Disparate Impact Analysis
What you'll learn: How to quantify unequal outcomes across demographic groups using the "80% rule" to detect potential discrimination in AI systems.
What is Disparate Impact?
Disparate impact occurs when an AI system produces significantly different outcomes for different demographic groups, even without explicit discriminatory intent. Unlike overt bias, disparate impact focuses on outcomes rather than process—if one group systematically receives favorable results at a much lower rate than another, that's a red flag for unfairness.
The 80% Rule (Four-Fifths Rule)
The classic test comes from employment law: compute the selection rate (ratio of favorable outcomes to total cases) for each group, then compare them. If the protected group's rate is less than 80% of the privileged group's rate, disparate impact may exist.
Formula:
Disparate Impact Ratio = (Selection Rate for Group A) / (Selection Rate for Group B)
If the ratio < 0.8, investigate further.
Example Without Code
Imagine a loan approval AI:
- Group A (protected): 60 approvals out of 100 applications = 60% approval rate
- Group B (reference): 90 approvals out of 100 applications = 90% approval rate
Disparate Impact Ratio = 60% ÷ 90% = 0.67
Since 0.67 < 0.8, this system shows disparate impact against Group A—they're approved at only 67% the rate of Group B, well below the 80% threshold.
Why This Matters
Disparate impact analysis is legally significant and ethically crucial. It provides a clear, quantifiable signal that your AI may be perpetuating systemic inequalities, even if individual predictions seem "fair." It helps you identify which groups are being harmed before deployment causes real-world damage.
Key Takeaway: Disparate impact analysis computes outcome ratios between groups—if the protected group receives favorable outcomes at less than 80% the rate of the reference group, your system likely exhibits problematic bias requiring investigation and mitigation.