ_M, one per feature, that distributes a quantity (a prediction, a loss, a fairness gap) among the M features.",
   "chapters": [
    72,
    73
   ]
  },
  {
   "term": "Feature construction for fairness",
   "definition": "The deliberate design of new variables that measure the legitimate quantity of interest more directly than an existing proxy-laden feature.",
   "chapters": [
    49
   ]
  },
  {
   "term": "Feature effect",
   "definition": "A description of how the model's prediction changes as a feature changes, such as a partial dependence or accumulated local effects curve. Feature importance ranks features by how much the model relies on them, with importance defined differently by each method (Molnar, 2025).",
   "chapters": [
    72
   ]
  },
  {
   "term": "Feature representation and encoding bias",
   "definition": "Bias introduced by the way measurements are transformed, normalized, binned, encoded or imputed in the final dataset. It concerns how a measurement is represented rather than what is measured.",
   "chapters": [
    25
   ]
  },
  {
   "term": "Features (X)",
   "definition": "The information about a decision subject that a model may use, for example income, credit history and debt-to-income ratio in lending. X may be a vector of many variables.",
   "chapters": [
    12
   ]
  },
  {
   "term": "Federal Constitution (BV/Cst.)",
   "definition": "The Federal Constitution of the Swiss Confederation of 18 April 1999 (SR 101). Its catalog of fundamental rights includes equality before the law and the prohibition of discrimination (Article 8).",
   "chapters": [
    87
   ]
  },
  {
   "term": "Federation",
   "definition": "Conveying the result of authentication and identity attributes from an identity provider to a relying party through a federation protocol.",
   "chapters": [
    98
   ]
  },
  {
   "term": "Feedback loop",
   "definition": "A process in which a system's outputs influence the data on which it is later trained or evaluated: which cases are observed, how they are labeled, how people behave or which data are generated. A runaway feedback loop drives the system toward an extreme regardless of the underlying truth (Ensign et al., 2018); in recommendation systems, recommendations shape user interactions, which are logged as training data for the next model (Mansoury et al., 2020).",
   "chapters": [
    8,
    27,
    66,
    114
   ]
  },
  {
   "term": "Fidelity",
   "definition": "How closely synthetic data reproduce the statistical properties of the real data that should be kept.",
   "chapters": [
    50
   ]
  },
  {
   "term": "Fidelity (faithfulness)",
   "definition": "The degree to which an explanation correctly reflects the behavior of the model it explains. LIME requires local fidelity: the explanation must correspond to how the model behaves in the vicinity of the instance being explained (Ribeiro, Singh and Guestrin, 2016).",
   "chapters": [
    70
   ]
  },
  {
   "term": "Fidelity gap",
   "definition": "A difference between groups in the fidelity of an explanation, that is, in how well the explanation imitates the model for members of each group (Balagopalan et al., 2022).",
   "chapters": [
    71,
    82
   ]
  },
  {
   "term": "Finding",
   "definition": "A documented conclusion about one metric for one comparison, stating the estimate, its interval, its denominator and its classification against the tolerance.",
   "chapters": [
    36
   ]
  },
  {
   "term": "First-party, second-party and third-party activity",
   "definition": "A conformity assessment activity is first-party when it is performed by the person or organization that provides or that is the object of conformity assessment; second-party when performed by a person or organization with a user interest in the object, such as a purchaser or a regulator implementing consumer protection legislation; and third-party when performed by a person or organization independent of the provider of the object and with no user interest in it (ISO/IEC, 2020). The labels describe roles in the assessment, not parties to a contract.",
   "chapters": [
    93
   ]
  },
  {
   "term": "First-, second- and third-party assessment",
   "definition": "In the usage adopted in Chapter 96, a first-party assessment is performed by the organization that provides the object assessed, a second-party assessment by a party with a user interest in it (such as a customer or a deployer assessing a vendor's tool), and a third-party assessment by a body independent of both. Costanza-Chock et al. (2022) use the same three categories for AI audits.",
   "chapters": [
    96
   ]
  },
  {
   "term": "Flaky test",
   "definition": "A test whose outcome changes between runs without a relevant change in the system, in the fairness tests of Chapter 60 because a fairness metric estimated from a finite sample fluctuates around its threshold.",
   "chapters": [
    60
   ]
  },
  {
   "term": "Flipset",
   "definition": "A list of actions, each of which would flip a person's prediction to the desired outcome (Ustun et al., 2019).",
   "chapters": [
    74
   ]
  },
  {
   "term": "Formal and substantive algorithmic fairness",
   "definition": "Green (2022) calls \"formal algorithmic fairness\" the methodology that restricts analysis to isolated decision procedures, and proposes \"substantive algorithmic fairness\", which takes a wider scope of analysis drawn from theories of substantive equality and asks whether and how algorithms can promote justice in practice.",
   "chapters": [
    115
   ]
  },
  {
   "term": "Formal equality",
   "definition": "The principle that when two persons are equal in a normatively relevant respect, they must be treated equally in that respect; Aristotle's \"treat like cases as like\" (Gosepath, 2021).",
   "chapters": [
    5
   ]
  },
  {
   "term": "Formal fairness",
   "definition": "Fairness as measured by a mathematical criterion of Part II (independence, separation, sufficiency, individual or counterfactual fairness). Deck et al. (2024) distinguish it from perceived fairness, the judgments and feelings people form about a system.",
   "chapters": [
    71
   ]
  },
  {
   "term": "Foundation model",
   "definition": "A model trained on broad data at scale that can be adapted to a wide range of downstream tasks (Bommasani et al., 2021). In EU law the corresponding legal category is the general-purpose AI model: a model that \"displays significant generality and is capable of competently performing a wide range of distinct tasks\" and can be integrated into downstream systems (Regulation (EU) 2024/1689, Article 3(63)).",
   "chapters": [
    67
   ]
  },
  {
   "term": "Four-fifths rule (80 percent rule)",
   "definition": "The guideline of the Uniform Guidelines on Employee Selection Procedures under which a selection rate for any race, sex or ethnic group less than four-fifths (80 percent) of the rate for the group with the highest rate will generally be regarded by federal enforcement agencies as evidence of adverse impact (29 C.F.R. § 1607.4(D)).",
   "chapters": [
    14,
    88
   ]
  },
  {
   "term": "Framework",
   "definition": "A structured, usually voluntary set of outcomes and practices for managing a type of risk, such as the NIST AI RMF.",
   "chapters": [
    89
   ]
  },
  {
   "term": "Framework integration",
   "definition": "A library designed to extend a popular machine learning or data framework (for example scikit-learn or pandas) so closely that it follows that framework's conventions, which minimizes adoption friction.",
   "chapters": [
    58
   ]
  },
  {
   "term": "Freshness",
   "definition": "How recent the status information is that a verifier relies on. The staleness window is the time between a status change by the issuer and the moment a given verifier acts on it.",
   "chapters": [
    99
   ]
  },
  {
   "term": "Fundamental problem of causal inference",
   "definition": "The impossibility of observing both Y_t(u) and Y_c(u) on the same unit (Holland, 1986).",
   "chapters": [
    33
   ]
  },
  {
   "term": "Fundamental rights impact assessment (FRIA)",
   "definition": "An assessment that certain deployers of high-risk AI systems must carry out before first use (Article 27 AI Act), describing the processes, affected groups, specific risks of harm, oversight measures and remedies.",
   "chapters": [
    9,
    86
   ]
  },
  {
   "term": "Gate",
   "definition": "A checkpoint at the end of a phase at which stated criteria must be met before the next phase begins; a failed gate sends the work back or stops the audit.",
   "chapters": [
    110
   ]
  },
  {
   "term": "General-purpose AI model",
   "definition": "The term of the EU AI Act for an AI model, including one trained on a large amount of data using self-supervision at scale, that displays significant generality, can competently perform a wide range of distinct tasks and can be integrated into a variety of downstream systems or applications (Regulation (EU) 2024/1689, Art. 3(63); European Parliament and Council of the European Union, 2024).",
   "chapters": [
    112
   ]
  },
  {
   "term": "Generalizability",
   "definition": "The extent to which results of a study apply in other contexts or populations that differ from the original one (National Academies of Sciences, Engineering, and Medicine, 2019, which adopts the definition used by the National Science Foundation).",
   "chapters": [
    96
   ]
  },
  {
   "term": "Generative AI",
   "definition": "Systems that produce new content (text, images, audio, video, code) from an input. Text-to-image models, which generate an image from a written description, are the second major family discussed in Chapter 112.",
   "chapters": [
    112
   ]
  },
  {
   "term": "Generative AI system",
   "definition": "A system that produces new content (text, images, audio, video, code) in response to an input. The NIST Generative AI Profile uses the abbreviation GAI (National Institute of Standards and Technology [NIST], 2024).",
   "chapters": [
    69
   ]
  },
  {
   "term": "Generative model",
   "definition": "A model that can draw new samples from an estimate of a data distribution. The main families used for tabular and image data are generative adversarial networks (GANs), in which a generator is trained against a discriminator that tries to tell real from generated samples (Goodfellow et al., 2014), and variational autoencoders (VAEs), which learn a latent representation and a decoder through a variational bound (Kingma and Welling, 2014). Simpler statistical generators, such as interpolation between neighboring records, are also used.",
   "chapters": [
    50
   ]
  },
  {
   "term": "Global explanation",
   "definition": "An explanation e(f) of the model's behavior over the whole data distribution, such as a permutation importance ranking or a partial dependence curve (Chapter 72).",
   "chapters": [
    76
   ]
  },
  {
   "term": "Goodhart's law",
   "definition": "The observation, as Thomas and Uminsky (2022) state it, that when a measure becomes a target, it ceases to be an effective measure.",
   "chapters": [
    45
   ]
  },
  {
   "term": "Governance framework",
   "definition": "A collection of one or more governance documents published by the governing body of an ecosystem or trust community, setting the policies on which actors may perform which actions on which data (Trust over IP Foundation, 2026). In practice such frameworks also fix the procedures for admission, supervision and removal.",
   "chapters": [
    100
   ]
  },
  {
   "term": "Governance gate",
   "definition": "A checkpoint in development at which fairness properties must be formally verified against stated criteria before work proceeds.",
   "chapters": [
    108
   ]
  },
  {
   "term": "Gradient reversal layer (GRL)",
   "definition": "A layer that passes its input unchanged in the forward pass and multiplies the gradient by a negative constant in the backward pass, so that one backpropagation step trains the adversary to minimize its loss and the predictor to maximize it (Ganin and Lempitsky, 2015).",
   "chapters": [
    53
   ]
  },
  {
   "term": "Group-conditional shift",
   "definition": "A change in the distribution of features, scores or outcomes within one group defined by A, for example a change in P(X | A = a), that may be small or invisible in the aggregate distribution.",
   "chapters": [
    61
   ]
  },
  {
   "term": "Group fairness",
   "definition": "A family of criteria that compare a statistical property of predictions or decisions (a rate, an error rate, a precision) across groups defined by A.",
   "chapters": [
    12
   ]
  },
  {
   "term": "Group fairness metric",
   "definition": "A function of decision records that compares a statistical property of predictions or decisions (a rate, an error rate, a predictive value) between groups defined by the protected attribute A. The notation follows Chapter 12: Y is the true outcome, Ŷ the prediction, D the decision, a and b two groups.",
   "chapters": [
    39
   ]
  },
  {
   "term": "Group label",
   "definition": "The value of the protected attribute, or of a combination of protected attributes, assigned to each record. An intersectional group is defined by a combination, for example sex and age band.",
   "chapters": [
    40
   ]
  },
  {
   "term": "Group-level explanation",
   "definition": "An explanation of the model's behavior on a defined subpopulation, or of a contrast between subpopulations. Examples are attributions averaged within a group, effect curves computed separately per group, an attribution of a fairness metric to features, or a summary of recourse options for a group.",
   "chapters": [
    76
   ]
  },
  {
   "term": "Group risk",
   "definition": "The expected loss of a model on the members of group k, written R_k. The overall risk is a weighted average of group risks with the group shares as weights.",
   "chapters": [
    26
   ]
  },
  {
   "term": "Group ROC curve",
   "definition": "The curve of operating points traced by a threshold as it moves over all values, computed separately within each group (Chapter 15).",
   "chapters": [
    55
   ]
  },
  {
   "term": "Group share",
   "definition": "The proportion of records in a dataset that belong to group g, written π_g. The corresponding share in the reference population is written π*_g.",
   "chapters": [
    37
   ]
  },
  {
   "term": "Group-specific regularization",
   "definition": "A penalty with separate weights, or separate forms, for different groups or intersectional subgroups.",
   "chapters": [
    52
   ]
  },
  {
   "term": "Group-specific threshold",
   "definition": "A threshold t_a that depends on the value a of the protected attribute A. The resulting classifier is Ŷ = 1 if R ≥ t_A.",
   "chapters": [
    55
   ]
  },
  {
   "term": "Guardrail",
   "definition": "A control placed around the model, such as an input classifier, an output classifier or a rule that routes a case to human review.",
   "chapters": [
    67
   ]
  },
  {
   "term": "Hallucination",
   "definition": "Generated text that is nonsensical or unfaithful to the provided source content, often while appearing fluent and natural (Ji et al., 2023).",
   "chapters": [
    67
   ]
  },
  {
   "term": "Hard and soft binding",
   "definition": "In content provenance, a hard binding is one or more cryptographic hashes that uniquely identify an asset or part of it; a soft binding is an identifier that is not statistically unique, such as a fingerprint, or an invisible watermark (C2PA, n.d.).",
   "chapters": [
    98
   ]
  },
  {
   "term": "Hard and soft fairness",
   "definition": "A hard constraint must be satisfied within its tolerance; a soft penalty adds a cost for disparity to the objective without guaranteeing any level of satisfaction (Chapter 52).",
   "chapters": [
    51
   ]
  },
  {
   "term": "Harmonised standard",
   "definition": "In EU law, a European standard adopted on request of the Commission for the application of Union harmonisation legislation. Under the AI Act, conformity with harmonised standards whose references are published in the Official Journal gives a presumption of conformity with the corresponding requirements (Regulation (EU) 2024/1689, Article 40(1)).",
   "chapters": [
    89
   ]
  },
  {
   "term": "Hash function",
   "definition": "A function that maps a message of arbitrary length to a fixed-length message digest. A secure hash function makes it computationally infeasible to find a message with a given digest or two messages with the same digest (National Institute of Standards and Technology, 2015).",
   "chapters": [
    98
   ]
  },
  {
   "term": "Herd privacy",
   "definition": "The privacy that a holder gains because a verifier fetches a list covering many credentials, so that the party serving the list cannot tell which credential is being checked. The IETF specification uses the term herd privacy; the W3C specification speaks of group privacy.",
   "chapters": [
    99
   ]
  },
  {
   "term": "Heterogeneous treatment effect",
   "definition": "A treatment effect that differs between subgroups. Fairness experiments are primarily about such heterogeneity.",
   "chapters": [
    64
   ]
  },
  {
   "term": "High-risk AI system",
   "definition": "An AI system that is a safety component of, or itself is, a product covered by the Union harmonization legislation in Annex I and subject to third-party conformity assessment (Article 6(1)), or an AI system in one of the use cases listed in Annex III (Article 6(2)), unless the derogation in Article 6(3) applies.",
   "chapters": [
    86
   ]
  },
  {
   "term": "High-risk profiling",
   "definition": "Profiling that poses a high risk to the data subject's personality or fundamental rights by matching data that allow an assessment of essential aspects of a natural person's personality (Article 5(g) FADP). The concept has no direct GDPR equivalent.",
   "chapters": [
    87
   ]
  },
  {
   "term": "Historical bias",
   "definition": "Bias that arises even if data are perfectly measured and sampled, because the world as it is or was leads a model to produce harmful outcomes. A system can reflect the world accurately and still harm a group, for example by reproducing a stereotype (Suresh and Guttag, 2021).",
   "chapters": [
    6,
    25,
    38
   ]
  },
  {
   "term": "Historical context assessment record",
   "definition": "The documented output of the Historical Context Assessment Method: questionnaire answers with sources, the completed matrix, information gaps, and recommendations.",
   "chapters": [
    11
   ]
  },
  {
   "term": "Historical pattern",
   "definition": "A specific, documented pattern of discrimination with historical evidence, stated in a form that can be connected to system components (for example \"geographic credit grading that excluded minority neighborhoods\" rather than \"racism in finance\").",
   "chapters": [
    11
   ]
  },
  {
   "term": "Homogenization",
   "definition": "The tendency of generative systems to produce skewed, overly uniform distributions of outputs, for example repetitive styles or reduced diversity, which the U.S. National Institute of Standards and Technology lists together with harmful bias as one generative AI risk (National Institute of Standards and Technology, 2024).",
   "chapters": [
    112
   ]
  },
  {
   "term": "Horizontal AI regulation",
   "definition": "Law that applies to AI systems across sectors, such as the EU AI Act, typically by attaching duties to use cases or risk tiers.",
   "chapters": [
    91
   ]
  },
  {
   "term": "Horizontal and sectoral regulation",
   "definition": "Horizontal rules apply across domains (the EU AI Act, South Korea's AI Basic Act); sectoral rules apply to one domain (New York City's rules on automated employment decision tools).",
   "chapters": [
    85
   ]
  },
  {
   "term": "Horizontal effect of fundamental rights",
   "definition": "The extent to which fundamental rights, which primarily bind the state, also shape relationships between private persons. Article 35(3) of the Federal Constitution instructs the authorities to ensure that fundamental rights, where appropriate, apply to such relationships.",
   "chapters": [
    87
   ]
  },
  {
   "term": "Human agency",
   "definition": "The capacity of people to make informed, autonomous decisions regarding AI systems, supported by the knowledge and tools to comprehend and interact with them and, where possible, to self-assess or challenge them (High-Level Expert Group on Artificial Intelligence, 2019).",
   "chapters": [
    9
   ]
  },
  {
   "term": "Human-AI interaction bias",
   "definition": "Bias that emerges from the dynamic relationship between a system and its users, where user behavior and system responses influence each other in ways that create or amplify unfairness.",
   "chapters": [
    28
   ]
  },
  {
   "term": "Human-centered explainable AI",
   "definition": "An approach that puts the human at the center of technology design and develops an understanding of who the human is by considering values, interpersonal dynamics and the socially situated nature of AI systems (Ehsan and Riedl, 2020).",
   "chapters": [
    78
   ]
  },
  {
   "term": "Human dignity",
   "definition": "The unearned worth or status that all humans share equally (Debes, 2023); in Kant, the worth of what is above all value and admits of no equivalent (Kant, 1785). It is the foundational value of the EU Charter, whose Article 1 states that human dignity is inviolable and must be respected and protected.",
