Disparate impact

In plain English

Imagine a warehouse that requires every applicant to be six feet tall, though the job involves no high shelves. The rule mentions no one's sex, yet it screens out far more women than men. That is disparate impact: a neutral rule with lopsided results and no good reason behind it.

AI makes this pattern common. A model may never see race or sex, yet lean on features that track them. Examples include ZIP code, employment gaps or the word "women's" on a résumé.

Example

Thornbury Staffing, a fictional US recruiter, uses a vendor's résumé-ranking model. A quarterly review finds the model advances 50% of male applicants and 35% of female applicants.

The ratio is 35 ÷ 50 = 70%, below the four-fifths threshold in 29 CFR 1607.4(D). That flags possible adverse impact, so Thornbury investigates.

It finds the model penalizes employment gaps longer than a year, which are more common among caregivers. Thornbury asks whether the gap feature is job-related. It tests a version without it, re-runs the numbers, documents the result and switches to the less discriminatory version. Thornbury stays liable as the employer, even though the vendor built the tool.

Why it matters

US employment. The Supreme Court recognized disparate impact in Griggs v. Duke Power (1971). Congress wrote it into Title VII in 1991. Once a plaintiff shows a practice causes disparate impact, the employer must prove it is job-related and consistent with business necessity. It still loses if it refused a less discriminatory alternative.

Housing and credit. The Fair Housing Act covers disparate impact (Inclusive Communities, 2015). The CFPB amended Regulation B to remove disparate impact liability under ECOA, effective 21 Jul 2026 and under court challenge. Intentional discrimination, including deliberate proxies, remains illegal.

Enforcement shift. Executive Order 14281 (April 2025) told agencies to deprioritize disparate impact. The EEOC closed its pending disparate impact investigations, and HUD proposed rescinding its rule in January 2026. Texas's TRAIGA says disparate impact alone does not show intent. The statutes and private suits remain.

EU. The Race Equality Directive defines indirect discrimination as an apparently neutral practice that puts a group at a particular disadvantage. It is lawful only if objectively justified by a legitimate aim, with appropriate and necessary means. High-risk providers must also examine training data for biases (EU AI Act Art. 10(2)(f)).

How the AIGP exam tests it

The BoK covers employment in lesson II.B.2 and credit, housing and insurance in lesson II.B.3.

Common traps:

• "Nobody intended it." Intent is the disparate treatment test. Impact needs none.
• "We removed the protected field." Proxies carry the same signal.
• "Below four-fifths means illegal." It flags a need to investigate.
• "The vendor is responsible." Contracts shift costs, not the employer's legal duties.
• "The EEOC withdrew its guidance." The statute did not change.

The tell: a neutral rule with lopsided outcomes is impact; a trait or proxy used on purpose is treatment. See disparate treatment vs disparate impact and algorithmic bias.