How Discrimination Law Reaches AI
An AI model that screens job applicants does not get a discrimination exemption because it runs on a server. Discrimination law applies to the outcome, not the mechanism, and it applied long before anyone drafted an AI Act. This sits at the heart of one Domain II topic on the AIGP, where the Body of Knowledge, the official blueprint of everything the exam can test, asks how existing laws reach AI. The point candidates miss is plain: nondiscrimination law reaches AI on its own terms, with no need for a single AI-specific clause.
Discrimination law did not wait for AI
Anti-discrimination regimes predate machine learning by decades. In the EU, the Racial Equality Directive and the Employment Equality Directive protect against unequal treatment in decisions that affect a person, whoever or whatever makes them. In the United States, Title VII, the Equal Credit Opportunity Act and the Fair Housing Act do similar work across hiring, lending and housing. None of these discrimination laws mentions AI. None needs to. They regulate the result, so a discriminatory result from a model is one the law already covers.
The AI Act sits on top of this, not in place of it. An organisation that has satisfied its AI Act duties has not thereby satisfied discrimination law; the two ask different questions and carry separate liability. Our breakdown of how existing laws apply to AI sets out the wider pattern.
Direct and indirect discrimination
The exam wants you to separate two forms. Direct discrimination, or disparate treatment, is the obvious kind: a system uses a protected characteristic as an input and treats people worse for it. Indirect discrimination, or disparate impact, is subtler and far more common in AI. A neutral rule produces a worse outcome for a protected group, with no protected attribute in the data and no intent behind it.
Intent is the trap. Indirect discrimination needs no malice and no awareness; the outcome carries the breach. A hiring model trained on a firm's past hires can learn that firm's historical bias and repeat it, while its builders believe they have engineered the bias out.
Indirect discrimination is not automatically unlawful. EU equality law allows a justification: a measure with a disparate impact can stand if it pursues a legitimate aim and the means are proportionate and necessary. For an AI system that becomes a real burden. You must show the model serves a genuine business need and that no less discriminatory model would do the job. Accuracy alone will not clear that bar, and a cheaper, fairer alternative left untried is a weak position to defend.
Where the law bites
The Body of Knowledge names the contexts that matter, and they are the contexts where automated decisions now cluster: employment, credit, lending, housing and insurance. These are the decisions that hand someone a job, a loan, a tenancy or a premium. Each is a context the exam singles out because the harm is concrete and the paper trail is often thin. A scoring model in any of them is a candidate for legal challenge.
Insurance deserves a flag. Pricing models run on correlation, and correlation is exactly where indirect discrimination hides. A variable that predicts risk well may also track a protected characteristic, and discrimination law does not accept predictive accuracy as a defence to a discriminatory effect.
Proxies do the damage
Stripping protected attributes from the training data feels like a fix. It is not. Models rebuild those attributes from proxies: a postcode that tracks ethnicity, a spending pattern that tracks disability, a career gap that tracks pregnancy. The model never sees the protected class and discriminates against it regardless.
This is why testing for a discriminatory effect matters more than auditing the input list. You measure outcomes across groups, not just the features you fed in. Regulators and courts increasingly expect that outcome testing as a matter of course, not a courtesy. A model can be blind by construction and biased in result, which is the failure our note on trustworthy and ethical AI keeps returning to.
Why discrimination law trips AIGP candidates
Two assumptions cost marks. The first is that AI-specific law is the only law in play; candidates reach for the AI Act and forget the older duty beneath it. The second is that the absence of a protected attribute means the absence of discrimination. Both are wrong, and scenario questions are built to expose exactly these reflexes.
When a stem describes a lending or hiring model, the examiner is usually testing whether you spot the existing discrimination law obligation before you reach for anything newer. Name the older duty first, then layer the AI-specific rules on top. Governance professionals will recognise the move from impact-assessment work, where bias testing and documented outcome analysis are the evidence that the obligation was met rather than assumed.
The governance answer is not exotic. Document the lawful aim, test outcomes across protected groups before and after deployment, keep the evidence, and retest when the model is retrained. That record is what turns a defensible decision into a provable one if a regulator or a claimant ever asks.
Treat nondiscrimination law as a live obligation in every AI use case you assess, not a footnote to the AI Act. More AIGP breakdowns are at 22academy.com/study.