   "chapters": [
    5,
    9
   ]
  },
  {
   "term": "Human oversight",
   "definition": "Arrangements through which natural persons monitor, interpret and, where necessary, override or stop an AI system (Regulation (EU) 2024/1689, Article 14).",
   "chapters": [
    9
   ]
  },
  {
   "term": "Human rights",
   "definition": "Rights held by every person by virtue of being human, protected in international instruments such as the European Convention on Human Rights. Fundamental rights is the term used in EU law for the rights in the Charter of Fundamental Rights of the European Union.",
   "chapters": [
    9
   ]
  },
  {
   "term": "Human rights due diligence",
   "definition": "Under the UN Guiding Principles, the ongoing process by which a business enterprise identifies, prevents, mitigates and accounts for how it addresses its adverse human rights impacts.",
   "chapters": [
    84
   ]
  },
  {
   "term": "Hyperparameters",
   "definition": "Settings fixed before or outside training, such as regularization strength, learning rate and schedule, model capacity (tree depth, network width and depth) and stopping criteria.",
   "chapters": [
    26
   ]
  },
  {
   "term": "Hypervolume indicator",
   "definition": "The size of the region of objective space that a set of points dominates, bounded by a reference point; a single-number measure of the quality of a front approximation (Guerreiro et al., 2022).",
   "chapters": [
    54
   ]
  },
  {
   "term": "Identification",
   "definition": "The property that a causal quantity can be computed from the distribution of observed data under stated assumptions.",
   "chapters": [
    31,
    33
   ]
  },
  {
   "term": "Identity and identifier",
   "definition": "An identity is the set of attributes by which a person, organization or thing is known in a context; an identifier is a value that refers to it. A decentralized identifier (DID) is an identifier whose controller can prove control cryptographically (Chapter 97).",
   "chapters": [
    98
   ]
  },
  {
   "term": "Identity proofing",
   "definition": "The processes used to collect, validate and verify information about a subject to establish assurance in the subject's claimed identity; identity evidence must support both validation (authenticity and accuracy) and verification (that the applicant is its true owner) (Temoshok et al., 2025).",
   "chapters": [
    98
   ]
  },
  {
   "term": "Impact assessment",
   "definition": "A structured evaluation of a system's potential effects before and during deployment. Variants include the algorithmic impact assessment of Canada's directive, the data protection impact assessment of Article 35 GDPR and the fundamental rights impact assessment of Article 27 of the EU AI Act.",
   "chapters": [
    85
   ]
  },
  {
   "term": "Impact ratio",
   "definition": "In the New York City bias audit rules, the selection rate (or scoring rate) of a category divided by that of the most selected (or highest scoring) category (New York City Department of Consumer and Worker Protection, 2023). It is the selection rate ratio of Chapter 14 computed against the most favored category.",
   "chapters": [
    88,
    96
   ]
  },
  {
   "term": "Impossibility theorem",
   "definition": "A mathematical result showing that several desirable fairness criteria cannot hold simultaneously except in special cases, such as equal base rates or perfect prediction (Chapter 19).",
   "chapters": [
    3,
    19
   ]
  },
  {
   "term": "In-processing",
   "definition": "A fairness intervention that acts on the learning algorithm itself, by changing its objective, its constraints, its architecture or its training algorithm, rather than on the data (pre-processing) or on the outputs (post-processing). It is also called a model-level intervention.",
   "chapters": [
    46,
    51,
    52
   ]
  },
  {
   "term": "Independence",
   "definition": "(1) As a fairness criterion, the requirement that a system's output be statistically independent of the protected attribute: R ⊥ A for a score, Ŷ ⊥ A for a prediction, D ⊥ A for a decision (Chapter 14). (2) Of an assessor (Chapter 96), the absence of relationships, interests and pressures that could compromise, or appear to compromise, objective and impartial judgment on the matters assessed.",
   "chapters": [
    14,
    96
   ]
  },
  {
   "term": "Independence, separation, sufficiency",
   "definition": "The three conditional-independence families of Barocas, Hardt and Narayanan (2023): R ⊥ A (independence), R ⊥ A | Y (separation) and Y ⊥ A | R (sufficiency).",
   "chapters": [
    19
   ]
  },
  {
   "term": "Indicator",
   "definition": "An observable sign that a bias type may be present and deserves investigation, for example demographic imbalance relative to the target population.",
   "chapters": [
    35
   ]
  },
  {
   "term": "Indigenous data sovereignty",
   "definition": "The claim of Indigenous Peoples to govern data about themselves, their lands and their knowledge, expressed for example in the CARE Principles (Collective benefit, Authority to control, Responsibility, Ethics) (Carroll et al., 2020).",
   "chapters": [
    4
   ]
  },
  {
   "term": "Indirect discrimination",
   "definition": "In EU law, the situation where \"an apparently neutral provision, criterion or practice would put persons\" sharing a protected characteristic \"at a particular disadvantage compared with other persons, unless that provision, criterion or practice is objectively justified by a legitimate aim and the means of achieving that aim are appropriate and necessary\" (Article 2(2)(b) of Directive 2000/43/EC, and the parallel provisions of the other directives). In the case law of the European Convention on Human Rights (Chapter 84), it covers disproportionately prejudicial effects of a general policy or measure couched in neutral terms, and does not necessarily require discriminatory intent. In causal fairness analysis (Chapter 18), indirect discrimination is the influence of A on the decision through intermediate variables that A causes.",
   "chapters": [
    18,
    23,
    84
   ]
  },
  {
   "term": "Indirect influence",
   "definition": "The influence of a feature on a model's outcomes that operates through other, related features, including the case where the feature is not referred to by the model at all (Adler et al., 2018).",
   "chapters": [
    71
   ]
  },
  {
   "term": "Indirect mechanism",
   "definition": "Influence of A on the decision through mediators, A → W → D.",
   "chapters": [
    30
   ]
  },
  {
   "term": "Individual, collective and societal harm",
   "definition": "Smuha (2021) distinguishes harm to an individual, harm to a collection of individuals, and harm to an interest of society at large that is not reducible to the sum of individual harms.",
   "chapters": [
    114
   ]
  },
  {
   "term": "Individual fairness",
   "definition": "A family of criteria that require similar individuals to be treated similarly, according to a task-specific similarity measure (Chapter 17).",
   "chapters": [
    12,
    17
   ]
  },
  {
   "term": "Inductive bias",
   "definition": "The set of assumptions a learning algorithm uses to generalize from limited data to unseen cases, such as linearity, smoothness, or the locality built into a convolutional architecture. Every learning algorithm has one; it is not a defect, but its fit can differ between groups.",
   "chapters": [
    26
   ]
  },
  {
   "term": "Inference",
   "definition": "A reasoning step that derives a claim (the conclusion) from a list of claims (the premises) under a specified context (International Organization for Standardization, International Electrotechnical Commission and Institute of Electrical and Electronics Engineers [ISO/IEC/IEEE], 2022).",
   "chapters": [
    94
   ]
  },
  {
   "term": "Infidelity",
   "definition": "The expected squared error between the change in model output predicted by an explanation for a perturbation and the actual change (Yeh et al., 2019).",
   "chapters": [
    77
   ]
  },
  {
   "term": "Influence function",
   "definition": "A measure of how an estimate, such as a model's parameters or its loss on a test point, would change if a training point were upweighted infinitesimally (Koh and Liang, 2017).",
   "chapters": [
    42
   ]
  },
  {
   "term": "Infrastructure and resource disparities",
   "definition": "Systematic differences in technological infrastructure, computing resources and technical expertise that create uneven access to AI systems across populations and regions.",
   "chapters": [
    28
   ]
  },
  {
   "term": "Infrastructure as code (IaC)",
   "definition": "Code, written in a declarative language, that provisions and configures the infrastructure resources hosting an application in a repeatable and consistent manner, and is versioned like application code instead of being configured by hand (Chandramouli, 2022).",
   "chapters": [
    60
   ]
  },
  {
   "term": "Inherent risk",
   "definition": "The risk a system poses by virtue of its purpose, function and context before any mitigating control is taken into account.",
   "chapters": [
    90
   ]
  },
  {
   "term": "Inherently interpretable model",
   "definition": "A model constrained in form so that it is useful to someone or obeys structural knowledge of the domain, such as monotonicity, additivity, sparsity or causal constraints (Rudin, 2019). Molnar (2025) calls this interpretability by design.",
   "chapters": [
    70
   ]
  },
  {
   "term": "Inherited bias",
   "definition": "Associations learned during pretraining that persist in a downstream model after fine-tuning.",
   "chapters": [
    65
   ]
  },
  {
   "term": "Instance influence",
   "definition": "How strongly a single training record affects the parameters a learning algorithm selects. Under unweighted empirical risk minimization every record contributes equally to the average loss.",
   "chapters": [
    47
   ]
  },
  {
   "term": "Instance weight (w_i)",
   "definition": "A non-negative number attached to record i that multiplies its contribution to the training loss. Equal weights reproduce ordinary training.",
   "chapters": [
    47
   ]
  },
  {
   "term": "Instrumental variable (IV)",
   "definition": "A variable that shifts the exposure, is independent of unmeasured confounders, and affects the outcome only through the exposure.",
   "chapters": [
    33
   ]
  },
  {
   "term": "Integrity",
   "definition": "(1) In trust research, the trustor's perception that the trustee adheres to a set of principles that the trustor finds acceptable (Mayer et al., 1995). (2) In data security (Chapter 98), the property that data has not been changed since a reference point, typically the moment it was hashed or signed; tamper-evident means a change is detectable, not that it is impossible.",
   "chapters": [
    1,
    98
   ]
  },
  {
   "term": "Inter-annotator agreement",
   "definition": "The extent to which independent annotators assign the same label to the same item, often measured by Cohen's kappa (Cohen, 1960).",
   "chapters": [
    37
   ]
  },
  {
   "term": "Interface and accessibility bias",
   "definition": "Barriers created by design choices, including visual layout, language, interaction modality and documentation, that disproportionately affect some groups' ability to use a system.",
   "chapters": [
    28
   ]
  },
  {
   "term": "Internal and external audit",
   "definition": "An internal audit is carried out by or for the organization that develops or deploys the system, with access to its artifacts; Raji et al. (2020) propose applying it throughout the development lifecycle. An external audit is carried out by a party outside the organization, with or without its cooperation, often only through the system's inputs and outputs (Metaxa et al., 2021).",
   "chapters": [
    36,
    96,
    110
   ]
  },
  {
   "term": "International Standard",
   "definition": "A document developed by consensus in ISO and IEC that provides requirements, specifications or guidelines. Some, such as ISO/IEC 42001, specify requirements against which an organization can be audited and certified; others, such as ISO/IEC 23894, provide guidance only.",
   "chapters": [
    89
   ]
  },
  {
   "term": "Interpretability",
   "definition": "In the machine learning literature, the ability to explain or to present in understandable terms to a human (Doshi-Velez and Kim, 2017); in NIST usage, the meaning of an AI system's output in the context of its designed functional purpose (National Institute of Standards and Technology, 2023). Both definitions place a human observer at the center: interpretability is a relation between a system and an audience, not a property of the system alone.",
   "chapters": [
    10,
    70
   ]
  },
  {
   "term": "Interpretability by design",
   "definition": "Training a model that is interpretable in itself, by constraining the learning algorithm to a class of interpretable models such as sparse linear models, short decision trees or rule lists (Molnar, 2025). Also called intrinsic or inherent interpretability; Arya et al. (2019) call such a model directly interpretable.",
   "chapters": [
    72
   ]
  },
  {
   "term": "Interruptibility",
   "definition": "The ability to halt an agent's operation, either a specific category of actions or the agent as a whole, in a safe state (Shavit et al., 2023; Regulation (EU) 2024/1689, Art. 14(4)(e); European Parliament and Council of the European Union, 2024).",
   "chapters": [
    113
   ]
  },
  {
   "term": "Intersectional counterfactual query",
   "definition": "A counterfactual question that changes several protected attributes at once, for example race and gender jointly.",
   "chapters": [
    32
   ]
  },
  {
   "term": "Intersectional group",
   "definition": "A group defined by a combination of protected attributes, for example women of color, written A = (A₁, A₂). Chapter 29 treats intersectionality in full.",
   "chapters": [
    37
   ]
  },
  {
   "term": "Intersectional (multiple) discrimination",
   "definition": "Discrimination that arises from a combination of protected characteristics, such as against Black women, and that may not be visible when each characteristic is examined on its own (Chapter 29).",
   "chapters": [
    83
   ]
  },
  {
   "term": "Intersectional subgroup",
   "definition": "A group defined by a combination of values of two or more protected attributes, such as women of a given ethnic origin over 65. With protected attributes A₁, ..., A_p, the finest intersectional subgroups are the cells of the cross-classification A = (A₁, ..., A_p). The term follows the intersectionality framework introduced by Crenshaw (1989).",
   "chapters": [
    29,
    49,
    76
   ]
  },
  {
   "term": "Intersectionality",
   "definition": "A framework, originating in legal scholarship and Black feminist thought, holding that the experience of discrimination of a person who belongs to several disadvantaged groups is shaped by the combination of those memberships and cannot be reduced to the sum of separate disadvantages (Crenshaw, 1989).",
   "chapters": [
    3,
    29
   ]
  },
  {
   "term": "Intersectionality matrix",
   "definition": "A retrospective exercise that tabulates which intersections of protected attributes were tested, with what sample, to expose blind spots.",
   "chapters": [
    105
   ]
  },
  {
   "term": "Intervention",
   "definition": "An operation, written do(X = x), that sets X to x by replacing its structural equation, so that X no longer responds to its usual causes while its effects respond to the new value (Pearl, 2009). Interventional quantities such as P(D = 1 | do(V = v)) cannot in general be computed from associations alone (Pearl, 2019).",
   "chapters": [
    18,
    30,
    75
   ]
  },
  {
   "term": "Intrinsic and extrinsic metric",
   "definition": "An intrinsic metric measures bias inside the model, in its embeddings or token probabilities; an extrinsic metric measures bias in the output of a task the model performs. The distinction matters because the two often disagree (Goldfarb-Tarrant et al., 2021).",
   "chapters": [
    112
   ]
  },
  {
   "term": "INVEST",
   "definition": "A mnemonic for good stories: Independent, Negotiable, Valuable, Estimable, Small and Testable (Wake, 2003).",
   "chapters": [
    103
   ]
  },
  {
   "term": "Issuer, holder, verifier",
   "definition": "The issuer asserts claims, creates the verifiable credential and transmits it to a holder. The holder possesses verifiable credentials and generates presentations from them; the holder is often, but not always, the subject (a parent may hold a child's credential). The verifier receives credentials or presentations for processing; other specifications call it a relying party.",
   "chapters": [
    97
   ]
  },
  {
   "term": "Joint, marginal and conditional distribution",
   "definition": "The joint distribution gives the probability of each combination of values, for example P(Ŷ = 1, A = a). A marginal distribution concerns one variable on its own, for example P(Ŷ = 1). A conditional distribution describes one variable within the subpopulation defined by another, for example P(Ŷ = 1 | A = a).",
   "chapters": [
    13
   ]
  },
  {
   "term": "Justification",
   "definition": "An account of why a decision is good, which does not necessarily describe the actual decision-making process (Miller, 2019, following Biran and Cotton). Justification and explanation are frequently confused; the difference is decisive for fairness, because a plausible justification can hide a discriminatory process.",
   "chapters": [
    70
   ]
  },
  {
   "term": "k-anonymity",
   "definition": "A release property under which the information for each person cannot be distinguished from that of at least k − 1 others whose information also appears in the release (Sweeney, 2002).",
   "chapters": [
    63
   ]
  },
  {
   "term": "Kantian ethics",
   "definition": "The deontological ethics of Kant, whose formula of humanity requires treating humanity, in oneself and in every other person, always as an end and never merely as a means, and which ascribes to persons a dignity that is above all value (Kant, 1785).",
   "chapters": [
    4
   ]
  },
  {
   "term": "Known limitations register",
   "definition": "A maintained list of identified fairness limitations of a system, with their status and monitoring.",
   "chapters": [
    107
   ]
  },
  {
   "term": "Label bias",
   "definition": "Systematic error in the outcome variable used for training or evaluation that differs across groups, whether it comes from historical decisions, from how the outcome is measured or from annotators' judgments.",
   "chapters": [
    37
   ]
  },
  {
   "term": "Label quality",
   "definition": "The extent to which the recorded outcome Y measures the construct of interest without group-dependent error; for example, health care costs are a biased proxy for health needs (Obermeyer et al., 2019).",
   "chapters": [
    22
   ]
  },
  {
   "term": "Label shift (prior probability shift)",
   "definition": "The label marginal changes but the conditional distribution of the inputs given the label does not (Lipton, Wang and Smola, 2018): P_T(Y) ≠ P_S(Y), P_T(X | Y) = P_S(X | Y). In fairness analysis: the base rates p_a change while P(X | Y, A) stays fixed.",
   "chapters": [
    34
   ]
  },
  {
   "term": "Lagrange multiplier (λ)",
   "definition": "A non-negative weight attached to a constraint in the Lagrangian; it expresses how much the objective would improve if the constraint were relaxed slightly.",
   "chapters": [
    51
   ]
  },
  {
   "term": "Large language model (LLM)",
   "definition": "A neural network, typically a transformer, trained on very large text corpora to predict tokens, which can generate text conditioned on a prompt and is usually adapted afterwards, for example by fine-tuning, instruction tuning and preference optimization. In the survey by Gallegos et al. (2024, Definition 1), an LLM is a model with an autoregressive, autoencoding or encoder-decoder architecture trained on a corpus of hundreds of millions to trillions of tokens.",
   "chapters": [
    67,
    80,
    112
   ]
  },
  {
   "term": "Layered communication",
   "definition": "Communication that gives each audience the level of detail and the form it needs, with consistent content across layers.",
   "chapters": [
    107
   ]
  },
  {
   "term": "Leakage",
   "definition": "The extent to which the protected attribute A can be recovered from a model's representations or outputs, or from transformed features (Wang et al., 2019). It is usually measured by training an auxiliary classifier, a probe or attacker, to predict A, and reported as its accuracy above the chance level.",
   "chapters": [
    48,
    53,
    65,
    68
   ]
  },
  {
   "term": "Learning bias",
   "definition": "Bias that arises when modeling choices amplify performance disparities across different examples in the data, for example when prioritizing one objective, such as overall accuracy, damages another, such as disparate impact (Suresh and Guttag, 2021).",
   "chapters": [
    26
   ]
  },
  {
   "term": "Learning to defer",
   "definition": "Training a model to decide when to pass a case to a downstream decision maker, taking that decision maker's accuracy and biases into account (Madras et al., 2018).",
   "chapters": [
    57
   ]
  },
  {
   "term": "Legal floor",
