What the law actually tests

Legal tests are not statistical definitions

Having spent three lessons on the mathematics, it is important to see that the law mostly does not use it. Legal systems test something narrower and older, and if you are deploying a system that affects people's opportunities, the legal test is the one that will be applied to you.

The central distinction in most anti-discrimination law is between two kinds of wrong.

Disparate treatment, or direct discrimination: treating someone worse because of a protected characteristic. Intent, or at least the explicit use of the characteristic, is central. Feeding caste or sex into a hiring model as a variable lands here.

Disparate impact, or indirect discrimination: a neutral rule that in practice disadvantages a protected group, and that cannot be justified as necessary. No intent required. This is the door almost every AI discrimination case walks through, because nobody puts the protected characteristic in the model, and the model reconstructs it anyway.

The four-fifths rule

The best-known operational test comes from the US Equal Employment Opportunity Commission's Uniform Guidelines, adopted in 1978, decades before any of this. It is arithmetic simple enough to do in your head.

Compute the selection rate for each group. Divide the lowest by the highest. If the ratio is below 0.8 — four-fifths — the guidelines treat that as evidence of adverse impact requiring justification.

Worked: 100 men apply, 40 are hired, so the selection rate is 40%. 100 women apply, 25 are hired, so 25%. The ratio is 25 ÷ 40 = 0.625, well below 0.8. That is a flag.

Three things to understand about this rule. It is a rule of thumb, not a law of nature — it appears in guidelines, courts treat it as one indicator among several, and with small numbers it triggers on noise. It is about outcomes, not mechanisms, so it does not care how your model works. And it is rebuttable: an employer can defend a disparity by showing the criterion is job-related and consistent with business necessity, at which point the argument moves to whether a less discriminatory alternative existed.

That last clause is where machine learning is exposed. If a slightly less accurate model produces a substantially smaller disparity, the existence of that alternative is legally relevant. Teams that never search for one cannot say they did.

Other jurisdictions, briefly

The shapes rhyme, the details differ, and this is not legal advice for any particular situation.

The UK Equality Act 2010 covers indirect discrimination with a proportionality defence. The EU has equal treatment directives with a similar structure, and now the AI Act layering safety and documentation duties on top of them. India approaches this through constitutional equality provisions and reservation policy rather than a general employment-discrimination statute with a private cause of action, so the practical route for an individual is different and often harder. Brazil, South Africa, Canada and others each have their own frameworks, and South Africa's employment equity regime is closer to an affirmative-action model than to the American disparate-impact one.

There is also a growing layer specific to automated decisions. New York City's Local Law 144, in force since 2023, requires an annual independent bias audit of automated employment decision tools, published, with selection-rate and impact-ratio figures by sex and race. Whatever you think of its scope, it is the first regime that makes the arithmetic in this module a published legal artefact.

The tension worth naming

Here is the awkward part. To check the four-fifths ratio you must know each applicant's group. But in several places, collecting that data is restricted, discouraged or culturally fraught, and in some legal readings using it to adjust a model is itself discrimination.

So organisations end up in a bind: legally exposed on outcomes they are discouraged from measuring. The usual resolution is to collect group data separately from the decision pipeline — held by a different team, used only for aggregate monitoring, never available to the model at decision time. It is a workable compromise and it requires deliberate design, because the default is to collect nothing and discover the problem from a lawsuit.

The practical minimum, if you deploy anything that selects people: compute the selection rate by group every quarter, keep the numbers, and keep the record of what you did when a ratio moved. That file is both your early warning and, if it ever comes to it, your evidence that you were looking.