Two days at HSLU on “Trustworthy and Explainable Artificial Intelligence in Practice,” and one lesson above the rest: explainability and fairness are not separate concerns but two views of the same decision under audit.
Why this matters now
Many AI systems make consequential decisions without being able to state, in plain language, why. That opacity is a trust problem and, increasingly, a legal one. The EU AI Act requires high-risk systems to be transparent enough for a deployer to interpret their output, and grants affected people a right to a meaningful explanation; GDPR has pointed the same way for years. The distance between “the model scored you 0.31” and “here is why, and here is what would have changed the outcome” is the distance this course set out to close.
It is also the distance Validant.ai was built to close. We treat trust as architected infrastructure rather than a compliance afterthought.
“Trust is the reason, fairness is the proof, identity is the bearer, and explainability is how the proof is shown.”
Day one: from ethics to attributions
The programme began where trust begins, with ethics. Erika Borcel introduced digital ethics as the foundation of trust, moving from humans to data, from data to decisions, and into the harder questions of dual use and misuse. This is the load-bearing layer, not a preamble. A faithful SHAP plot attached to a system that should never have been built is still a failure of trust. The Iceberg framework, published at the Swiss Conference on Data Science 2026, makes the same point structurally: trust failures occur at the seams between the agency, engineering, governance, and institutional layers, and strength in three layers does not compensate for weakness in the fourth.
Luis Terán then set out explainable AI itself, what it is, why businesses need it, its legal implications, and its limitations, before moving into the model-agnostic methods that anchor the field, LIME and SHAP. These are precisely the methods Validant wrapped behind its vfairness library for the new XAI view. SHAP matters for a reason beyond local attributions: because SHAP values are additive and share the units of the model output, a group-fairness disparity can be decomposed into the sum of per-feature contributions. That is the mathematical bridge from “the model is unfair” to “this feature is driving the unfairness, and by this much,” and it is the basis of the claim that fairness and explainability are two views of one decision.
“A notebook that produces a SHAP plot is a demonstration. A system that produces a reproducible, audit-ready explanation for every decision is a product.”
Kenneth A. Bonfo led the hands-on session, working LIME and SHAP use cases in Google Colab. The emphasis on practice over slideware underlined a distinction that defines production work, and closing the gap between the two is most of the engineering and most of the value.
Day two: interpretability, validation, and the toolbox
Day two went deeper into foundations and industrial reality. Prof. José María Alonso Moral, Tenured Professor at the University of Santiago de Compostela, drew a distinction practitioners often blur: interpretability, explainability, and trustworthiness are not synonyms. He illustrated interpretability-by-design with fuzzy sets and systems, showed how interpretable systems can become self-explaining, and gave real weight to the validation of explanations through intrinsic and extrinsic methods, reproducibility, and user studies.
This mirrors the Validant approach. An explanation that is persuasive but unfaithful is worse than none, because it manufactures unearned confidence; that is why the XAI view ships faithfulness and stability diagnostics alongside attributions, and why every assessment writes a replayable audit artifact.
“Trust without verifiability is marketing.”
José A. Mancera closed with the industrial toolbox and best practice: InterpretML, ALIBI, DiCE, and ELI5, with a comparison guide and concrete guidelines for designing explainable machine learning systems. This is the landscape the Validant facade is built to navigate. Rather than betting on one library, vfairness composes best-of-breed tools behind a single interface, SHAP and LIME for attribution, DiCE for actionable counterfactuals, and others, so the method fits the model and the use case. One thread matched our own engineering closely: selecting an XAI toolkit is a licensing decision as much as a technical one, since several well-known toolkits have moved away from permissive open-source licenses, which matters for anything shipped commercially.
What it contributes to systems for trust and agency
Validant treats fairness, digital trust, self-sovereign identity, and now explainability as four views of one trust posture. Each is its own lens, kept separate for focus and retrieval precision, yet all resolve to one coherent assessment of a single audited system. The XAI view this course addressed lets a user run LIME, SHAP, and counterfactual analysis, decompose a fairness metric into the features that drive it, and produce explanations tiered to the audience: a plain-language sentence and a counterfactual for the affected person, a risk summary for the board, a reproducible evidence trail for the auditor, and the full diagnostic suite for the engineer.
The reason this matters for agency is direct. A person cannot meaningfully contest a decision they cannot see inside, and an organization cannot stand behind a decision it cannot explain.
“Explainability is what makes a decision contestable, and contestability is what keeps automated decision-making compatible with human agency rather than corrosive to it.”
That is the standard the EU AI Act is moving toward, with high-risk enforcement arriving in August 2026, and the standard Validant.ai builds to.
With thanks
Our thanks to the lecturers, Erika Borcel, Luis Terán, Kenneth A. Bonfo, José María Alonso Moral, and José A. Mancera, and to HSLU for a course that balanced ethics, theory, practice, and industrial reality, and that took the validation of explanations as seriously as their generation. For organizations preparing for the EU AI Act, or simply intent on making their AI decisions as explainable as they are accurate and as fair as they are explainable, that is the work we do at Validant.ai.
“Fairness is the proof. Explainability is how we show it.”
Two Views of One Decision
Preparing for the EU AI Act, or set on making your AI decisions as explainable as they are accurate and as fair as they are explainable? See an independent, evidenced read of what your system actually does, and how it explains itself.
ResearchOpen to readDigital Trust Is an Orbit, Not a Pillar
Trust is not one more pillar to stack. It is the orbit three bodies trace together: the model, the person and the organisation. Why the three-body problem is the honest metaphor for trustworthy AI, and how to tell where you are in the orbit.
Read
