Minimally discriminating AI applications — From legal duty to technical design choices (MIN-DISC)
We will connect legal concepts of non-discrimination with technical design choices of AI developers, for making sure that AI is compliant with non-discrimination laws. This will support developers in creating legally compliant AI applications, and it will help authorities, auditors, and citizens to evaluate AI applications for legal discrimination.
Beschreibung
The research on Algorithmic Fairness has started with focusing on technical issues (first wave, since ~2015): how to measure
unfairness of AI decision systems, and how to avoid unfairness measured this way. A second wave (since ~2020) integrated
philosophical aspects, e.g. answering questions like “What is fair under which conditions?”, “How to choose an appropriate
fairness metric?”.
Currently, new regulation on AI (most notably the EU AI Act), creates new questions, centered on “How to generate
compliance with existing non-discrimination laws?”. This legal question is the most pressing issue today for practitionners -
but current research lacks answers.
Our project focuses on legal issues in the European and Swiss context, based on the results of the first two waves and our
own research of the last 4 years. The goal is to develop (a) approaches for assessing legal compliance, and (b)
methodologies helping developers to achieve legal compliance. For many applications (e.g. in banking, insurance, public
services, etc.), non-discrimination is now legally required in EU, and similar regulation is expected for Switzerland soon.
However, only few and preliminary answers are currently available.
The project concentrates on one central aspect of the legal framework: Disparities, e.g. between men and women, are not
forbidden per se. Legally, they are not qualified as discrimination if they can be “objectively justified” - a fact that has been
largely ignored in the algorithmic fairness research so far. Since disparities are often necessary for achieving a good
performance, legally justifying them is key for any application. This, however, requires (among other less critical conditions) a
proof that there are no less discriminatory alternatives (LDAs) to achieve the goal.
In the proposed project, we focus on the question: How can developers make sure that there are no LDAs with similar
performance? How can auditors in a lawsuit prove that there are LDAs, and that a given solution is more discriminatory than
necessary? These are imperative legal questions, and at the same time hard technical questions.
Eckdaten
Projektleitung
Co-Projektleitung
Prof. Dr. Florent Thouvenin (Universität Zürich UZH)
Projektteam
Projektpartner
Universität Zürich UZH / Rechtswissenschaftliche Fakultät
Projektstatus
laufend, gestartet 05/2025
Institut/Zentrum
Institut für Data Science (IDS)
Drittmittelgeber
Digitalisierungsinitiative der Zürcher Hochschulen DIZH / DIZH Fellowship 2025
Projektvolumen
299'550 CHF