   "definition": "The minimum set of obligations that applicable law imposes on a system at a given time and place. It varies by jurisdiction, sector and use case, and it changes over time.",
   "chapters": [
    92
   ]
  },
  {
   "term": "Legal regime",
   "definition": "The set of legal instruments that apply to the decision, such as anti-discrimination law, sector rules and AI regulation.",
   "chapters": [
    22
   ]
  },
  {
   "term": "Less discriminatory alternative",
   "definition": "An alternative practice that would serve the respondent's legitimate interest with less disparate impact. If the complaining party demonstrates one and the respondent refuses to adopt it, liability can follow even when the original practice was justified.",
   "chapters": [
    83,
    88
   ]
  },
  {
   "term": "Leveling down",
   "definition": "Meeting a parity criterion by worsening the outcomes or error rates of the better-served group without improving those of the worse-served group.",
   "chapters": [
    20,
    55
   ]
  },
  {
   "term": "Libertarianism",
   "definition": "The view that a distribution is just if everyone is entitled to the holdings they possess, where entitlement derives from just acquisition, just transfer and the rectification of past injustice, rather than from the pattern of the distribution (Lamont and Favor, 2017).",
   "chapters": [
    3
   ]
  },
  {
   "term": "Likelihood (L)",
   "definition": "How often the pattern appears in similar systems, on a three-level scale.",
   "chapters": [
    11
   ]
  },
  {
   "term": "Limited assurance engagement",
   "definition": "An assurance engagement in which engagement risk is acceptable in the circumstances but greater than for reasonable assurance, and whose conclusion conveys whether, based on the procedures performed, anything has come to the practitioner's attention to cause it to believe the subject matter information is materially misstated. Its procedures are limited compared with reasonable assurance but must still be planned to obtain a level of assurance that is meaningful, that is, likely to enhance users' confidence to a degree clearly more than inconsequential (IAASB, 2013).",
   "chapters": [
    101
   ]
  },
  {
   "term": "Lipschitz condition",
   "definition": "The requirement that the output distance never exceed L times the input distance, for a constant L (the Lipschitz constant).",
   "chapters": [
    17
   ]
  },
  {
   "term": "LLM-as-a-judge",
   "definition": "The use of a language model to evaluate or rank outputs, including the outputs of other models (Zheng et al., 2023).",
   "chapters": [
    69
   ]
  },
  {
   "term": "Local and global attribution",
   "definition": "A local attribution explains one prediction f(x). A global attribution describes the model's behavior over a population, for example the mean absolute attribution of each feature or the drop in accuracy when a feature is permuted.",
   "chapters": [
    73
   ]
  },
  {
   "term": "Local and global explanation",
   "definition": "A local explanation is for a single prediction; a global explanation describes the behavior of the entire model (Arya et al., 2019). Applying a global method to a subset of the data gives a group-wise or regional explanation (Molnar, 2025).",
   "chapters": [
    72
   ]
  },
  {
   "term": "Local and global interpretability",
   "definition": "Global interpretability means knowing what patterns are present in the model in general; local interpretability means knowing the reasons for a specific decision, such as why a particular loan application was rejected (Doshi-Velez and Kim, 2017).",
   "chapters": [
    70
   ]
  },
  {
   "term": "Local explanation",
   "definition": "An explanation e(f, x) of the model output f(x) for one instance x, such as a vector of feature attributions or a counterfactual (Chapter 70).",
   "chapters": [
    76
   ]
  },
  {
   "term": "Look-back review",
   "definition": "The systematic re-examination of all decisions made during the period in which an incident was active, to identify every affected person.",
   "chapters": [
    109
   ]
  },
  {
   "term": "Luck egalitarianism",
   "definition": "The view that inequalities resulting from people's free and informed choices may be acceptable, while those resulting from brute luck (circumstances beyond a person's control) should be corrected (Binns, 2018).",
   "chapters": [
    3
   ]
  },
  {
   "term": "Manipulation",
   "definition": "A deliberate change to the model, its inputs or its deployment that alters explanations while leaving the model's behavior on real data essentially unchanged (Chapters 71 and 73).",
   "chapters": [
    77
   ]
  },
  {
   "term": "Marginal analysis",
   "definition": "An analysis that compares groups defined by one protected attribute at a time.",
   "chapters": [
    76
   ]
  },
  {
   "term": "Marginal group",
   "definition": "A group defined by one value of one protected attribute (for example, all women), irrespective of the values of the other attributes.",
   "chapters": [
    29
   ]
  },
  {
   "term": "Markov equivalence class",
   "definition": "The set of directed acyclic graphs that imply the same conditional independences. Two DAGs are Markov equivalent when they imply the same conditional independencies; data alone cannot distinguish them (Textor et al., 2016).",
   "chapters": [
    31,
    33
   ]
  },
  {
   "term": "Masking",
   "definition": "Intentional discrimination disguised as a neutral, data-driven process (Barocas and Selbst, 2016).",
   "chapters": [
    23
   ]
  },
  {
   "term": "Massaging",
   "definition": "Changing the class labels of selected records (promoting some members of the disadvantaged group from the negative to the positive class and demoting the same number of members of the favored group) until the training data show no disparity (Kamiran and Calders, 2012). It is a relabeling method, not a weighting method, and is treated alongside weighting methods in Chapter 47 because it is the natural comparison.",
   "chapters": [
    47
   ]
  },
  {
   "term": "Maximum Mean Discrepancy (MMD)",
   "definition": "A kernel-based distance between two probability distributions, estimated from two samples; used as a two-sample test statistic for shift detection (Gretton et al., 2012).",
   "chapters": [
    61
   ]
  },
  {
   "term": "Meaningful human control",
   "definition": "The condition that an autonomous system remains responsive to the relevant moral reasons of the humans who design and deploy it and to the relevant facts of its environment (tracking), and that its outcomes can always be traced back to at least one human along the chain of design and operation who understands the system and accepts responsibility (tracing) (Santoni de Sio and van den Hoven, 2018).",
   "chapters": [
    113
   ]
  },
  {
   "term": "Meaningful information about the logic involved",
   "definition": "The information that GDPR Articles 13(2)(f), 14(2)(g) and 15(1)(h) require for automated decisions referred to in Article 22(1) and (4), together with the significance and envisaged consequences of the processing.",
   "chapters": [
    79
   ]
  },
  {
   "term": "Measurement bias",
   "definition": "Bias that arises when choosing, collecting or computing the features and labels of a prediction problem. A feature or label is usually a proxy, a concrete measurement chosen to approximate a construct that cannot be observed directly. Proxies become problematic when they reflect the construct poorly or are generated differently across groups (Suresh and Guttag, 2021).",
   "chapters": [
    25,
    37
   ]
  },
  {
   "term": "Measurement model",
   "definition": "The assumptions by which an unobservable construct, such as creditworthiness or need, is inferred from observable properties; also called an operationalization (Jacobs and Wallach, 2021).",
   "chapters": [
    45
   ]
  },
  {
   "term": "Mechanistic interpretability",
   "definition": "Reverse engineering of the computations and representations inside a neural network into human-understandable algorithms and concepts, aiming at a causal understanding (Bereska and Gavves, 2024).",
   "chapters": [
    80
   ]
  },
  {
   "term": "Mediation analysis",
   "definition": "The decomposition of a total causal effect into a direct effect and indirect effects that pass through intermediate variables (mediators).",
   "chapters": [
    64
   ]
  },
  {
   "term": "Mediator",
   "definition": "A variable on a causal path between a cause and an effect. In causal mediation analysis of a neural network, internal components such as neurons or attention heads are treated as mediators between input and output (Vig et al., 2020).",
   "chapters": [
    30,
    75
   ]
  },
  {
   "term": "Membership inference",
   "definition": "An attack that infers whether a given data point was part of a model's training set; Shokri et al. (2021) use it to quantify how much explanations leak about training data.",
   "chapters": [
    82
   ]
  },
  {
   "term": "Merkle tree",
   "definition": "A binary tree whose leaves are hashes of data entries and whose internal nodes are hashes of their children; the root commits to all entries at once (Laurie et al., 2021).",
   "chapters": [
    98
   ]
  },
  {
   "term": "Meta-metric",
   "definition": "A single number summarizing a base metric across many groups, such as the difference between the largest and smallest group rate or the variance of group rates (Lum et al., 2022).",
   "chapters": [
    43
   ]
  },
  {
   "term": "Metric function",
   "definition": "Code that takes predictions, true outcomes and group labels and returns a fairness quantity, such as a per-group rate or a disparity between groups.",
   "chapters": [
    40
   ]
  },
  {
   "term": "Metric report",
   "definition": "A table giving, for each group and each metric, the count, the denominator, the estimate and an interval, together with the chosen disparity summaries and the conventions used.",
   "chapters": [
    39
   ]
  },
  {
   "term": "Metric specification",
   "definition": "The complete, written definition of a metric for a given assessment: population, outcome, groups and intersections, reference group, scale, sign convention, denominators, estimator, interval method and tolerance.",
   "chapters": [
    44
   ]
  },
  {
   "term": "Mid-Sprint fairness checkpoint",
   "definition": "A short, scheduled verification point between Sprint Planning and Sprint Review, with entry and exit criteria, at which fairness properties of work in progress are checked.",
   "chapters": [
    105
   ]
  },
  {
   "term": "Minimal intervention",
   "definition": "An action modeled as a causal intervention on one or more variables in a structural causal model, chosen at minimal cost so that, after all downstream consequences, the decision changes (Karimi, Schölkopf and Valera, 2021).",
   "chapters": [
    74
   ]
  },
  {
   "term": "Minimax game",
   "definition": "A training objective in which one player minimizes and the other maximizes the same quantity. Its solutions are saddle points rather than minima, which is why ordinary gradient descent can oscillate.",
   "chapters": [
    53
   ]
  },
  {
   "term": "Minimum detectable difference (MDD)",
   "definition": "The smallest true disparity that a test of given level would detect with a stated power, for given denominators.",
   "chapters": [
    43
   ]
  },
  {
   "term": "Minimum detectable effect (MDE)",
   "definition": "The smallest true effect that an experiment can detect with stated significance and power.",
   "chapters": [
    64
   ]
  },
  {
   "term": "Minimum group size",
   "definition": "A threshold below which a metric is computed but not reported as a finding, or not computed at all, because its estimate would be too imprecise to act on.",
   "chapters": [
    40
   ]
  },
  {
   "term": "Miscalibration",
   "definition": "A systematic gap between the score and the observed outcome rate. Underestimation means outcomes occur more often than the score says; overestimation, less often.",
   "chapters": [
    56
   ]
  },
  {
   "term": "Missing at random",
   "definition": "A condition defined by Rubin (1976): the probability that a value is missing may depend on the data that are observed but not on the missing values themselves. Rubin showed that under this condition (together with further conditions that depend on the type of inference) the process that causes missing data can be ignored. When missingness depends on the unobserved values, or when whole groups are absent by design, the condition fails and analyses that ignore the missingness can be biased.",
   "chapters": [
    7
   ]
  },
  {
   "term": "Missing-attribute strategy",
   "definition": "A rule for records whose protected attribute is unknown: exclude them, or treat \"missing\" as a group of its own.",
   "chapters": [
    40
   ]
  },
  {
   "term": "Modality dominance",
   "definition": "A situation in which one modality contributes disproportionately to the output, so that the model effectively ignores the others (Gat et al., 2020).",
   "chapters": [
    68
   ]
  },
  {
   "term": "Model-agnostic and model-specific",
   "definition": "A model-agnostic method queries the model only through its predictions (LIME, Kernel SHAP, permutation importance). A model-specific method uses the model's structure (TreeSHAP for tree ensembles, integrated gradients for differentiable models).",
   "chapters": [
    73
   ]
  },
  {
   "term": "Model-agnostic method",
   "definition": "A method that ignores the model's internals and analyzes only how outputs change with inputs. Molnar (2025) describes such methods by the SIPA principle: sample from the data, perform an intervention on the data, get the predictions for the manipulated data, and aggregate the results.",
   "chapters": [
    72
   ]
  },
  {
   "term": "Model card",
   "definition": "A short document accompanying a trained model that reports its intended use and its evaluated performance, disaggregated across relevant groups and their intersections (Mitchell et al., 2019).",
   "chapters": [
    107
   ]
  },
  {
   "term": "Model collapse",
   "definition": "Irreversible defects that arise when models are trained on content generated by earlier models, in which the tails of the original data distribution disappear (Shumailov et al., 2024).",
   "chapters": [
    69,
    114
   ]
  },
  {
   "term": "Model developer, system deployer and user",
   "definition": "The three human parties in an agent's life cycle (Shavit et al., 2023): the developer builds the underlying model; the deployer builds and operates the system on top of it, including its system prompt, tool access and interface; the user initiates a specific instance and gives it its goals.",
   "chapters": [
    113
   ]
  },
  {
   "term": "Model (f)",
   "definition": "The function being explained. In the compendium notation (Chapter 12) the model maps features X to a score R = f(X); for a classifier, f(x) is usually the predicted probability of the positive class or its log odds.",
   "chapters": [
    73
   ]
  },
  {
   "term": "Model-specific method",
   "definition": "A post hoc method that works only for a particular model class because it analyzes the model's internals, such as Gini importance in random forests or the features learned by a neural network (Molnar, 2025).",
   "chapters": [
    72
   ]
  },
  {
   "term": "Monitored metric",
   "definition": "A metric tracked and reported without being optimized, to detect disparities on non-prioritized dimensions.",
   "chapters": [
    22
   ]
  },
  {
   "term": "Monitoring metric",
   "definition": "A metric tracked to detect changes or trade-offs (for example selection rate parity when equal opportunity is primary) without being an acceptance criterion.",
   "chapters": [
    44
   ]
  },
  {
   "term": "Monitoring window",
   "definition": "The set of recent records over which a metric is computed. A sliding window contains the last N records or the last T units of time and moves forward with every new record; a tumbling window partitions time into consecutive, non-overlapping blocks (for example, calendar days).",
   "chapters": [
    61
   ]
  },
  {
   "term": "Moral equality",
   "definition": "The idea that every person has the same dignity and is owed the same respect, which in its modern form is the widely held conception of substantive, universal equality (Gosepath, 2021).",
   "chapters": [
    5
   ]
  },
  {
   "term": "Multi-agent system",
   "definition": "A setting in which several AI agents, possibly owned by different parties, interact, cooperate or compete.",
   "chapters": [
    69
   ]
  },
  {
   "term": "Multi-class prediction",
   "definition": "A prediction Ŷ with more than two possible values.",
   "chapters": [
    21
   ]
  },
  {
   "term": "Multi-dimensional fairness metric",
   "definition": "A fairness measure that is evaluated across all intersectional subgroups at once, or across a structured class of subgroups, rather than across the values of a single attribute.",
   "chapters": [
    29
   ]
  },
  {
   "term": "Multi-group fairness",
   "definition": "Fairness assessed across more than two groups, including groups defined by intersections of several protected attributes.",
   "chapters": [
    21
   ]
  },
  {
   "term": "Multi-modal model",
   "definition": "A model that processes two or more modalities, such as image and text, or video, audio and text. A vision-language model learns joint representations of images and text; CLIP, for example, was pre-trained to predict which caption goes with which image on 400 million image-text pairs collected from the internet (Radford et al., 2021).",
   "chapters": [
    68
   ]
  },
  {
   "term": "Multi-objective optimization problem",
   "definition": "The task of minimizing a vector of objectives F(θ) = (f_1(θ), ..., f_k(θ)) over model parameters θ, when no single θ minimizes all of them at once.",
   "chapters": [
    54
   ]
  },
  {
   "term": "Multi-scale analysis",
   "definition": "Decomposition of a metric series into components at different timescales (for example, trend, weekly cycle and residual) with separate detection on each component.",
   "chapters": [
    62
   ]
  },
  {
   "term": "Multiaccuracy",
   "definition": "The weaker requirement that predictions be accurate on average on every identifiable subgroup, achievable by black-box post-processing (Kim et al., 2019).",
   "chapters": [
    56
   ]
  },
  {
   "term": "Multicalibration",
   "definition": "Calibration required simultaneously on every subgroup in a rich, possibly overlapping collection of computationally identifiable subgroups (Hébert-Johnson et al., 2018).",
   "chapters": [
    56
   ]
  },
  {
   "term": "Multistakeholder recommendation",
   "definition": "A setting in which the end user is not the only party whose interests the recommender must take into account (Abdollahpouri et al., 2020).",
   "chapters": [
    66
   ]
  },
  {
   "term": "Mutual information",
   "definition": "A measure of statistical dependence between two variables that is zero exactly when they are independent and that, unlike a correlation coefficient, also captures non-linear and non-monotonic relationships (Cover and Thomas, 2006). For a feature Z and a protected attribute A, it is the reduction in uncertainty about A obtained by observing Z.",
   "chapters": [
    8,
    37
   ]
  },
  {
   "term": "Near miss",
   "definition": "A fairness failure that was caught before it affected any decision about a person, for example a disparity found in a pre-release test after a retraining. Near misses are recorded as hazards.",
   "chapters": [
    109
   ]
  },
  {
   "term": "Necessity and sufficiency",
   "definition": "A factor is necessary for an outcome if, had it been absent, the outcome would not have occurred; sufficient if its presence would produce the outcome. Probabilistic versions measure how likely each is (Galhotra, Pradhan and Salimi, 2021; Watson et al., 2021).",
   "chapters": [
    75
   ]
  },
  {
   "term": "Negative predictive value (NPV)",
   "definition": "Of group a: NPV_a = P(Y = 0 | Ŷ = 0, A = a).",
   "chapters": [
    16,
    19
   ]
  },
  {
   "term": "Notation for penalties",
   "definition": "Because Chapter 12 reserves R for a score and D for a decision, Chapter 52 writes a fairness penalty as Ω (Ω_fair, Ω_PR, Ω_DP) and a disparity measure as Δ.",
   "chapters": [
    52
   ]
  },
  {
   "term": "Null hypothesis (H₀)",
   "definition": "The hypothesis that the disparity being tested is absent, for example that a fairness metric M takes the same value in two groups, M(a) = M(b).",
   "chapters": [
    42
   ]
  },
  {
   "term": "Numerical and proportional equality",
   "definition": "Numerical equality treats all persons identically, giving each the same quantity; proportional equality treats each person in relation to their due, so that allocations are in the same ratio as the relevant claims. Numerical equality is just only when persons are equal in the relevant respects (Gosepath, 2021).",
   "chapters": [
    5
   ]
  },
  {
   "term": "Object of assessment",
   "definition": "The quantity whose fairness is examined: a score R, a prediction Ŷ, a decision D or downstream outcomes (Chapter 12). An assessment states its object explicitly.",
   "chapters": [
    36
   ]
  },
  {
   "term": "Objective",
   "definition": "A quantity to be minimized or maximized, for example the error rate, one minus the area under the ROC curve (AUC), the demographic parity difference, or the gap in true positive rates (TPR_a minus TPR_b). In Chapter 54 every objective is written so that smaller is better.",
