AI and non-discrimination: equal-treatment law alongside the AI Act

Short answer: An AI system that systematically treats groups of people differently engages two regimes at once. The AI Act imposes requirements on data quality and bias control for high-risk systems, but the actual ban on discrimination comes from equal-treatment law — and that applies regardless of your system's risk level.

Two regimes side by side

The AI Act is product regulation: it sets out how to build and place a system on the market safely and carefully. Equal-treatment law — including the Racial Equality Directive (2000/43/EC) and national equal-treatment statutes — prohibits direct and indirect discrimination on grounds such as origin, sex, age or disability.

Those prohibitions applied before the AI Act and continue to apply in full. An employer using AI to screen applicants, or a bank scoring credit with a model, can therefore discriminate in the legal sense — regardless of whether the system is AI Act compliant.

Direct and indirect discrimination

Direct discrimination is rare in AI: you take a prohibited ground explicitly as a feature. Far greater is the risk of indirect discrimination: a seemingly neutral characteristic (postcode, language use, a gap in a CV) that in practice disadvantages a protected group.

Indirect discrimination is prohibited unless there is an objective justification that is proportionate and necessary. That the bias "came from the data" is no justification. The party responsible for the decision must be able to explain why a disadvantageous effect is justified.

What the AI Act does add here

For high-risk systems — recruitment, lending, education, social benefits — the AI Act requires representative datasets, examination of possible bias and human oversight, among other things. These are tools to prevent discrimination, but they do not replace the ban. See the overview of high-risk obligations.

The AI Act also bans some applications outright, such as certain forms of social scoring; see prohibited AI practices. And the GDPR sets its own limits on profiling and automated decisions about individuals.

Burden of proof and transparency

Equal-treatment law often shifts the burden of proof: a person who makes a disadvantage plausible forces the other party to prove there was no discrimination. With an opaque model that is hard — which is precisely why logging and explainability become evidentially important. Transparency towards data subjects may also flow from Article 50 of the AI Act.

What to do

• Test for indirect effects: measure outcomes per group, not just whether a prohibited ground sits in the model.
• Document a justification for any disadvantageous effect that remains — proportionate and necessary.
• Keep evidence: datasets, versions and decision logs determine whether you can rebut a presumption of discrimination.
• Assign responsibility to a person who can explain and correct the decision.
• Combine the regimes: AI Act compliance and an equal-treatment test, not one or the other.

Ticking off only the AI Act does not cover the discrimination risk. The ban applies independently of your system's risk level.