How do we detect and mitigate algorithmic bias?

Algorithmic bias occurs when an AI system produces outcomes that are systematically less favorable to individuals in protected classes, even if the protected characteristic is not an explicit input. This can occur through proxy variables (inputs such as zip code that closely track a protected trait), historical bias in training data, or model design choices that have disparate impact.

For AI used in hiring, lending, housing, or access to services, algorithmic bias creates exposure under Title VII, the Fair Credit Reporting Act (FCRA), the Equal Credit Opportunity Act, the Fair Housing Act, and their state equivalents. Regulators including the Equal Employment Opportunity Commission (EEOC), the Consumer Financial Protection Bureau (CFPB), and the Department of Housing and Urban Development (HUD) have all signaled active enforcement interest in AI-driven discrimination.

Bias testing should be conducted before deployment and at regular intervals thereafter. The standard approach is disparate impact analysis: measure the selection rate for each protected group and compare to the most favored group. A ratio below 0.8 (the "four-fifths rule") indicates potential adverse impact requiring further investigation.

Testing should cover protected characteristics including race, sex, age, national origin, religion, and disability status, as well as intersectional combinations. Use both the training data and real-world outputs for testing. A model that performs well on historical data may behave differently on the population it is actually deployed against.

For high-risk systems, consider engaging an independent third party to conduct bias audits. New York City Local Law 144 requires independent bias audits for automated tools used in hiring or promotion for jobs located in New York City, and similar requirements are likely to spread to other jurisdictions.

When bias is detected, mitigation options include retraining on more representative data, adjusting the score a case must reach to be approved, correcting outputs after the model produces them, or replacing the system. Document every mitigation step taken: what was found, what was changed, and what the outcome was after remediation.

Maintain an audit trail that allows you to reconstruct the testing history of any deployed system. If a claim is filed or a regulator investigates, you need to demonstrate that you tested for bias, found the results acceptable, and took action when they were not. Regulators and plaintiffs' attorneys read missing records as a sign that testing never happened.