   "chapters": [
    54
   ]
  },
  {
   "term": "Objective and reasonable justification",
   "definition": "The test under which a difference in treatment is not discriminatory: it pursues a legitimate aim and there is a reasonable relationship of proportionality between the means employed and the aim sought.",
   "chapters": [
    84
   ]
  },
  {
   "term": "Objective justification",
   "definition": "The European test for indirect discrimination: the practice pursues a legitimate aim, and the means of achieving it are appropriate and necessary.",
   "chapters": [
    83
   ]
  },
  {
   "term": "Observational criterion",
   "definition": "A fairness criterion defined only on the joint distribution of the protected attribute, the outcome and the score or prediction, such as demographic parity, equalized odds or calibration (Chapters 14 to 16).",
   "chapters": [
    45
   ]
  },
  {
   "term": "OECD Recommendation",
   "definition": "An OECD legal instrument adopted by the Council that is not legally binding; it represents a political commitment to the principles it contains and entails an expectation that Adherents will do their best to implement it (OECD, 2024, note on OECD legal instruments). It differs from an OECD Decision, which is legally binding on all Members except those that abstain at the time of adoption.",
   "chapters": [
    89
   ]
  },
  {
   "term": "Operating model",
   "definition": "The way fairness expertise and authority are distributed between a central function and business units: centralized, embedded or hybrid.",
   "chapters": [
    106
   ]
  },
  {
   "term": "Operating point",
   "definition": "(1) In threshold optimization (Chapter 55), the pair (FPR_a, TPR_a) that a decision rule produces in group a; each threshold yields one operating point per group. (2) In multi-objective model selection (Chapter 54), the model, and so the point on the front, that is finally selected for deployment.",
   "chapters": [
    54,
    55
   ]
  },
  {
   "term": "Operational fairness",
   "definition": "The practice of maintaining fairness properties during operation through a loop of detection and response (recalibration, retraining, human review or halting predictions).",
   "chapters": [
    61
   ]
  },
  {
   "term": "Optimal transport",
   "definition": "The problem of moving one probability distribution onto another at minimal total cost, where moving mass from x to x′ costs c(x, x′).",
   "chapters": [
    48
   ]
  },
  {
   "term": "Organizational implementation context",
   "definition": "The institutional priorities, workflows, incentives and decision processes that determine how a system's outputs are turned into decisions and who bears the costs and benefits of automation.",
   "chapters": [
    28
   ]
  },
  {
   "term": "Other status",
   "definition": "The open end of the lists of grounds in the ICCPR, the ECHR and the Universal Declaration, which allows characteristics not named in the text to be protected.",
   "chapters": [
    84
   ]
  },
  {
   "term": "Outcome homogenization and systemic failure",
   "definition": "Outcome homogenization is the extent to which particular individuals or groups experience negative outcomes from all decision makers; systemic failure is the event that an individual is failed (for example misclassified or rejected) by every system they encounter (Bommasani et al., 2022; Toups et al., 2023).",
   "chapters": [
    114
   ]
  },
  {
   "term": "Output distance (d_Y)",
   "definition": "A measure of how different two individuals' outcomes are, for example the absolute difference of two scores or the total variation distance between two distributions over decisions. When outcomes are distributions it is written D, which in Chapter 17 follows Dwork et al. (2012) and denotes a distance between distributions, not the decision D of Chapter 12.",
   "chapters": [
    17
   ]
  },
  {
   "term": "Overall evaluation criterion (OEC)",
   "definition": "The quantitative measure, or set of measures, that an experiment is designed to improve and on which the decision is based (Kohavi et al., 2020). In a fairness experiment it combines business and fairness metrics with explicit trade-off rules.",
   "chapters": [
    64
   ]
  },
  {
   "term": "Overreliance",
   "definition": "Accepting an AI suggestion even when it is wrong (Buçinca et al., 2021).",
   "chapters": [
    28
   ]
  },
  {
   "term": "Override rate",
   "definition": "The share of cases in which the final decision differs from the system's recommendation, computed per group.",
   "chapters": [
    28
   ]
  },
  {
   "term": "Overtrust and distrust",
   "definition": "Overtrust is poor calibration in which trust exceeds system capabilities; with distrust, trust falls short of them (Lee and See, 2004).",
   "chapters": [
    2
   ]
  },
  {
   "term": "Parents, descendants, ancestors",
   "definition": "The parents of V are the nodes with an edge into V; a descendant is reachable from V along directed edges; an ancestor is a node from which V is reachable.",
   "chapters": [
    31
   ]
  },
  {
   "term": "Pareto dominance",
   "definition": "Model θ′ dominates model θ if θ′ is no worse on every objective and strictly better on at least one.",
   "chapters": [
    20,
    54
   ]
  },
  {
   "term": "Pareto frontier",
   "definition": "In a problem with several objectives, the set of candidate solutions for which no other candidate is at least as good on every objective and strictly better on one.",
   "chapters": [
    20,
    26,
    64
   ]
  },
  {
   "term": "Pareto optimal (non-dominated) model",
   "definition": "A model that no other feasible model dominates. A model is weakly Pareto optimal if no other model is strictly better on every objective simultaneously; every Pareto optimal model is weakly Pareto optimal, but not the reverse (Wei & Niethammer, 2022). Chapter 20 calls it a Pareto optimal (efficient) model.",
   "chapters": [
    20,
    54
   ]
  },
  {
   "term": "Pareto set and Pareto front",
   "definition": "The Pareto set is the set of Pareto optimal solutions in decision (parameter) space; its image in objective space is the Pareto front (Guerreiro et al., 2022). In fairness work the front is often called the fairness-accuracy frontier.",
   "chapters": [
    54
   ]
  },
  {
   "term": "Partial pooling (shrinkage)",
   "definition": "Estimating each subgroup's rate as a compromise between its own data and information from the other subgroups, with more weight on its own data the larger it is.",
   "chapters": [
    33,
    43
   ]
  },
  {
   "term": "Participatory AI",
   "definition": "Approaches that involve affected communities in the design, development, deployment or governance of AI systems (Birhane et al., 2022).",
   "chapters": [
    115
   ]
  },
  {
   "term": "Pass criterion",
   "definition": "A measurable condition that a system must meet to pass a gate, including the rule for handling statistical uncertainty.",
   "chapters": [
    108
   ]
  },
  {
   "term": "Path",
   "definition": "A sequence of edges between two nodes, regardless of direction. A directed path follows the arrows. A back-door path from A to an outcome is a path that begins with an arrow into A.",
   "chapters": [
    31
   ]
  },
  {
   "term": "Path-specific counterfactual fairness",
   "definition": "The requirement that the prediction be the same as in the counterfactual world in which A takes a baseline value only along the unfair paths, with fair paths left at their actual values (Chiappa, 2019).",
   "chapters": [
    32
   ]
  },
  {
   "term": "Path-specific effect",
   "definition": "The effect of changing A that is transmitted only along a designated set of causal paths (Pearl, 2001; Avin et al., 2005).",
   "chapters": [
    18
   ]
  },
  {
   "term": "Patient care decision support tool",
   "definition": "Under the U.S. rule implementing Section 1557 of the Affordable Care Act, a tool used by a covered health program to support clinical decision making; covered entities must not discriminate through its use (45 C.F.R. § 92.210).",
   "chapters": [
    91
   ]
  },
  {
   "term": "Per-group confusion matrix",
   "definition": "For a binary prediction and a binary outcome, the four counts of a group: true positives (TP, Ŷ = 1 and Y = 1), false negatives (FN, Ŷ = 0 and Y = 1), false positives (FP, Ŷ = 1 and Y = 0) and true negatives (TN, Ŷ = 0 and Y = 0). Every metric in Chapter 39 is a function of these counts.",
   "chapters": [
    39
   ]
  },
  {
   "term": "Perfect prediction",
   "definition": "A predictor that assigns score 1 to every positive and score 0 to every negative, or a classifier with FPR = FNR = 0.",
   "chapters": [
    19
   ]
  },
  {
   "term": "Performative prediction",
   "definition": "A prediction that influences the outcome it aims to predict, because predictions support decisions; when performativity is ignored, it surfaces as distribution shift that is routinely addressed by retraining (Perdomo et al., 2020). A credit score that determines who receives credit, and thereby who builds a repayment history, is performative.",
   "chapters": [
    27,
    34,
    114
   ]
  },
  {
   "term": "Performative stability",
   "definition": "An equilibrium in which predictions are calibrated not against past outcomes but against the future outcomes that manifest from acting on them (Perdomo et al., 2020).",
   "chapters": [
    27
   ]
  },
  {
   "term": "Permutation test",
   "definition": "A test that obtains the null distribution of a statistic by recomputing it after randomly reassigning group labels.",
   "chapters": [
    42
   ]
  },
  {
   "term": "Phase",
   "definition": "A unit of the Fairness Audit Playbook workflow with one purpose, a responsible lead and a defined output artifact.",
   "chapters": [
    110
   ]
  },
  {
   "term": "Pipeline extension",
   "definition": "An integration pattern in which fairness components are added as stages (transformers, estimators, validators) of a standard machine learning pipeline.",
   "chapters": [
    58
   ]
  },
  {
   "term": "Plausibility",
   "definition": "How convincing an explanation is to humans: coherent, consistent with the output and seemingly logical (Jacovi and Goldberg, 2020; Agarwal et al., 2024). Plausibility is a property of the explanation and its reader, not of the model; an explanation can be highly plausible and unfaithful.",
   "chapters": [
    77,
    80
   ]
  },
  {
   "term": "Policy as code",
   "definition": "Declarative code that generates the rules and configuration parameters realizing security objectives at runtime, in the usage of Chandramouli (2022). By extension, fairness thresholds and alerting rules kept as versioned, declarative configuration can be managed as policy as code.",
   "chapters": [
    60
   ]
  },
  {
   "term": "Popularity bias",
   "definition": "The tendency of recommendation algorithms to recommend a few popular items frequently while most items, the long tail, are rarely recommended or ignored (Abdollahpouri, Burke and Mobasher, 2017; Mansoury et al., 2020).",
   "chapters": [
    27,
    66
   ]
  },
  {
   "term": "Population benchmark",
   "definition": "An external reference distribution of group membership, such as census data or local government statistics, used as the expected composition in representation tests.",
   "chapters": [
    38
   ]
  },
  {
   "term": "Position bias",
   "definition": "The tendency of users to examine and click top-ranked items more, regardless of relevance. Craswell et al. (2008) found that a cascade model, in which users read results from top to bottom and stop at the first worthwhile one, best explained position effects in early ranks of a web search engine.",
   "chapters": [
    21,
    66
   ]
  },
  {
   "term": "Positive action",
   "definition": "Specific measures to prevent or compensate for disadvantages linked to a protected characteristic, adopted with a view to ensuring full equality in practice, in the terms of the directive of the Council of the European Union (2000).",
   "chapters": [
    5
   ]
  },
  {
   "term": "Positive predictive value (PPV)",
   "definition": "Also called precision, of group a: PPV_a = P(Y = 1 | Ŷ = 1, A = a), the share of positive predictions that are correct.",
   "chapters": [
    16,
    19
   ]
  },
  {
   "term": "Post hoc explanation",
   "definition": "An explanation produced by a method applied after the model has been trained (Molnar, 2025), as opposed to an inherently interpretable model whose structure is itself the explanation (Rudin, 2019).",
   "chapters": [
    72,
    73
   ]
  },
  {
   "term": "Post hoc interpretability",
   "definition": "Information extracted from a model after training without necessarily elucidating how the model works, for example verbal explanations, visualizations, local explanations or explanations by example (Lipton, 2018).",
   "chapters": [
    70
   ]
  },
  {
   "term": "Post-processing",
   "definition": "Any intervention that changes how the outputs of a trained model are turned into decisions, including thresholds and threshold optimization, score transformations, recalibration and randomized decision rules; the model itself is not retrained. It is also called a prediction-level intervention.",
   "chapters": [
    46,
    55
   ]
  },
  {
   "term": "Postmortem",
   "definition": "A written record of an incident, its impact, the actions taken to resolve it, its root causes and the follow-up actions to prevent recurrence (Lunney and Lueder, 2017).",
   "chapters": [
    109
   ]
  },
  {
   "term": "Potential outcomes",
   "definition": "For a unit u and two conditions t and c, the outcomes Y_t(u) and Y_c(u) that would be observed under each condition. The unit-level causal effect is their difference.",
   "chapters": [
    33
   ]
  },
  {
   "term": "Power",
   "definition": "The probability that a test detects a disparity of a stated size when it exists.",
   "chapters": [
    42,
    43
   ]
  },
  {
   "term": "Pre-commit hook",
   "definition": "A script that Git runs on the developer's machine before a commit is created; if it exits with a non-zero status, the commit is aborted (Git, n.d.).",
   "chapters": [
    59
   ]
  },
  {
   "term": "Pre-processing (data-level intervention)",
   "definition": "An intervention that changes the training data: the weight of records, the values or representation of features, the labels, or the composition of the data set through generated records.",
   "chapters": [
    46
   ]
  },
  {
   "term": "Pre-training and fine-tuning",
   "definition": "Pre-training fits the model to a large unlabeled corpus. Fine-tuning continues training on a smaller, task-specific or value-specific dataset.",
   "chapters": [
    67
   ]
  },
  {
   "term": "Prediction (Ŷ)",
   "definition": "The model's guess of the outcome. For a binary task Ŷ ∈ {0, 1}, often produced by thresholding a score: Ŷ = 1 if R ≥ t.",
   "chapters": [
    12
   ]
  },
  {
   "term": "Prediction transformation",
   "definition": "A function T that maps a model output (and possibly the protected attribute) to a new output, T(R, A), used in place of the original R.",
   "chapters": [
    57
   ]
  },
  {
   "term": "Predictive equality",
   "definition": "Equal FPR across groups (the mirror relaxation; the name follows Corbett-Davies et al., 2017).",
   "chapters": [
    15
   ]
  },
  {
   "term": "Predictive parity",
   "definition": "Equal PPV across groups: PPV_a = PPV_b. Chouldechova (2017) distinguishes it from test fairness, her term for the score-level condition that a given score carries the same outcome probability in every group.",
   "chapters": [
    16,
    19
   ]
  },
  {
   "term": "Predictor",
   "definition": "The model that performs the primary task, predicting Y from X (and in some designs from A). Its parameters are written W.",
   "chapters": [
    53
   ]
  },
  {
   "term": "Preexisting, technical and emergent bias",
   "definition": "A classic taxonomy of bias in computer systems: preexisting bias has its roots in social institutions, practices and attitudes; technical bias arises from technical constraints or considerations; emergent bias arises in a context of use (Friedman and Nissenbaum, 1996).",
   "chapters": [
    6
   ]
  },
  {
   "term": "Preference articulation",
   "definition": "The way the decision maker's priorities enter the process: before the search (a priori), after the front has been computed (a posteriori), or repeatedly during the search (interactive) (Emmerich & Deutz, 2018).",
   "chapters": [
    54
   ]
  },
  {
   "term": "Preferential sampling",
   "definition": "Resampling in which records close to the decision boundary, as judged by a ranker, are duplicated or skipped first (Kamiran and Calders, 2012).",
   "chapters": [
    47
   ]
  },
  {
   "term": "Prejudice index (PI)",
   "definition": "The mutual information between the prediction and the protected attribute, estimated from the data (Kamishima et al., 2012). Its normalized form (NPI) lies between 0 and 1.",
   "chapters": [
    52
   ]
  },
  {
   "term": "Presentation and verifiable presentation (VP)",
   "definition": "Data derived from one or more verifiable credentials, possibly from several issuers, that is shared with a specific verifier. A verifiable presentation is tamper-evident; some verifiable presentations contain data synthesized from, but not containing, the original credentials, as with zero-knowledge proofs.",
   "chapters": [
    97
   ]
  },
  {
   "term": "Presumption of equality",
   "definition": "A procedural principle that everyone should receive an equal share of goods suitable for public distribution unless differences can be justified by reasons acceptable to all (Gosepath, 2021).",
   "chapters": [
    5
   ]
  },
  {
   "term": "Prima facie case",
   "definition": "The initial showing a complaining party must make (for disparate impact, that a particular practice causes a disparity on a protected ground) before the burden moves to the respondent.",
   "chapters": [
    83
   ]
  },
  {
   "term": "Primary and secondary criteria",
   "definition": "The fairness criterion that is enforced as a constraint (primary), and the criteria that are encouraged through regularization or monitored with tolerances (secondary).",
   "chapters": [
    20
   ]
  },
  {
   "term": "Primary chapter",
   "definition": "The chapter of the compendium that carries a scenario's full treatment, with its method, numbers and corrections.",
   "chapters": [
    111
   ]
  },
  {
   "term": "Primary criterion",
   "definition": "The criterion enforced as a constraint and reported as the main fairness claim.",
   "chapters": [
    22
   ]
  },
  {
   "term": "Primary metric",
   "definition": "The metric that operationalizes the primary fairness criterion selected for the system and on which acceptance decisions are based.",
   "chapters": [
    44
   ]
  },
  {
   "term": "Principal and delegation",
   "definition": "The principal is the person or organization on whose behalf an agent acts; delegation is the transfer of authority to act. The economic theory of agency describes the costs that arise when an agent's interests or information diverge from the principal's (Jensen and Meckling, 1976).",
   "chapters": [
    113
   ]
  },
  {
   "term": "Principal reasons",
   "definition": "In US credit law, the specific reasons for an adverse action that a creditor must disclose; they must relate to the factors actually considered or scored.",
   "chapters": [
    79
   ]
  },
  {
   "term": "Principle",
   "definition": "A high-level normative commitment, such as \"respect for human rights and democratic values, including fairness and privacy\". Principles state what matters; they do not prescribe procedures.",
   "chapters": [
    89
   ]
  },
  {
   "term": "Prior and posterior",
   "definition": "The prior is a probability distribution expressing what is assumed about a quantity before the data are seen; the posterior combines the prior with the data through Bayes' rule.",
   "chapters": [
    41
   ]
  },
  {
   "term": "Prioritarianism",
   "definition": "The view that benefits matter more the worse off the person who receives them; formally, outcomes are ranked by the sum of a strictly increasing and strictly concave transformation of individual well-being (Adler and Holtug, 2025).",
   "chapters": [
    3,
    4
   ]
  },
  {
   "term": "Priority score",
   "definition": "A score used to rank bias risks into priority bands. In the Historical Context Assessment Method (Chapter 11), P is the product S × L × R; in the Bias Source Identification Method (Chapter 35), it is the weighted sum of the five dimension scores, used to place a bias source in a high, medium or low priority category.",
   "chapters": [
    11,
    35
   ]
  },
  {
   "term": "Probabilistic proxy model",
   "definition": "A model that estimates the probability of protected class membership from observable proxies, such as surname and geolocation for race, used to impute missing labels for assessment (Chen, Kallus, Mao, Svacha and Udell, 2019).",
   "chapters": [
    24
   ]
  },
  {
   "term": "Probe (attacker)",
   "definition": "A classifier trained after the fact, on frozen outputs or representations, to test for leakage. Unlike the adversary, it plays no role in training.",
   "chapters": [
    53
   ]
  },
  {
   "term": "Procedural fairness",
