Disparate treatment vs disparate impact

Side by side

When each applies

Disparate treatment applies when someone chose to use a protected trait, or a stand-in for it. An explicit age cutoff coded into hiring software is treatment. So is deliberately using ZIP codes as a substitute for race. Automation does not dilute intent; it scales it.

Disparate impact applies when no one intended harm but the outcome is unequal. This is how AI usually discriminates. Models learn from biased history, rely on proxies such as commute distance or school, or perform worse for groups they rarely saw. Removing the protected trait from the inputs does not fix proxies.

Under Title VII, the impact framework has three steps. The plaintiff shows adverse impact from a particular practice. The employer proves the practice is job-related and consistent with business necessity. The plaintiff can still win by showing a less discriminatory alternative the employer refused to adopt. The four-fifths rule flags possible adverse impact when one group's selection rate is under 80% of the highest group's.

The 2025 to 2026 enforcement shift. Executive Order 14281 (23 April 2025) told federal agencies to deprioritize disparate-impact liability. The EEOC removed its AI guidance. The statute is unchanged, and private lawsuits continue. Texas's TRAIGA, in effect since 1 January 2026, says disparate impact alone does not show intent to discriminate under that law.

In the EU, Directive 2000/78/EC and its sister directives ban both direct and indirect discrimination. The EU AI Act adds a design duty: high-risk providers must examine training data for biases and act on them (Art. 10(2)(f) and (g)).

Worked example

Pellworth Logistics, a fictional US company, licenses a vendor's résumé-screening model. An internal audit finds two problems.

Problem one: impact. The model advances 45% of male applicants and 30% of female applicants. The ratio is 30 ÷ 45, or 67%, below four-fifths. Sex is not an input. The audit traces the gap to a feature that penalizes career breaks over a year. That is a classic proxy, and a disparate impact problem. Pellworth must show the feature is job-related, or adopt a less discriminatory alternative, such as dropping it.

Problem two: treatment. A hiring manager had asked the vendor to drop applicants who graduated before 1995. His stated reason: "to keep the team young." That is a deliberate proxy for age, used on purpose. It is disparate treatment under the Age Discrimination in Employment Act, which protects workers aged 40 and over.

The vendor contract says the vendor handles "all legal compliance." It does not matter. Employers answer for the selection tools they use. Pellworth removes the filter, retests the model by group and documents the fixes. An AI audit on a set schedule would have caught both problems earlier.

The exam trap

The classic trap is "the model never uses race or sex, so it can't discriminate." In fact, proxies produce disparate impact without the protected trait as an input.

The tells:

• An explicit rule or a proxy chosen on purpose points to treatment.
• A neutral rule with unequal results, and no proof of motive, points to impact.
• The four-fifths rule triggers investigation. It does not prove illegality.
• "The vendor built it" does not shift the employer's legal duty.
• Withdrawn agency guidance does not repeal a statute.

Equal opportunity, a fairness metric, compares recall across groups. See precision vs recall and algorithmic bias. Practice the hiring cases in lesson II.B.2 and credit and housing in lesson II.B.3.