Disparate Impact
A legal concept: A seemingly neutral rule or practice that disproportionately negatively affects a protected group.
Disparate impact occurs when a neutral algorithm disadvantages protected groups – the 80% rule is the classic fairness test.
Explanation
The 80% Rule (Four-Fifths Rule): If the selection rate of a group is less than 80% of the highest group rate, potential disparate impact exists. In ML: When a model systematically produces different outcomes for protected groups.
Marketing Relevance
Critical for marketing AI: Ad targeting, credit scoring, recruiting tools – wherever algorithms distribute access or resources.
Example
An ad targeting algorithm shows housing ads less to minorities although ethnicity is not a feature – proxy variables cause disparate impact.
Common Pitfalls
Proxy variables (zip code, name) can cause disparate impact without explicit protected attributes. Hard to prove without disaggregated data.
Origin & History
The term originates from US employment law (Griggs v. Duke Power, 1971). The 80% rule was formalized by the EEOC. In AI context, disparate impact became critical through studies on Facebook ad targeting (2019) and Amazon recruiting (2018).
Comparisons & Differences
Disparate Impact vs. Disparate Treatment
Disparate treatment is intentional discrimination (e.g., explicit feature "gender"); Disparate impact is unintentional but systematic.
Disparate Impact vs. Demographic Parity
Demographic parity is a fairness metric; Disparate impact is a legal standard using the 80% threshold.