   "definition": "Fairness of the process by which a decision is made (who decides, on what information, under which rules, with which possibilities of challenge), as distinct from the pattern of outcomes the process produces.",
   "chapters": [
    3
   ]
  },
  {
   "term": "Profiling",
   "definition": "Any form of automated processing of personal data consisting of the use of personal data to evaluate certain personal aspects relating to a natural person, in particular to analyze or predict aspects concerning performance at work, economic situation, health, personal preferences, interests, reliability, behavior, location or movements (Article 4(4) GDPR). The AI Act adopts this definition. The Swiss FADP uses a definition whose wording matches it (Article 5(f) FADP).",
   "chapters": [
    9,
    79,
    86,
    87
   ]
  },
  {
   "term": "Progressive disclosure",
   "definition": "A design principle that shows a small amount of essential information first and reveals detail on demand (drill-down), so that users are guided from summary to specifics.",
   "chapters": [
    63,
    107
   ]
  },
  {
   "term": "Prompt, system prompt and context window",
   "definition": "The prompt is the text given to the model. A system prompt is an instruction set fixed by the deployer and placed before user input. The context window is the total text the model conditions on, including any examples supplied in the prompt (in-context learning).",
   "chapters": [
    67
   ]
  },
  {
   "term": "Prompting paradigm and fine-tuning paradigm",
   "definition": "Zhao et al. (2024) organize explanation methods for language models by how the model is used: models that are fine-tuned for a downstream task (a classifier built on a pre-trained encoder) and models that are used through prompts, with no task-specific training. The explanation targets and methods differ between the two.",
   "chapters": [
    80
   ]
  },
  {
   "term": "Propensity score",
   "definition": "The conditional probability of assignment to a particular treatment given observed covariates (Rosenbaum and Rubin, 1983).",
   "chapters": [
    33
   ]
  },
  {
   "term": "Propensity to trust",
   "definition": "A stable characteristic of the trustor, a general willingness to trust others, which affects how much trust the trustor has before any information about the trustee is available (Mayer et al., 1995).",
   "chapters": [
    1
   ]
  },
  {
   "term": "Protected attribute (A)",
   "definition": "The characteristic with respect to which fairness is assessed, such as sex, ethnic origin, age band or disability status; its values define groups, and a and b denote two generic values. In law, a protected attribute (protected ground) is a personal characteristic on the basis of which unfavorable treatment is prohibited in a defined domain; which attributes are protected depends on the jurisdiction and the domain (Chapter 24).",
   "chapters": [
    12,
    83
   ]
  },
  {
   "term": "Protected attribute at inference",
   "definition": "Whether the intervention needs the protected attribute A of a person at the time a decision about that person is made, as opposed to only during development.",
   "chapters": [
    46
   ]
  },
  {
   "term": "Protected characteristic (protected ground)",
   "definition": "A characteristic on the basis of which discrimination is prohibited by a legal instrument, such as racial or ethnic origin, sex, religion or belief, disability, age or sexual orientation. Lists differ by instrument and domain (Chapter 24). In the formal notation of Chapter 12 it is the protected attribute A.",
   "chapters": [
    23,
    24
   ]
  },
  {
   "term": "Provenance",
   "definition": "In the W3C PROV model, a record that describes the people, institutions, entities and activities involved in producing, influencing or delivering a piece of data or a thing (Moreau and Missier, 2013).",
   "chapters": [
    98
   ]
  },
  {
   "term": "Provider",
   "definition": "A natural or legal person, public authority, agency or other body that develops an AI system, or has it developed, and places it on the market or puts it into service under its own name or trademark, whether for payment or free of charge (Article 3(3) AI Act). Most high-risk requirements fall on the provider.",
   "chapters": [
    86
   ]
  },
  {
   "term": "Proximity and sparsity",
   "definition": "Proximity means x′ is close to x under a chosen distance; sparsity means that few features change.",
   "chapters": [
    74
   ]
  },
  {
   "term": "Proxy",
   "definition": "(1) A variable whose values are correlated with a protected characteristic, so that decisions based on it track group membership (Chapter 23; see proxy variable). (2) For a construct (Chapter 7), a measured variable used in place of a concept that is not measured directly, for example health care cost as a stand-in for health need; such a proxy is biased with respect to a group when its relationship to the intended concept differs across groups. (3) In the formal causal sense of Kilbertus et al. (2017), a descendant of A, such as a name or a visual feature, designated as a variable that should not influence the prediction (Chapters 18, 30); causal analysis is needed to say whether a correlated feature is such a descendant.",
   "chapters": [
    7,
    18,
    23,
    30
   ]
  },
  {
   "term": "Proxy discrimination",
   "definition": "Disadvantage for a protected group that operates through proxy variables, features statistically associated with a protected attribute, rather than through the attribute itself; it can be intentional or, more often, an unintended by-product of optimizing for accuracy. Chapter 24 distinguishes a causal sense (Kilbertus et al., 2017), a sense based on the causal influence of protected class correlates (Datta, Fredrikson, Ko, Mardziel and Sen, 2017), and a legal sense in which the usefulness of a facially neutral practice derives, at least in part, from the very fact that it produces a disparate impact (Prince and Schwarcz, 2020).",
   "chapters": [
    8,
    24,
    48,
    49
   ]
  },
  {
   "term": "Proxy metric",
   "definition": "A fairness metric computed without the protected attribute itself, using legally permissible variables that are correlated with it (for example, geography), or using properties of the items affected (for example, the origin of artists in a music catalog).",
   "chapters": [
    61
   ]
  },
  {
   "term": "Proxy variable",
   "definition": "A feature that is not itself a protected attribute but is statistically associated with one, so that its use can reproduce group differences or let a model treat groups differently even when the protected attribute is excluded. Chapter 24 discusses which proxies are legally relevant and how they arise.",
   "chapters": [
    6,
    24,
    37,
    38,
    49
   ]
  },
  {
   "term": "Pull request (merge request)",
   "definition": "A proposal to merge a branch into another branch, typically the main branch, which is reviewed before it is accepted. GitHub uses the term pull request; GitLab uses merge request.",
   "chapters": [
    59
   ]
  },
  {
   "term": "Quality gate",
   "definition": "A point in the workflow at which a change can proceed only if specified checks pass.",
   "chapters": [
    59
   ]
  },
  {
   "term": "Quantile function",
   "definition": "For group a with cumulative distribution function F_a of a feature, the quantile function F_a^{-1}(u) returns the value below which a share u of the group lies. The value F_a(x) is the within-group rank of x.",
   "chapters": [
    48
   ]
  },
  {
   "term": "Quantile mapping",
   "definition": "Alignment that maps a score to the value at the same within-group quantile of a common target distribution, preserving rank order within each group.",
   "chapters": [
    57
   ]
  },
  {
   "term": "RACI matrix",
   "definition": "A table that assigns, for each decision or task, the roles that are Responsible (do the work), Accountable (answer for the result), Consulted (give input before the decision) and Informed (are told after it).",
   "chapters": [
    106
   ]
  },
  {
   "term": "Random variable",
   "definition": "A quantity whose value depends on which decision subject is drawn from the population, such as A, Y or Ŷ.",
   "chapters": [
    13
   ]
  },
  {
   "term": "Randomized classifier",
   "definition": "A distribution over a set of classifiers; to predict, one classifier is drawn and applied. Reductions methods return such mixtures.",
   "chapters": [
    51
   ]
  },
  {
   "term": "Randomized threshold",
   "definition": "A rule that accepts everyone above an upper threshold, rejects everyone below a lower threshold and accepts with a fixed probability in between. It reaches operating points that lie on the straight line between two points of the ROC curve.",
   "chapters": [
    55
   ]
  },
  {
   "term": "Rank preservation",
   "definition": "A transformation strongly preserves rank if every record keeps its within-group rank: F_a(x̃) = F_a(x) (Feldman et al., 2015).",
   "chapters": [
    48
   ]
  },
  {
   "term": "Rank-preserving (monotone) transformation",
   "definition": "A transformation that is non-decreasing in the score within each group, so that within-group order is unchanged.",
   "chapters": [
    57
   ]
  },
  {
   "term": "Ranking",
   "definition": "An ordering of a set of candidates (people, items, documents) presented to a user, who attends mostly to the top positions.",
   "chapters": [
    21
   ]
  },
  {
   "term": "Rawlsian justice",
   "definition": "The theory of justice as fairness developed by Rawls (1971), whose principles are those that parties would choose behind a \"veil of ignorance\" about their own position. Its second principle requires fair equality of opportunity and permits social and economic inequalities only if they are to the greatest benefit of the least advantaged (the difference principle).",
   "chapters": [
    3
   ]
  },
  {
   "term": "Re-performance",
   "definition": "A second party carries out the assessor's computations again, from the same inputs, to check that the reported figures follow from them.",
   "chapters": [
    96
   ]
  },
  {
   "term": "Real-time fairness monitoring",
   "definition": "The continuous computation of fairness metrics on the predictions and decisions a deployed system makes, per protected group, with alerts when a metric leaves its acceptable range. \"Real time\" ranges in practice from every prediction to micro-batches of minutes; the defining feature is that measurement is continuous and automatic.",
   "chapters": [
    61
   ]
  },
  {
   "term": "Reasonable accommodation",
   "definition": "Under the disability convention (CRPD), necessary and appropriate modification and adjustments, not imposing a disproportionate or undue burden, needed in a particular case to ensure persons with disabilities enjoy rights on an equal basis with others. Its denial is a form of discrimination.",
   "chapters": [
    5,
    84
   ]
  },
  {
   "term": "Reasonable and limited assurance",
   "definition": "Terms from assurance engagements under the international standard ISAE 3000 (Revised). In a reasonable assurance engagement the practitioner reduces engagement risk to an acceptably low level as the basis for a positively worded conclusion; in a limited assurance engagement the practitioner collects less evidence, and the conclusion is worded negatively (\"nothing came to our attention\") (Institute of Chartered Accountants in England and Wales [ICAEW], n.d.).",
   "chapters": [
    93
   ]
  },
  {
   "term": "Reasonable assurance engagement",
   "definition": "An assurance engagement in which the practitioner reduces engagement risk to an acceptably low level as the basis for a conclusion that conveys the practitioner's opinion on the measured or evaluated subject matter against the criteria (IAASB, 2013).",
   "chapters": [
    101
   ]
  },
  {
   "term": "Recalibration map",
   "definition": "A function fitted on held-out data that transforms raw scores into calibrated probabilities. A group-specific map is fitted separately for each group.",
   "chapters": [
    56
   ]
  },
  {
   "term": "Recommender system",
   "definition": "A system that suggests or prioritizes items for a user. The EU Digital Services Act defines it as a fully or partially automated system used by an online platform to suggest in its online interface specific information to recipients of the service or prioritize that information, including as a result of a search initiated by the recipient or otherwise determining the relative order or prominence of information displayed (Regulation (EU) 2022/2065, Article 3(s)).",
   "chapters": [
    66
   ]
  },
  {
   "term": "Reconciliation",
   "definition": "Checking that the explanation communicated for a decision is consistent with the logged inputs, model version and output of that decision.",
   "chapters": [
    81
   ]
  },
  {
   "term": "Recourse",
   "definition": "The ability of a person who received an unfavorable decision to change the model's inputs, by altering actionable input variables, in a way that obtains a favorable decision (Ustun, Spangher and Liu, 2019). A recourse gap is a difference between groups in how far, or at what cost, their members are from such a change (Gupta et al., 2019). In incident response (Chapter 109), recourse is the ability of an individual to change a decision about themselves, through action or appeal; redress is collective and backward-looking, recourse individual and often forward-looking.",
   "chapters": [
    9,
    10,
    71,
    109
   ]
  },
  {
   "term": "Recourse and recourse gap",
   "definition": "Recourse is the ability of a person to change their circumstances to obtain a favorable decision; a recourse gap is a difference between groups in its availability or cost (Gupta et al., 2019; Chapter 74).",
   "chapters": [
    82
   ]
  },
  {
   "term": "Recourse summary",
   "definition": "A global or group-level counterfactual explanation that describes, for a whole subpopulation of rejected people, a small set of actions that would change their outcomes and how often each action succeeds.",
   "chapters": [
    76
   ]
  },
  {
   "term": "Red teaming",
   "definition": "Systematic adversarial testing in which people, or other models, try to elicit harmful or biased outputs (Perez et al., 2022; Ganguli et al., 2022).",
   "chapters": [
    67
   ]
  },
  {
   "term": "Redress",
   "definition": "Measures that put right the situation of people affected by the incident: re-examination of their decisions, notification, explanation, correction of records, compensation and apology.",
   "chapters": [
    109
   ]
  },
  {
   "term": "Reductions approach",
   "definition": "A method that solves a fairness-constrained classification problem through a sequence of cost-sensitive (weighted) classification problems, solved by an unmodified base learner (Agarwal et al., 2018).",
   "chapters": [
    51
   ]
  },
  {
   "term": "Redundant encoding",
   "definition": "The situation in which membership of a protected class is encoded in other data, so that removing the protected attribute leaves the information available (Barocas and Selbst, 2016; Dwork et al., 2012).",
   "chapters": [
    17,
    24,
    49,
    71
   ]
  },
  {
   "term": "Reference (baseline, background)",
   "definition": "The point or distribution against which a prediction is compared. Every attribution explains a difference: between f(x) and the prediction at a baseline input, or between f(x) and the average prediction over a background distribution. Sundararajan, Taly and Yan (2017) argue that a baseline is implicit in every attribution, because humans assign blame by comparing a situation with the absence of the cause.",
   "chapters": [
    73
   ]
  },
  {
   "term": "Reference group",
   "definition": "The group against which other groups' metric values are compared when disparities are expressed as differences or ratios: for example the largest group, the most favored group, or the population as a whole.",
   "chapters": [
    44,
    58
   ]
  },
  {
   "term": "Reference group and comparison group",
   "definition": "The group against which others are compared. In legal practice it is often the group with the highest selection rate; in toolkits it is often the group designated as privileged. The reference group must be stated, because signs and ratios depend on it.",
   "chapters": [
    39
   ]
  },
  {
   "term": "Reference population",
   "definition": "The population against which the composition of a dataset is compared, such as census figures, the relevant labor market or the population a system is intended to serve. Its choice is a substantive decision and must be stated.",
   "chapters": [
    37
   ]
  },
  {
   "term": "Reference window and current window",
   "definition": "The period whose behavior is taken as normal (often the validation period or a stable period after launch) and the recent period compared against it.",
   "chapters": [
    62
   ]
  },
  {
   "term": "Regression",
   "definition": "Prediction of a real-valued outcome Y with a predictor f(X) and a loss function ℓ(Y, f(X)), such as squared error.",
   "chapters": [
    21
   ]
  },
  {
   "term": "Regularization",
   "definition": "Any constraint or penalty that limits model complexity to reduce overfitting, such as an L1 or L2 penalty on parameters, or implicit forms such as early stopping.",
   "chapters": [
    26
   ]
  },
  {
   "term": "Regularization strength (λ or η)",
   "definition": "The weight on the fairness penalty. It sets the exchange rate between prediction loss and unfairness. The literature uses both symbols; Chapter 52 writes λ except when quoting a formulation that uses η.",
   "chapters": [
    52
   ]
  },
  {
   "term": "Regulatory change",
   "definition": "Any change in the legal floor: new legislation, amendment, postponement of application dates, repeal or replacement, new or withdrawn guidance, and new case law.",
   "chapters": [
    92
   ]
  },
  {
   "term": "Regulatory sandbox",
   "definition": "A controlled framework set up by a competent authority in which AI systems can be developed and tested under supervision. Under the AI Act, each Member State must ensure that at least one national AI regulatory sandbox is operational by 2 August 2027 (Regulation (EU) 2024/1689, Article 57(1), as amended).",
   "chapters": [
    92
   ]
  },
  {
   "term": "Reinforcement learning from human feedback (RLHF)",
   "definition": "An adaptation method in which human raters rank model outputs, a reward model is trained on the rankings, and the language model is optimized against the reward model (Ouyang et al., 2022).",
   "chapters": [
    67
   ]
  },
  {
   "term": "Reject option classification",
   "definition": "A post-processing method that, in a confidence band around the decision boundary, assigns favorable outcomes to members of the unprivileged group and unfavorable outcomes to members of the privileged group (Kamiran et al., 2012).",
   "chapters": [
    57
   ]
  },
  {
   "term": "Relational equality",
   "definition": "The view that a society is just if and only if its members relate to one another as equals, without domination, exploitation, marginalization or demeaning treatment (Gosepath, 2021).",
   "chapters": [
    5
   ]
  },
  {
   "term": "Relevance (R)",
   "definition": "How directly the pattern applies to the system under assessment, on a three-level scale.",
   "chapters": [
    11
   ]
  },
  {
   "term": "Reliability diagram",
   "definition": "A plot of the observed outcome rate against the mean score within score bins; a calibrated score lies on the diagonal (Guo et al., 2017). It is also called a calibration curve.",
   "chapters": [
    16,
    56
   ]
  },
  {
   "term": "Reliance",
   "definition": "The behavior of depending on something or someone, in AI the behavior of depending on or following an automation's output, for example by accepting a recommendation. Trust guides reliance, especially when complexity makes a complete understanding of the automation impractical (Lee and See, 2004); reliance can be disappointed, whereas trust, in addition, can be betrayed (McLeod, 2020). Chapter 78 distinguishes overreliance (accepting AI advice when it is wrong), underreliance (rejecting it when it is right) and appropriate reliance.",
   "chapters": [
    1,
    2,
    78
   ]
  },
  {
   "term": "Remediation",
   "definition": "Changes to the system, its data, its decision rule or its operating process that end the failure and reduce the chance of recurrence.",
   "chapters": [
    109
   ]
  },
  {
   "term": "Remedy",
   "definition": "The means by which a violation of a right is redressed, including complaint mechanisms, review and compensation. Article 47 of the EU Charter guarantees an effective remedy before a tribunal.",
   "chapters": [
    9
   ]
  },
  {
   "term": "Repair",
   "definition": "The term of Feldman et al. (2015) for transforming non-protected attributes so that the protected attribute cannot be predicted from them. Full repair makes the distribution of each repaired feature identical across groups; partial repair moves it part of the way, controlled by a parameter λ ∈ [0, 1].",
   "chapters": [
    48
   ]
  },
  {
   "term": "Repair level (λ)",
   "definition": "A number between 0 and 1 that sets how far a partial repair moves each group's feature distribution towards a common target distribution; λ = 0 leaves the data unchanged and λ = 1 is the full repair.",
   "chapters": [
    49
