Disparate Impact: When Fair Rules Aren't Fair

WHY IT EXISTS Disparate impact was established as a legal concept to address discrimination that isn't overt or intentional. It provides a way to challenge practices in areas like employment and housing that are fair on their face but systemically disadvantage a protected group in practice.

THE MENTAL MODEL A policy's fairness is judged by its results, not its intentions. A rule applied equally to everyone can still be discriminatory if it creates an unequal, adverse outcome for a group protected by law (based on race, sex, religion, etc.). Think of it as judging the effect, not the motive.

HOW IT WORKS Legally, a plaintiff first demonstrates that a practice has a disproportionately harmful effect on their protected group. The burden of proof then shifts to the defendant (e.g., an employer) to show the practice is a legitimate business necessity and directly related to the job or goal. If they can't, the practice is deemed discriminatory.

WHEN TO USE IT This concept is critical for auditing machine learning models in high-stakes domains like hiring, credit scoring, and housing. Data scientists use statistical tests for disparate impact to measure and mitigate algorithmic bias, ensuring their models don't illegally perpetuate or amplify societal inequities.

WHEN NOT TO USE IT The concept of disparate impact is distinct from "disparate treatment," which is intentional discrimination. Also, a practice with a disparate impact might be legally permissible if it can be proven to be a genuine business necessity for which no less-discriminatory alternative exists.

ONE CANONICAL EXAMPLE A tech company requires all applicants for a software engineering role to pass a physical fitness test. The test results show that female applicants pass at a significantly lower rate than male applicants. Even with no discriminatory intent, this practice has a disparate impact. Because physical fitness is not a business necessity for a programming job, the practice would likely be found discriminatory.