How to Conduct an AI Bias Audit

Begin with the decision, system boundary, and people affected. Identify what the model predicts, how a score changes a real decision, the deployment setting, and which group differences could cause harm. A bias audit is not just a single fairness metric: it should document data sources, selection stages, target definitions, model versions, decision thresholds, and human review. Choose metrics tied to the harm, such as selection rates, false-positive or false-negative rates, calibration, or error severity. Different metrics can conflict, so explain why the chosen measures fit the decision.

Disaggregate results by legally and contextually relevant groups, including intersections where sample size and privacy permit. Report counts and uncertainty, not only percentages. Examine data coverage, proxy features, missingness, and whether labels measure the intended construct. The 2019 Obermeyer study, for example, found a health-management algorithm that used cost as a proxy for need assigned lower risk scores to Black patients with comparable illness. An audit should test the target and process, not just the model’s final output.

For employment tools covered by New York City Local Law 144, employers and employment agencies must arrange an independent bias audit within one year before use, publish a summary, and provide required notices. That specific local law does not define every AI audit everywhere. Other laws and policies can require different methods or records. An auditor should disclose scope, data limitations, metrics, and whether the system was tested under conditions matching actual use.

Close the loop: identify findings, assign remediation owners, decide whether to alter data, model, threshold, or process, and retest before relying on the system. Keep a dated report and version history. A passing metric is not proof of fairness, legal compliance, or validity. If critical gaps remain, limit use or do not deploy until evidence and safeguards are adequate.