   ]
  },
  {
   "term": "Replicability",
   "definition": "Obtaining consistent results across studies aimed at answering the same scientific question, each of which has obtained its own data (National Academies of Sciences, Engineering, and Medicine, 2019).",
   "chapters": [
    96
   ]
  },
  {
   "term": "Report-only mode",
   "definition": "A configuration in which failing fairness checks produce warnings and reports but do not block merges or deployments.",
   "chapters": [
    60
   ]
  },
  {
   "term": "Reporting factory",
   "definition": "An architecture in which one validated metrics store feeds several generated reports or views tailored to different roles, so that all views agree.",
   "chapters": [
    63
   ]
  },
  {
   "term": "Representation",
   "definition": "The vector h = φ(x), often written Z, that a network computes from an input x at some internal layer, and from which the final layers compute a score R or prediction Ŷ. The encoder φ is learned during training. It is also called an embedding.",
   "chapters": [
    53,
    65
   ]
  },
  {
   "term": "Representation bias",
   "definition": "Bias that arises when the development sample underrepresents part of the population, so that the model generalizes poorly for a subset of the use population (Suresh and Guttag, 2021). When it is caused by a limited or uneven sampling method, statisticians call it sampling bias.",
   "chapters": [
    25,
    38
   ]
  },
  {
   "term": "Representation disparity",
   "definition": "The phenomenon of low overall risk together with high risk for a group with a small population share, which results from optimizing for average loss (Hashimoto et al., 2018).",
   "chapters": [
    26,
    114
   ]
  },
  {
   "term": "Representation disparity and disparity amplification",
   "definition": "Representation disparity is low overall risk together with high risk for a small group. Disparity amplification is its growth over time when users who suffer high error leave a system, shrinking their group's share of the data (Hashimoto et al., 2018).",
   "chapters": [
    27
   ]
  },
  {
   "term": "Representation learning",
   "definition": "Training in which the network learns its own features rather than receiving them from a designer. The learned features abstract from surface inputs, can create correlations that are not visible in any single input, can encode protected attributes implicitly, and interact nonlinearly.",
   "chapters": [
    65
   ]
  },
  {
   "term": "Representation ratio",
   "definition": "The share of a group in a dataset divided by its share in a reference population, ρ_g = π_g / π*_g. A value of 1 means proportional representation; values below 1 indicate underrepresentation relative to the benchmark.",
   "chapters": [
    7,
    25,
    37
   ]
  },
  {
   "term": "Representational fairness",
   "definition": "The requirement that a learned representation carry little or no information about the protected attribute, so that any downstream model built on it is constrained in how unfair it can be.",
   "chapters": [
    53
   ]
  },
  {
   "term": "Representational harm",
   "definition": "A harm in which a system denigrates or subordinates a social group in how it represents it, independently of any allocation of resources: derogatory language, disparate system performance, erasure, exclusionary norms, misrepresentation, stereotyping and toxicity (Gallegos et al., 2024, Table 1).",
   "chapters": [
    112
   ]
  },
  {
   "term": "Representational harm and allocational harm",
   "definition": "A representational harm occurs when a system portrays a group in a demeaning, stereotyped or invisible way; an allocational harm occurs when a system distributes a resource or opportunity unequally. LLMs can cause both, often in the same output.",
   "chapters": [
    67
   ]
  },
  {
   "term": "Required status check",
   "definition": "A check, typically a continuous integration job, that must succeed before a pull request can be merged into a protected branch (GitHub, n.d.-a).",
   "chapters": [
    59
   ]
  },
  {
   "term": "Requirement-driven documentation",
   "definition": "Documentation designed from the specific obligations that apply to a system, so that each artifact exists because a requirement needs it and each requirement has corresponding evidence.",
   "chapters": [
    95
   ]
  },
  {
   "term": "Requirements inventory",
   "definition": "A consolidated list of every legal requirement that applies to a system across its deployment jurisdictions, with the provision it comes from and the principle it expresses.",
   "chapters": [
    85
   ]
  },
  {
   "term": "Resampling",
   "definition": "Building a new training set by duplicating some records (oversampling) and omitting others (undersampling), so that groups or group-outcome cells reach target sizes.",
   "chapters": [
    47
   ]
  },
  {
   "term": "Research agenda",
   "definition": "An ordered set of research questions with their rationale, methods and expected outputs, intended to coordinate work across researchers and institutions.",
   "chapters": [
    115
   ]
  },
  {
   "term": "Residual",
   "definition": "In regression, the difference between prediction and true value, Ŷ − Y.",
   "chapters": [
    40
   ]
  },
  {
   "term": "Residual risk",
   "definition": "Risk remaining after risk treatment (NIST, 2023, section 1.2.3, adopting the definition of ISO Guide 73).",
   "chapters": [
    90
   ]
  },
  {
   "term": "Resolving variable",
   "definition": "A variable influenced by A in a way that is accepted as non-discriminatory (Kilbertus et al., 2017).",
   "chapters": [
    18
   ]
  },
  {
   "term": "Reviewability",
   "definition": "A framework that treats automated decision-making as a socio-technical process from commissioning to investigation and determines which record-keeping and logging mechanisms are needed to permit meaningful review of individual decisions and of the process as a whole (Cobbe, Lee and Singh, 2021).",
   "chapters": [
    95
   ]
  },
  {
   "term": "Revocation",
   "definition": "A permanent withdrawal of a credential by its issuer before its validity period ends, because the issuer no longer stands by the claim or because the credential or its keys were compromised.",
   "chapters": [
    99
   ]
  },
  {
   "term": "Reweighing",
   "definition": "The specific scheme of Kamiran and Calders (2012) that assigns each combination of protected attribute value and class label the weight that would make attribute and label statistically independent in the weighted data. The spelling \"reweighing\" is the authors' own and is kept for this method.",
   "chapters": [
    47
   ]
  },
  {
   "term": "Reweighting",
   "definition": "Assigning instance weights that depend on group membership, outcome or both, so that the weighted training set has a chosen composition. The dataset's size and records stay the same.",
   "chapters": [
    47
   ]
  },
  {
   "term": "Risk",
   "definition": "The EU AI Act defines risk as \"the combination of the probability of an occurrence of harm and the severity of that harm\" (Regulation (EU) 2024/1689, Article 3(2)). The NIST AI Risk Management Framework uses a closely related notion: a composite measure of an event's probability of occurring and the magnitude or degree of its consequences (NIST, 2023, section 1.1).",
   "chapters": [
    90
   ]
  },
  {
   "term": "Risk-based regulation",
   "definition": "A regulatory strategy that sets priorities and allocates oversight according to an assessment of the risks that regulated activities pose to the regulator's objectives, so that obligations scale with risk and higher-risk systems face stricter requirements and oversight. Baldwin and Black (2016) show that identifying, selecting and prioritizing risks involves inescapably normative and political choices.",
   "chapters": [
    85,
    90
   ]
  },
  {
   "term": "Risk-calibrated governance",
   "definition": "Governance whose depth (number of gates, level of review, approval authority) is set by a documented assessment of the application's risk.",
   "chapters": [
    108
   ]
  },
  {
   "term": "Risk classification matrix",
   "definition": "The table in which each pattern is recorded with its affected components, its three ratings, its priority score and band, and the required action.",
   "chapters": [
    11
   ]
  },
  {
   "term": "Risk factor",
   "definition": "A dimension along which a system is assessed, such as domain impact, autonomy, decision impact, demographic risk or scale.",
   "chapters": [
    90
   ]
  },
  {
   "term": "Risk score",
   "definition": "A real-valued output R, often in [0, 1], intended to order people by the likelihood or size of an outcome. It may be used directly or thresholded into decisions.",
   "chapters": [
    21
   ]
  },
  {
   "term": "Risk tier",
   "definition": "A category of a risk classification scheme to which a defined set of obligations is attached, for example \"high-risk\" under the EU AI Act or impact level III under Canada's Directive on Automated Decision-Making.",
   "chapters": [
    90
   ]
  },
  {
   "term": "Risk tolerance",
   "definition": "An organization's or actor's readiness to bear risk in order to achieve its objectives; NIST (2023, section 1.2.2) stresses that the AI RMF does not prescribe a tolerance and that acceptable levels are highly contextual.",
   "chapters": [
    90
   ]
  },
  {
   "term": "Robustness",
   "definition": "(1) Of a fairness evaluation (Chapter 42), the stability and reliability of fairness metrics across data subsets, time periods and environmental conditions. (2) Of an explanation (Chapter 77), also called stability or continuity, the property that similar inputs give rise to similar explanations (Alvarez-Melis and Jaakkola, 2018); its opposite is fragility.",
   "chapters": [
    42,
    77
   ]
  },
  {
   "term": "ROC curve",
   "definition": "Of group a: the set of points (FPR, TPR) traced by the thresholded score as the threshold varies, computed within group a (Fawcett, 2006).",
   "chapters": [
    15
   ]
  },
  {
   "term": "Root cause analysis",
   "definition": "The investigation that traces an incident to the mechanisms that produced it, structured in this compendium by the seven bias types and the life cycle stages of Chapter 35.",
   "chapters": [
    109
   ]
  },
  {
   "term": "Runaway feedback",
   "definition": "A feedback loop in which a system's estimates diverge from ground truth and reinforce themselves, so that the system's own predictions increasingly determine the data that confirm them (Ensign et al., 2018).",
   "chapters": [
    27
   ]
  },
  {
   "term": "Saddle point",
   "definition": "A pair of parameters and multipliers at which the Lagrangian is minimized over the parameters and maximized over the multipliers; it corresponds to a solution of the constrained problem under suitable conditions.",
   "chapters": [
    51
   ]
  },
  {
   "term": "SAFE",
   "definition": "The validant.ai structure for fairness-enhanced stories (Specific protected attributes, Actionable fairness definition, Feature integration points, Expected outcome measures).",
   "chapters": [
    103
   ]
  },
  {
   "term": "Sampling distribution and standard error",
   "definition": "The sampling distribution is the distribution an estimate would have over repeated samples of the same design. Its standard deviation is the standard error (SE), the basic measure of an estimate's precision.",
   "chapters": [
    41
   ]
  },
  {
   "term": "Sampling uncertainty",
   "definition": "The variation of an estimate across hypothetical repetitions of the sampling process. It is summarized by a standard error or an interval.",
   "chapters": [
    13
   ]
  },
  {
   "term": "Scaffolding attack",
   "definition": "A construction in which a biased classifier is wrapped so that its predictions on real inputs remain biased while the inputs used by perturbation-based explainers receive innocuous predictions, so that the explanations look innocuous (Slack et al., 2020).",
   "chapters": [
    82
   ]
  },
  {
   "term": "Scalarization",
   "definition": "Turning a vector of objectives into a single number to be optimized, for example a weighted sum. Each choice of the scalarization's parameters yields one candidate point on the front.",
   "chapters": [
    54
   ]
  },
  {
   "term": "Scope",
   "definition": "What the claim is about. In the validant.ai terminology this is the pointing of an assessment: which body (model, affected person, organization), which pathway (predictive, generative, agentic, multi-agent) and which audience the finding addresses (Glinz, 2026).",
   "chapters": [
    101
   ]
  },
  {
   "term": "Scope statement",
   "definition": "The part of a fairness finding that states what was measured, on which population and period, against which construct, and what the finding does not establish.",
   "chapters": [
    45
   ]
  },
  {
   "term": "Score (R)",
   "definition": "A real-valued output of a model, typically an estimate of the probability that Y = 1, such as a credit score or a risk score.",
   "chapters": [
    12
   ]
  },
  {
   "term": "Scorer",
   "definition": "In machine learning libraries, a callable that evaluates a fitted model on data, so that it can be used in cross-validation and model selection.",
   "chapters": [
    40
   ]
  },
  {
   "term": "Scrum",
   "definition": "A lightweight framework that helps people, teams and organizations generate value through adaptive solutions for complex problems (Schwaber and Sutherland, 2020).",
   "chapters": [
    102
   ]
  },
  {
   "term": "Scrum artifacts",
   "definition": "The Product Backlog, the Sprint Backlog and the Increment, each with a commitment (the Product Goal, the Sprint Goal and the Definition of Done).",
   "chapters": [
    102
   ]
  },
  {
   "term": "Scrum events",
   "definition": "The Sprint (a fixed-length event of one month or less that contains all other events), Sprint Planning, the Daily Scrum, the Sprint Review and the Sprint Retrospective. Practitioners commonly call the four events inside the Sprint \"ceremonies\"; this compendium uses both words.",
   "chapters": [
    102
   ]
  },
  {
   "term": "Scrum Team",
   "definition": "One Scrum Master, one Product Owner and Developers, with no sub-teams or hierarchies.",
   "chapters": [
    102
   ]
  },
  {
   "term": "Secondary criterion",
   "definition": "A criterion encouraged through regularization or held within a tolerance, without being enforced exactly.",
   "chapters": [
    22
   ]
  },
  {
   "term": "Secondary metric",
   "definition": "A metric that captures an additional fairness concern, such as the cost of a specific error type, and is reported and assessed alongside the primary metric.",
   "chapters": [
    44
   ]
  },
  {
   "term": "Sectoral regime",
   "definition": "The body of law and supervisory rules that governs a type of decision or service, such as consumer credit, employment, insurance, medical devices or social benefits, independently of whether AI is used.",
   "chapters": [
    91
   ]
  },
  {
   "term": "Seeing",
   "definition": "How well the assessor could observe the subject: the access tier, the evidence grade and the validity grade of a finding (Glinz, 2026). Chapter 99 treats validity; Chapter 101 treats access and evidence.",
   "chapters": [
    101
   ]
  },
  {
   "term": "Selection bias",
   "definition": "A form of representation bias in which inclusion in the data depends on a variable related to the outcome or to the protected attribute, for example when repayment is observed only for applicants who were approved. When the outcome is observed only after a favorable decision, this is the selective labels problem (Lakkaraju et al., 2017; Chapter 12).",
   "chapters": [
    25,
    30
   ]
  },
  {
   "term": "Selection decision record",
   "definition": "The document produced by the Fairness Definition Selection Method, stating the criteria selected, their formal definitions, the rationale, the criteria not satisfied, and the monitoring plan.",
   "chapters": [
    22
   ]
  },
  {
   "term": "Selection diagram",
   "definition": "A causal diagram augmented with selection nodes S that mark the variables whose mechanisms or background factors differ between source and target populations (Bareinboim and Pearl, 2016).",
   "chapters": [
    33
   ]
  },
  {
   "term": "Selection rate",
   "definition": "The probability of a positive prediction or decision within a group, SR_a = P(Ŷ = 1 | A = a).",
   "chapters": [
    12,
    14
   ]
  },
  {
   "term": "Selection rate and impact ratio",
   "definition": "In the New York City rule, the selection rate is the rate at which individuals in a category are selected to move forward or assigned a classification; the impact ratio is the selection rate for a category divided by the selection rate of the most selected category (or the analogous ratio of scoring rates) (6 RCNY § 5-300).",
   "chapters": [
    91
   ]
  },
  {
   "term": "Selective classification",
   "definition": "Classification with the option to abstain; coverage is the share of cases decided automatically and selective risk the error rate on those cases.",
   "chapters": [
    57
   ]
  },
  {
   "term": "Selective disclosure",
   "definition": "The ability of a holder to make fine-grained decisions about what information to share. Unlinkable disclosure is selective disclosure in which presentations cannot be correlated between verifiers.",
   "chapters": [
    97
   ]
  },
  {
   "term": "Selective labels",
   "definition": "The situation in which the true outcome Y is observed only for subjects who received a particular decision, for example repayment is observed only for approved applicants.",
   "chapters": [
    12
   ]
  },
  {
   "term": "Selective optimization",
   "definition": "The design, tuning and testing of a technology for the needs, bodies or contexts of dominant groups, which produces systematically worse performance for others.",
   "chapters": [
    6
   ]
  },
  {
   "term": "Self-explaining model",
   "definition": "A model that itself generates local explanations without necessarily being directly interpretable, for example by generating rationales in text (Arya et al., 2019).",
   "chapters": [
    72
   ]
  },
  {
   "term": "Self-explanation",
   "definition": "An explanation that a model generates about its own output, such as a rationale, a list of important words or a counterfactual (Madsen et al., 2024).",
   "chapters": [
    80
   ]
  },
  {
   "term": "Sensitive attribute",
   "definition": "In this compendium, an umbrella term for a personal characteristic that is a protected ground, a special category of data, or both. In the fairness literature the terms sensitive attribute and protected attribute are often used interchangeably; the compendium keeps protected attribute A for the variable on which a fairness criterion is evaluated (Chapter 12).",
   "chapters": [
    24
   ]
  },
  {
   "term": "Sensitive personal data",
   "definition": "Data on religious, philosophical, political or trade union views or activities; on health, the private sphere or affiliation to a race or ethnicity; genetic data; biometric data that uniquely identifies a person; data on administrative and criminal proceedings or sanctions; and data on social assistance measures (Article 5(c) FADP). The Swiss list is wider than the GDPR's special categories in that it includes criminal and social assistance data.",
   "chapters": [
    87
   ]
  },
  {
   "term": "Sensitivity",
   "definition": "The largest change in an explanation within a neighborhood of an input (Yeh et al., 2019).",
   "chapters": [
    77
   ]
  },
  {
   "term": "Sensitivity analysis",
   "definition": "An analysis of how much a causal conclusion would change under specified departures from its assumptions, in particular unmeasured confounding.",
   "chapters": [
    31,
    33
   ]
  },
  {
   "term": "Separation",
   "definition": "The prediction (or score) is independent of the protected attribute conditional on the true outcome.",
   "chapters": [
    15
   ]
  },
  {
   "term": "Sequential test",
   "definition": "A test that processes observations one at a time and decides, after each, whether to stop with an alarm or continue; examples are the sequential probability ratio test (SPRT) and the cumulative sum (CUSUM) chart.",
   "chapters": [
    62
   ]
  },
  {
   "term": "Sequential testing",
   "definition": "Analysis that monitors results as they accumulate and allows early stopping while controlling error rates.",
   "chapters": [
    64
   ]
  },
  {
   "term": "Serious incident",
   "definition": "Under Article 3, point (49), of the EU AI Act, an incident or malfunctioning of an AI system that directly or indirectly leads to the death of a person or serious harm to a person's health; a serious and irreversible disruption of critical infrastructure; the infringement of obligations under Union law intended to protect fundamental rights; or serious harm to property or the environment (Regulation (EU) 2024/1689). Point (c) brings discriminatory outcomes within the scope of the term.",
   "chapters": [
    109
   ]
  },
  {
   "term": "Severity class",
   "definition": "A category of fairness issue (for example critical, major, minor) that determines the escalation route and response time.",
   "chapters": [
    108
   ]
  },
  {
   "term": "Severity (S)",
   "definition": "The impact of the pattern if it were perpetuated by the system, on a three-level scale.",
   "chapters": [
    11
   ]
  },
  {
   "term": "Side-warrant",
   "definition": "In CAE, the explicit statement of the property that makes an argument step valid, for example that a property distributes over the parts of a system; the side-warrant is itself a claim that can be challenged and supported (Bloomfield and Netkachova, 2014).",
   "chapters": [
    94
   ]
  },
  {
   "term": "Signed gap, absolute gap and ratio",
   "definition": "For a rate M, the signed gap M_b − M_a shows direction; the absolute gap |M_b − M_a| shows size only; the ratio M_b / M_a shows relative size. They answer different questions and can lead to different conclusions (Chapter 14).",
   "chapters": [
    39
   ]
  },
  {
   "term": "Similarity-based fairness",
   "definition": "Individual fairness expressed by a Lipschitz condition.",
   "chapters": [
    17
   ]
  },
  {
   "term": "Simpson's paradox",
   "definition": "A reversal in the direction of an association between two variables when data are aggregated over a third variable, so that a relationship holding within every stratum is reversed in the pooled data.",
   "chapters": [
    29
   ]
  },
  {
   "term": "Single-axis analysis",
   "definition": "An analysis of discrimination that considers one protected attribute at a time. Crenshaw (1989) used the term \"single-axis framework\" for the dominant approach of anti-discrimination doctrine.",
   "chapters": [
    29
   ]
  },
  {
   "term": "Small-cell suppression",
   "definition": "Not reporting a statistic for a group below a minimum size.",
   "chapters": [
    63
   ]
  },
  {
   "term": "Small sample problem",
   "definition": "The statistical instability of fairness estimates for intersectional subgroups that contain few individuals, which makes it hard to distinguish genuine disparities from noise.",
   "chapters": [
    29
   ]
  },
  {
   "term": "Smooth surrogate",
   "definition": "A differentiable function that approximates a non-differentiable quantity, such as a sigmoid in place of the indicator 1[R ≥ t].",
   "chapters": [
    52
   ]
  },
  {
   "term": "Smoothing (borrowing strength)",
   "definition": "Estimation methods that stabilize a small subgroup's estimate by combining its own data with information from a prior distribution or from related subgroups.",
   "chapters": [
    29
   ]
  },
  {
   "term": "Social burden",
   "definition": "The cost that individuals must incur to obtain a favorable decision under a classifier designed to be robust to strategic behavior (Milli et al., 2019).",
   "chapters": [
    114
   ]
  },
  {
   "term": "Social group and protected attribute",
   "definition": "A social group is a subset of the population that shares an identity trait, which may be fixed, contextual or socially constructed; the protected attribute is the trait that determines membership (Gallegos et al., 2024, Definitions 3 and 4). This matches the compendium's A (Chapter 12), with the caution that in text the attribute is usually signaled indirectly, through a name, a pronoun, a dialect or a described appearance.",
   "chapters": [
    112
   ]
  },
  {
   "term": "Social stratification by technology",
   "definition": "The role of a technology in distributing benefits and burdens across social groups, and thereby maintaining or challenging existing hierarchies.",
   "chapters": [
    6
   ]
  },
  {
   "term": "Social transparency",
   "definition": "A perspective that incorporates the socio-organizational context, such as how other people used and judged the system's outputs, into explaining AI-mediated decisions (Ehsan et al., 2021).",
   "chapters": [
    78
   ]
  },
  {
   "term": "Sociotechnical system",
   "definition": "A system whose behavior depends jointly on technical components (data, models, software) and on the people, organizations, institutions, rules, practices and values around them. Selbst et al. (2019) argue that fairness-aware machine learning fails when it abstracts away this social context.",
   "chapters": [
    6,
    28,
    115
   ]
  },
  {
   "term": "Source attribution (citation)",
   "definition": "A link from a statement in generated text to a document that supports it. Citation recall is the share of statements that are fully supported by their citations; citation precision is the share of citations that support their associated statement (Liu et al., 2023).",
   "chapters": [
    80
   ]
  },
  {
   "term": "Sparse data paradox",
   "definition": "The groups most exposed to harm, especially intersectional ones, often have the least data, so their metrics are the noisiest and the hardest to monitor reliably.",
   "chapters": [
    62
   ]
  },
  {
   "term": "Sparse subgroup",
   "definition": "A group, often an intersection of several protected attributes, whose denominator for a metric of interest is too small for the metric to be estimated with the precision the analysis plan requires.",
   "chapters": [
    43
   ]
  },
  {
   "term": "Special categories of personal data",
   "definition": "Personal data revealing racial or ethnic origin, political opinions, religious or philosophical beliefs or trade union membership, and genetic data, biometric data for unique identification, health data and data concerning sex life or sexual orientation (Article 9(1) GDPR). Their processing is prohibited unless an exception in Article 9(2) applies. The list overlaps with, but is not the same as, the protected grounds of anti-discrimination law.",
   "chapters": [
    24,
    83,
    86
   ]
  },
  {
   "term": "Special (temporary special) measures",
   "definition": "Measures taken to accelerate de facto equality for a disadvantaged group, which the conventions declare not to be discrimination provided they are temporary and do not maintain separate standards.",
   "chapters": [
    84
   ]
  },
  {
   "term": "Specialized toolkit",
   "definition": "A library that concentrates on doing one part of the fairness lifecycle well, for example bias auditing or algorithmic mitigation, rather than covering the whole lifecycle.",
   "chapters": [
    58
   ]
  },
  {
   "term": "Sprint Backlog",
   "definition": "The Sprint Goal, the Product Backlog items selected for the Sprint, and the plan for delivering them.",
   "chapters": [
    104
   ]
  },
  {
   "term": "Sprint Planning",
   "definition": "The event that initiates the Sprint by laying out the work to be performed. It addresses three topics: why the Sprint is valuable, what can be done this Sprint, and how the chosen work will get done. It is timeboxed to a maximum of eight hours for a one-month Sprint (Schwaber and Sutherland, 2020).",
   "chapters": [
    104
   ]
  },
  {
   "term": "Sprint Retrospective",
   "definition": "The event whose purpose is to plan ways to increase quality and effectiveness; the team inspects individuals, interactions, processes, tools and its Definition of Done. It is timeboxed to a maximum of three hours for a one-month Sprint (Schwaber and Sutherland, 2020).",
   "chapters": [
    105
   ]
  },
  {
   "term": "Sprint Review",
   "definition": "The event whose purpose is to inspect the outcome of the Sprint and determine future adaptations; the Scrum Team presents the results of its work to key stakeholders. It is timeboxed to a maximum of four hours for a one-month Sprint (Schwaber and Sutherland, 2020).",
   "chapters": [
    105
   ]
  },
  {
   "term": "Spurious association",
   "definition": "Association between A and the decision through common causes rather than through any causal effect of A.",
   "chapters": [
    18
   ]
  },
  {
   "term": "Spurious mechanism",
   "definition": "Association between A and the decision through common causes (back-door paths), with no causal effect of A (Zhang and Bareinboim, 2018).",
   "chapters": [
    30
   ]
  },
  {
   "term": "Stakeholder",
   "definition": "Any person or group affected by, involved in or responsible for a decision system: developers, operators, decision subjects, regulators and the wider public. In recommender systems (Chapter 66), the stakeholders are the users who receive recommendations, the providers whose items are recommended and the platform operator; Burke (2017) calls fairness for consumers C-fairness, for providers P-fairness, and for both CP-fairness.",
   "chapters": [
    3,
    66
   ]
  },
  {
   "term": "Standalone audit tool",
   "definition": "An integration pattern in which a tool takes a trained model's predictions (and labels and group attributes) as input and produces a fairness report, separately from model development.",
   "chapters": [
    58
   ]
  },
  {
   "term": "Statement of applicability",
   "definition": "In ISO/IEC 42001 (clause 3.26), the documentation of all necessary controls and the justification for including or excluding controls.",
   "chapters": [
    89
   ]
  },
  {
   "term": "Static and interactive explanation",
   "definition": "A static explanation does not change in response to feedback from the consumer; an interactive explanation lets the consumer drill down or ask for different kinds of explanation, for example through dialog (Arya et al., 2019).",
   "chapters": [
    72
   ]
  },
  {
   "term": "Static fairness evaluation",
   "definition": "Computing fairness metrics once, on a held-out data set, before deployment. It establishes a baseline but says nothing about later behavior.",
   "chapters": [
    61
   ]
  },
  {
   "term": "Statistical disparity analysis",
   "definition": "The use of hypothesis tests and effect sizes to judge whether an observed difference between groups in a distribution, data quality measure or outcome is larger than chance variation would explain.",
   "chapters": [
    38
   ]
  },
  {
   "term": "Statistical parity difference (SPD)",
   "definition": "The difference between two groups' selection rates.",
   "chapters": [
    14
   ]
  },
  {
   "term": "Statistical power",
   "definition": "The probability that a test detects a difference of a given size when that difference truly exists.",
   "chapters": [
    38
   ]
  },
  {
   "term": "Status list",
   "definition": "A compressed array in which each issued credential owns an index, and the value stored at that index encodes the credential's current status. The status list is itself signed and published, so that verifiers can fetch the status of many credentials at once (Looker et al., 2026; Sporny et al., 2025b).",
   "chapters": [
    99
   ]
  },
  {
   "term": "Strategic classification",
   "definition": "A setting in which the individuals being classified know something about the classifier and change their features to obtain a better outcome, which can degrade the classifier's performance (Hardt et al., 2016).",
   "chapters": [
    114
   ]
  },
  {
   "term": "Strategic ignorance",
   "definition": "Used in Chapter 7 for systematic, non-random patterns in which populations, variables or phenomena are under-recorded or not recorded at all, as distinct from random technical gaps. The term emphasizes that what is not measured often reflects priorities and power.",
   "chapters": [
    7
   ]
  },
  {
   "term": "Stratified bootstrap",
   "definition": "A bootstrap in which resampling is done separately within each group, so that every bootstrap sample keeps the observed group sizes.",
   "chapters": [
    41
   ]
  },
  {
   "term": "Stratified evaluation",
   "definition": "Reporting performance separately for each combination of group and condition, rather than one aggregate figure.",
   "chapters": [
    68
   ]
  },
  {
   "term": "Stratified randomization",
   "definition": "Randomization carried out separately within strata (for example, demographic groups or city types), which guarantees balance on the stratifying variables.",
   "chapters": [
    64
   ]
  },
  {
   "term": "Strong and weak null hypothesis",
   "definition": "The strong (or sharp) null states that the whole joint distribution of the data is the same in both groups; the weak null states only that a chosen parameter of the distributions, such as a true positive rate, is equal (DiCiccio, Vasudevan, Basu, Kenthapadi and Agarwal, 2020).",
   "chapters": [
    42
   ]
  },
  {
   "term": "Structural causal model (SCM)",
   "definition": "A model consisting of exogenous (background) variables U, endogenous (observed) variables V, a set of structural equations F that determine each endogenous variable from its parents and its own background variables, and a probability distribution P(U) over the exogenous variables (Pearl, 2009).",
   "chapters": [
    18,
    32,
    75
   ]
  },
  {
   "term": "Structural equation",
   "definition": "An equation V = f_V(parents of V, ε_V) stating how a variable is determined by its direct causes and a background disturbance (Pearl, 1995).",
   "chapters": [
    31
   ]
  },
  {
   "term": "Studentized statistic",
   "definition": "A statistic divided by an estimate of its standard error, so that its sampling distribution does not depend on unknown variances.",
   "chapters": [
    42
   ]
  },
  {
   "term": "Subgroup fairness",
   "definition": "Fairness evaluated over many structured subgroups (for example intersections of attributes) rather than over a few marginal groups (Kearns et al., 2018).",
   "chapters": [
    8
   ]
  },
  {
   "term": "Subgroup (slice)",
   "definition": "A subset of the population defined by a predicate over features or protected attributes, for example \"women over 50 with less than two years of tenure\". When the predicate is a conjunction of simple conditions it is also called an itemset or a pattern.",
   "chapters": [
    76
   ]
  },
  {
   "term": "Subject",
   "definition": "A thing about which claims are made: a person, an organization, an animal, a product, a dataset or an AI system.",
   "chapters": [
    97
   ]
  },
  {
   "term": "Subpopulation shift",
   "definition": "Test domains are subpopulations seen during training, but their proportions change (Koh et al., 2021). For protected groups: P_T(A) ≠ P_S(A) while P(X, Y | A) is unchanged.",
   "chapters": [
    34
   ]
  },
  {
   "term": "Substantial modification",
   "definition": "In the EU AI Act, a change to an AI system after it is placed on the market or put into service which was not foreseen or planned in the initial conformity assessment and which affects compliance with the requirements for high-risk systems or modifies the intended purpose for which the system was assessed (Regulation (EU) 2024/1689, Art. 3(23)).",
   "chapters": [
    99
   ]
  },
  {
   "term": "Substantive equality",
   "definition": "Equality understood not only as equal treatment but as responsiveness to people who are disadvantaged, demeaned, excluded or ignored (Fredman, 2016).",
   "chapters": [
    5
   ]
  },
  {
   "term": "Sufficiency",
   "definition": "The view that what matters to justice is that everyone has enough, with special importance attached to bringing people above a threshold (Bidadanure and Axelsen, 2025). As a fairness criterion (Chapter 16), sufficiency is the requirement that the outcome be independent of the protected attribute conditional on the score.",
   "chapters": [
    5,
    16
   ]
  },
  {
   "term": "Surrogate model",
   "definition": "A second, usually directly interpretable model that approximates a more complex model (Arya et al., 2019). A global surrogate imitates the model everywhere; a local surrogate imitates it near one instance.",
   "chapters": [
    72
   ]
  },
  {
   "term": "Suspension",
   "definition": "A temporary invalidity that the issuer may reverse. The IETF Token Status List defines the status value SUSPENDED as \"temporarily invalid\", usually temporary, next to VALID and INVALID (Looker et al., 2026).",
   "chapters": [
    99
   ]
  },
  {
   "term": "Sycophancy",
   "definition": "The tendency of a model to give responses that match the user's stated views rather than accurate ones (Sharma et al., 2023).",
   "chapters": [
    67
   ]
  },
  {
   "term": "Synthetic bias injection",
   "definition": "Deliberately inserting known patterns of fairness drift into historical data to measure a detector's delay and false alarm rate.",
   "chapters": [
    62
   ]
  },
  {
   "term": "Synthetic data",
   "definition": "Records produced by a model of the data distribution rather than collected from real individuals.",
   "chapters": [
    50
   ]
  },
  {
   "term": "Synthetic oversampling",
   "definition": "Creating new artificial records for an underrepresented cell, for example by interpolation between neighbors as in SMOTE (Chawla et al., 2002). Generation of synthetic data is treated in full in Chapter 50.",
   "chapters": [
    47
   ]
  },
  {
   "term": "System-level and decision-level explanation",
   "definition": "A system-level (or general) explanation describes how a system works for everyone: its inputs, its logic, its purpose. A decision-level (or individual) explanation describes why a particular output was produced for a particular person. The distinction matters because some provisions are satisfied by the first and others require the second.",
   "chapters": [
    79
   ]
  },
  {
   "term": "Systemic bias",
   "definition": "In the NIST taxonomy, bias that results from the procedures and practices of institutions that advantage some social groups and disadvantage others, without requiring conscious prejudice; NIST notes it is also referred to as institutional or historical bias (Schwartz et al., 2022).",
   "chapters": [
    6
   ]
  },
  {
   "term": "Systemic discrimination",
   "definition": "In this compendium, a pattern of disadvantage for a protected group that is produced by the interaction of institutions, practices and historical conditions rather than by one identifiable decision. It is not a separate legal category in the EU equal treatment directives, which define only direct and indirect discrimination, harassment and instructions to discriminate; it is a descriptive category whose effects are assessed through the direct and indirect tests.",
   "chapters": [
    23
   ]
  },
  {
   "term": "Tamper-evident log",
   "definition": "A log constructed so that any later alteration or deletion of entries can be detected, for example by chaining cryptographic hashes of entries (Schneier and Kelsey, 1999).",
   "chapters": [
    95
   ]
  },
  {
   "term": "Target criterion",
   "definition": "The fairness criterion that an intervention is designed to satisfy, such as demographic parity (Chapter 14), equal opportunity or equalized odds (Chapter 15), or calibration within groups (Chapter 16).",
   "chapters": [
    46
   ]
  },
  {
   "term": "Target distribution",
   "definition": "The composition (of groups, or of group-outcome cells) that a weighting or sampling scheme is designed to produce. Choosing it is a normative and domain decision, not a statistical one.",
   "chapters": [
    47
   ]
  },
  {
   "term": "Target population, development sample, use population",
   "definition": "The target population is the set of people (or things) a dataset is meant to represent; the development sample is the data actually collected from it; and the use population is the population the deployed model will see. Bias can arise at each step from one to the next.",
   "chapters": [
    25
   ]
  },
  {
   "term": "Task-specific similarity metric (d_X)",
   "definition": "A function assigning a non-negative distance to each pair of individuals, small for pairs that should be treated alike for the task at hand. Dwork et al. (2012) require only symmetry, non-negativity and zero self-distance.",
   "chapters": [
    17
   ]
  },
  {
   "term": "Technical instantiation of historical bias",
   "definition": "The process by which a historical social pattern becomes a concrete computational property of a system, typically through the choice of target variable, the composition of training data, or the features made available.",
   "chapters": [
    8
   ]
  },
  {
   "term": "Technical Report (TR)",
   "definition": "An ISO or IEC deliverable of an informative kind. ISO/IEC TR 24027 is a Technical Report that \"articulates current best practices\" for detecting and treating bias, rather than setting requirements.",
   "chapters": [
    89
   ]
  },
  {
   "term": "Technological continuity of discrimination",
   "definition": "The tendency of a new technology to carry forward the discriminatory patterns of the practices and infrastructures it replaces, often while appearing neutral or progressive.",
   "chapters": [
    6
   ]
  },
  {
   "term": "Test-driven bias prevention",
   "definition": "The application of test-driven development (Beck, 2003), in which tests are written before the functional code, to fairness: fairness requirements are written as executable tests before data processing or model training begins.",
   "chapters": [
    60
   ]
  },
  {
   "term": "Test statistic and p-value",
   "definition": "A test statistic summarizes the evidence against H₀; the p-value is the probability, computed under H₀, of a statistic at least as extreme as the one observed.",
   "chapters": [
    42
   ]
  },
  {
   "term": "Testable implication",
   "definition": "A conditional independence, or another constraint on the observed distribution, that the model entails and the data can contradict.",
   "chapters": [
    31
   ]
  },
  {
   "term": "TEVV",
   "definition": "Test, evaluation, verification and validation, the family of activities through which the NIST AI Risk Management Framework expects AI risks and trustworthiness characteristics to be measured (National Institute of Standards and Technology [NIST], 2023).",
   "chapters": [
    93
   ]
  },
  {
   "term": "Text-to-image model",
   "definition": "A generative model that produces an image from a written description, for example a diffusion model.",
   "chapters": [
    69
   ]
  },
  {
   "term": "Third party (counterparty)",
   "definition": "A person or organization affected by an agent's actions who is not its principal, such as a job applicant contacted by a recruiting agent or a seller dealing with a purchasing agent.",
   "chapters": [
    113
   ]
  },
  {
   "term": "Threshold rule",
   "definition": "A decision rule that acts on everyone whose score exceeds a threshold; a group-specific threshold rule uses different thresholds for different groups.",
   "chapters": [
    20
   ]
  },
  {
   "term": "Tolerance",
   "definition": "The largest disparity in a metric that is accepted without further action, fixed in advance (in the analysis plan) from the application's requirements, the applicable law and stakeholder considerations, for example |M_b − M_a| ≤ ε for a difference or a bound such as 0.8 for a ratio. A tolerance is a policy choice, not a statistical result.",
   "chapters": [
    36,
    39,
    44,
    103
   ]
  },
  {
   "term": "Tool use and retrieval",
   "definition": "An agent calls external functions (search engines, calculators, calendars, databases) and may retrieve documents into its context before generating, as in retrieval-augmented generation (Lewis et al., 2020; Schick et al., 2023).",
   "chapters": [
    69
   ]
  },
  {
   "term": "Top-k ranking",
   "definition": "The selection and ordering of the k best candidates, as in a shortlist.",
   "chapters": [
    21
   ]
  },
  {
   "term": "Total variation (TV)",
   "definition": "The observed difference in the expected decision or score between two groups, the quantity most group metrics are built from (Zhang and Bareinboim, 2018).",
   "chapters": [
    30
   ]
  },
  {
   "term": "Traceability",
   "definition": "The property that every finding can be linked backward to the metric, criterion, bias source and historical pattern that motivated it, and forward to the decision, remedy and owner that respond to it.",
   "chapters": [
    110
   ]
  },
  {
   "term": "Traceability matrix",
   "definition": "An explicit, bidirectional mapping between requirements and the system components, tests and evidence artifacts that implement, verify and document them, showing which artifacts support each requirement and which requirements each artifact addresses.",
   "chapters": [
    85,
    95
   ]
  },
  {
   "term": "Trade-off",
   "definition": "A situation in which improving one desired property of a system requires worsening another.",
   "chapters": [
    20
   ]
  },
  {
   "term": "Trajectory",
   "definition": "The ordered sequence of observations, decisions and actions an agent takes in pursuing a goal, from the first instruction to the final state.",
   "chapters": [
    69,
    113
   ]
  },
  {
   "term": "Trajectory and action ledger",
   "definition": "The sequence of an agent's reasoning, tool calls, observations and actions in a task, and a record of the actions taken, which can be reviewed after the fact (Shavit et al., 2023).",
   "chapters": [
    80
   ]
  },
  {
   "term": "Transfer",
   "definition": "The use of a lesson learned in one domain for a system in another. Transfer is justified when the mechanism, not merely the surface, is shared.",
   "chapters": [
    111
   ]
  },
  {
   "term": "Transfer learning",
   "definition": "Reusing a model pretrained on one task or corpus as the starting point for another task, by fine-tuning all its weights, part of them, or small added modules (adapters).",
   "chapters": [
    65
   ]
  },
  {
   "term": "Transparency",
   "definition": "In the NIST usage, the extent to which information about an AI system and its outputs is available to individuals interacting with it (National Institute of Standards and Technology, 2023). In the machine learning literature, a model is transparent if it is understandable by itself (Barredo Arrieta et al., 2020); Lipton (2018) distinguishes three levels: simulatability, decomposability and algorithmic transparency.",
   "chapters": [
    10,
    70
   ]
  },
  {
   "term": "Transparency log",
   "definition": "An append-only, publicly auditable log built on a Merkle tree, whose operator can prove that an entry is included and that the log has only grown.",
   "chapters": [
    98
   ]
  },
  {
   "term": "Transportability",
   "definition": "The property that a causal effect learned in a source population can be computed for a target population from the available data, given a model of how the populations differ (Bareinboim and Pearl, 2016).",
   "chapters": [
    33
   ]
  },
  {
   "term": "Trimmed mean and Winsorized mean",
   "definition": "Means computed after removing (trimming) or capping (Winsorizing) a fixed proportion of the smallest and largest values.",
   "chapters": [
    42
   ]
  },
  {
   "term": "True negative rate (TNR)",
   "definition": "Also called specificity: 1 − FPR.",
   "chapters": [
    15
   ]
  },
  {
   "term": "True outcome (Y)",
   "definition": "The quantity the decision is meant to track, for example whether a loan is repaid, whether a candidate performs well, or whether a patient has a condition. In binary problems Y ∈ {0, 1}; the compendium codes Y = 1 as the outcome the decision process is looking for (repays, qualified, has the condition). Y is also called the target variable or label.",
   "chapters": [
    12
   ]
  },
  {
   "term": "True positive rate (TPR)",
   "definition": "Of group a, also called recall or sensitivity: TPR_a = P(Ŷ = 1 | Y = 1, A = a).",
   "chapters": [
    15,
    19
   ]
  },
  {
   "term": "Trust",
   "definition": "In the integrative model used throughout this compendium, trust is the willingness of a party to be vulnerable to the actions of another party, based on the expectation that the other will perform a particular action important to the trustor, irrespective of the ability to monitor or control that other party (Mayer et al., 1995). Trust is an attitude or intention, not yet an action.",
   "chapters": [
    1
   ]
  },
  {
   "term": "Trust anchor",
   "definition": "An entity, or its public key, that a relying party accepts without further proof and from which other trust is derived. OpenID Federation defines a Trust Anchor as an entity that represents a trusted third party (Hedberg et al., 2026).",
   "chapters": [
    100
   ]
  },
  {
   "term": "Trust chain",
   "definition": "A sequence of signed statements linking the subject of a credential (or the issuer of a credential) to a trust anchor. In OpenID Federation, a sequence of Entity Statements starting at the subject's Entity Configuration and ending in a Trust Anchor (Hedberg et al., 2026).",
   "chapters": [
    100
   ]
  },
  {
   "term": "Trust in automation",
   "definition": "The attitude that an agent will help achieve an individual's goals in a situation characterized by uncertainty and vulnerability (Lee and See, 2004). The agent may be automation or a person.",
   "chapters": [
    2
   ]
  },
  {
   "term": "Trust mark",
   "definition": "A statement of conformance to a well-scoped set of trust or interoperability requirements as determined by an accreditation authority (Hedberg et al., 2026). The eIDAS Regulation separately defines an EU trust mark that qualified trust service providers may display once their qualified status appears in a trusted list (Regulation (EU) No 910/2014, Art. 23).",
   "chapters": [
    100
   ]
  },
  {
   "term": "Trust registry",
   "definition": "A system that serves authoritative information about which entities are recognized or authorized within a trust domain. The Trust over IP Trust Registry Query Protocol describes a trust registry as not the ecosystem itself but a capability within it that serves as the authoritative source of its authorization state (Trust over IP Foundation, 2026). The same specification notes that trust registries are also known as trust lists.",
   "chapters": [
    100
   ]
  },
  {
   "term": "Trust theatre",
   "definition": "The presentation of assurance artifacts that look authoritative but do not support the claim they are taken to support, for example quoting a benchmark tuned on one kind of system as evidence about another (Glinz, 2026).",
   "chapters": [
    101
   ]
  },
  {
   "term": "Trusted list",
   "definition": "In EU law, the signed, machine-processable list that each member state publishes of the qualified trust service providers for which it is responsible and the qualified trust services they provide (Regulation (EU) No 910/2014, Art. 22).",
   "chapters": [
    100
   ]
  },
  {
   "term": "Trusted time-stamp",
   "definition": "A signed statement by a time-stamping authority that a datum existed before a particular time (Adams et al., 2001).",
   "chapters": [
    98
   ]
  },
  {
   "term": "Trusting behavior (risk taking in relationship)",
   "definition": "The behavioral manifestation of trust: actually assuming the risk, for example by delegating a task or disclosing sensitive information. Mayer et al. (1995) call it risk taking in relationship (RTR).",
   "chapters": [
    1
   ]
  },
  {
   "term": "Trustor and trustee",
   "definition": "The trustor is the party who trusts; the trustee is the party who is trusted. In AI settings the trustee may be a person, an organization, an institution or, in a qualified sense discussed in Chapter 2, a technical system.",
   "chapters": [
    1
   ]
  },
  {
   "term": "Trustworthiness",
   "definition": "The property of the trustee that makes trust appropriate. Philosophical accounts describe trust as an attitude and trustworthiness as a property, and they require a trustworthy party to be both competent and willing to do what is expected (McLeod, 2020). The integrative model decomposes perceived trustworthiness into ability, benevolence and integrity.",
   "chapters": [
    1
   ]
  },
  {
   "term": "Trustworthy AI",
   "definition": "In the European Commission's High-Level Expert Group guidelines, AI that is lawful, ethical and robust throughout its life cycle (High-Level Expert Group on Artificial Intelligence, 2019). The NIST AI Risk Management Framework instead lists characteristics of trustworthy AI systems (National Institute of Standards and Technology, 2023).",
   "chapters": [
    2
   ]
  },
  {
   "term": "Two-sample test",
   "definition": "A test of whether two samples come from the same distribution, such as the Kolmogorov-Smirnov (KS) test or a kernel test based on MMD (Chapter 61).",
   "chapters": [
    62
   ]
  },
  {
   "term": "Type I and Type II error",
   "definition": "A Type I error (false positive) declares a disparity where none exists; a Type II error (false negative) misses a disparity that exists. The significance level α is the tolerated probability of a Type I error in a single test.",
   "chapters": [
    41
   ]
  },
  {
   "term": "Ubuntu",
   "definition": "A relational Sub-Saharan African philosophy in which personhood is constituted through relationships with others rather than by individual rationality (Mhlambi, 2020).",
   "chapters": [
    4
   ]
  },
  {
   "term": "Unconditional and conditional metrics",
   "definition": "An unconditional metric (statistical parity) compares predictions without reference to the outcome. A conditional metric conditions on the true outcome (equal opportunity, equalized odds) or on the prediction (predictive parity).",
   "chapters": [
    39
   ]
  },
  {
   "term": "Undefined rate",
   "definition": "A rate whose denominator is zero, for example the true positive rate of a group with no members for whom Y = 1. It has no value and must be reported as missing (NaN), not as 0 or 1.",
   "chapters": [
    40
   ]
  },
  {
   "term": "Undeveloped element",
   "definition": "A goal or strategy that is intentionally not yet developed in the argument (ACWG, 2021); in ISO/IEC/IEEE 15026-2, a placeholder argument to be replaced by a concrete argument.",
   "chapters": [
    94
   ]
  },
  {
   "term": "UNESCO Recommendation",
   "definition": "An instrument adopted by UNESCO's General Conference in which Member States are invited to adopt measures to give effect to its principles. Like the OECD instrument, it is not a treaty.",
   "chapters": [
    89
   ]
  },
  {
   "term": "Unfair practice",
   "definition": "Under Section 5 of the FTC Act, an act or practice that causes or is likely to cause substantial injury to consumers that is not reasonably avoidable by them and not outweighed by countervailing benefits to consumers or competition (15 U.S.C. § 45(n)).",
   "chapters": [
    88
   ]
  },
  {
   "term": "Uniform sampling",
   "definition": "Resampling in which every record of a cell has the same chance of being duplicated or skipped (Kamiran and Calders, 2012).",
   "chapters": [
    47
   ]
  },
  {
   "term": "Unmeasured confounding",
   "definition": "Confounding by variables not recorded in the data, which adjustment for measured variables cannot remove.",
   "chapters": [
    33
   ]
  },
  {
   "term": "Unsupervised bias detection",
   "definition": "Methods that search for groups with systematically worse outcomes without predefined demographic labels, for example by clustering.",
   "chapters": [
    38
   ]
  },
  {
   "term": "Unwanted bias",
   "definition": "Bias that is not intended in the objectives of an AI system and that can lead to unfair results, as distinguished from desired bias that the system needs to fulfill its objectives (ISO/IEC TR 24027, Introduction and 5.1).",
   "chapters": [
    89
   ]
  },
  {
   "term": "Use, misuse, disuse and abuse",
   "definition": "Use is the voluntary activation or disengagement of automation by human operators; misuse is overreliance on automation; disuse is neglect or underutilization of automation; abuse is the automation of functions by designers and implementation by managers without due regard for the consequences for human performance (Parasuraman and Riley, 1997).",
   "chapters": [
    2
   ]
  },
  {
   "term": "User story",
   "definition": "A short description of a capability from the point of view of a user role (Cohn, 2004), commonly written in the role-feature-reason format \"As a [role], I want [functionality] so that [benefit]\", which originated at the company Connextra in 2001 and which the Agile Alliance glossary describes as the most commonly recommended aid for writing stories (Agile Alliance, n.d.).",
   "chapters": [
    102,
    103
   ]
  },
  {
   "term": "Utilitarianism",
   "definition": "The paradigm consequentialist theory, whose classic proponents were Bentham, Mill and Sidgwick: the right act or policy is the one that maximizes the net total of well-being (good over bad) for all affected, so that outcomes are ranked by the sum of individual well-being (Sinnott-Armstrong, 2023; Adler and Holtug, 2025).",
   "chapters": [
    3,
    4
   ]
  },
  {
   "term": "Utility",
   "definition": "The value of a decision policy to the decision maker or to society, for example expected profit in lending, expected public safety in pretrial release, or expected health benefit in care allocation.",
   "chapters": [
    20
   ]
  },
  {
   "term": "Utopia point",
   "definition": "The vector of the best values each objective can reach on its own. It is usually unattainable, since the objectives conflict, and serves as a reference for several methods.",
   "chapters": [
    54
   ]
  },
  {
   "term": "Validation",
   "definition": "The application of a verifier's business rules to decide whether a particular use of a verified credential is proper. The same specification treats the validity period, the status of a credential, its schema and its fitness for purpose as components of validation, performed after verification (Sporny et al., 2025a).",
   "chapters": [
    97,
    99
   ]
  },
  {
   "term": "Validity",
   "definition": "A counterfactual is valid if the model in fact assigns it the desired outcome (Verma et al., 2024).",
   "chapters": [
    74
   ]
  },
  {
   "term": "Validity grade",
   "definition": "A disclosure of how current a finding is: whether the assessed subject was pinned to a version, whether the finding carries an expiry, whether it is re-tested on schedule, and whether it is monitored in production. The validant.ai grading of V0 to V3 is introduced in Chapter 99 (Glinz, 2026).",
   "chapters": [
    99
   ]
  },
  {
   "term": "Validity period",
   "definition": "The interval during which an issuer states that the information in a credential is valid. In the W3C data model it is expressed by validFrom and validUntil; in JSON Web Tokens and SD-JWT credentials by claims such as exp (expiration time). If neither validFrom nor validUntil is present, the W3C data model treats the credential as valid indefinitely from its creation (Sporny et al., 2025a).",
   "chapters": [
    99
   ]
  },
  {
   "term": "Value function (set function)",
   "definition": "In Shapley-based methods, the function v(S) that gives the model's output when only the features in the set S are \"present\". How the absent features are filled in defines the method.",
   "chapters": [
    73
   ]
  },
  {
   "term": "Value pluralism",
   "definition": "The condition that stakeholders hold genuinely different and reasonable ethical views, so that consensus cannot be assumed.",
   "chapters": [
    4
   ]
  },
  {
   "term": "Velocity",
   "definition": "The amount of work, usually in story points, that a team completes per Sprint, used to forecast capacity.",
   "chapters": [
    104
   ]
  },
  {
   "term": "Verifiable credential (VC)",
   "definition": "A tamper-evident credential whose authorship can be cryptographically verified. The only difference between a credential and a verifiable credential is that the latter is secured by a cryptographic mechanism.",
   "chapters": [
    97
   ]
  },
  {
   "term": "Verifiable data registry",
   "definition": "In the W3C credentials model, a role a system performs by mediating the creation and verification of identifiers, verification material, credential schemas, revocation registries and similar data; all parties expect it to be tamper-evident and to be a correct record of which data is controlled by which entities (Sporny et al., 2025a).",
   "chapters": [
    97,
    100
   ]
  },
  {
   "term": "Verification",
   "definition": "The evaluation of whether a credential or presentation is an authentic and current statement of its issuer or presenter: it conforms to the specification, its securing mechanism (the cryptographic proof) is satisfied and the credential has not been altered, and, where present, its status check succeeds. In the W3C Verifiable Credentials Data Model 2.0 a verifier checks, among other things, that the public key used is not suspended, revoked or expired (Sporny et al., 2025a).",
   "chapters": [
    97,
    99
   ]
  },
  {
   "term": "Virtue ethics",
   "definition": "An approach that makes the virtues, practical wisdom (phronesis) and human flourishing (eudaimonia) central, and addresses the questions of how one should live and what kind of person one should become (Hursthouse and Pettigrove, 2026).",
   "chapters": [
    4
   ]
  },
  {
   "term": "Vision model",
   "definition": "A model that takes images or video as input, typically a convolutional neural network (CNN) or a vision transformer, which extracts features in successive layers from low-level patterns (edges, textures, color) to abstract concepts.",
   "chapters": [
    68
   ]
  },
  {
   "term": "Warranted trust",
   "definition": "Trust that is justified or well grounded in the trustee's actual trustworthiness (McLeod, 2020). For AI, trust in a model that is caused by the model's trustworthiness, as opposed to unwarranted trust caused by other factors (Jacovi et al., 2021).",
   "chapters": [
    1,
    2
   ]
  },
  {
   "term": "Wasserstein barycenter",
   "definition": "The distribution that minimizes the weighted average of the squared 2-Wasserstein (optimal transport) distances to a set of distributions; for one-dimensional scores, its quantile function is the weighted average of the groups' quantile functions.",
   "chapters": [
    48,
    57
   ]
  },
  {
   "term": "Wasserstein distance",
   "definition": "The minimal transport cost between two distributions for cost |x − x′|^p, raised to the power 1/p (W_1 is also called the earth mover's distance).",
   "chapters": [
    48
   ]
  },
  {
   "term": "Weakly informative prior",
   "definition": "A prior that rules out implausible values and stabilizes estimates but is deliberately less informative than the available substantive knowledge would allow.",
   "chapters": [
    41
   ]
  },
  {
   "term": "Worst-group performance",
   "definition": "The lowest value of a performance metric across a predefined set of groups, the quantity that group distributionally robust optimization targets (Sagawa et al., 2020).",
   "chapters": [
    65
   ]
  },
  {
   "term": "Worst-group risk",
   "definition": "The largest group risk, R_max = max over k of R_k. Objectives that minimize it are called minimax or distributionally robust objectives.",
   "chapters": [
    26
   ]
  },
  {
   "term": "Wrapper pattern",
   "definition": "An integration pattern in which a fairness component wraps an existing estimator and adds fairness behavior without changes to the estimator's code.",
   "chapters": [
    58
   ]
  }
 ]
}