Bias Audit

Before a model enters production, a bias audit acts as a mandatory conformity assessment under the EU AI Act. The audit quantifies disparate impact across protected attributes like race and gender.

• Validates equalized odds and statistical parity thresholds
• Generates immutable audit trail evidence for regulatory submission
• Blocks deployment if the Four-Fifths Rule is violated

The COMPAS algorithm audit serves as a canonical case study. A third-party bias audit revealed that the tool falsely flagged Black defendants as high-risk at nearly twice the rate of white defendants.

• Applied causal fairness analysis to isolate discriminatory paths
• Demonstrated failure of equalized odds in criminal justice
• Led to mandatory human-in-the-loop oversight protocols

The Gender Shades audit evaluated commercial classifiers from Microsoft, IBM, and Face++. It exposed that darker-skinned females had error rates up to 34.7% higher than lighter-skinned males.

• Utilized intersectional fairness methodology
• Tested representation bias in training data
• Triggered immediate model retraining and transparency disclosures

Bias audits in HR tech scan resume screening models for disparate impact against protected classes. An audit might reveal that a model systematically downgrades candidates from specific zip codes.

• Measures compliance with the Four-Fifths Rule
• Analyzes counterfactual fairness by swapping demographic indicators
• Provides algorithmic recourse instructions for rejected candidates

Financial regulators require bias audits to ensure credit scoring models do not engage in redlining by proxy. Audits test if statistical parity holds between majority and minority applicants with identical financial profiles.

• Detects historical bias embedded in repayment data
• Applies adversarial debiasing to blind the model to protected attributes
• Ensures compliance with Equal Credit Opportunity Act

A bias audit is not a one-time event. Continuous compliance monitoring pipelines run fairness metrics on live prediction logs to detect model drift that introduces new biases.

• Tracks accuracy-fairness trade-off fluctuations over time
• Triggers automated model rollback if fairness thresholds breach
• Logs decisions for data subject rights automation requests

The foundational study of designing systems that make decisions without unjustified discrimination. It provides the theoretical basis for why a bias audit is necessary, moving beyond mere accuracy to evaluate the ethical distribution of outcomes across protected attributes like race or gender.

A legal doctrine central to a bias audit. It identifies a facially neutral policy or algorithm that disproportionately harms a protected class. An audit quantifies this using the Four-Fifths Rule: a selection rate for any group less than 80% of the highest group's rate is a statistical red flag requiring a compelling business justification.

A structured transparency document that is the primary output of a bias audit. It reports the model's intended use, evaluation results, and ethical considerations. A complete model card must disclose disaggregated performance across intersectional demographic groups, turning the audit's findings into a public, accountable artifact.

A causal definition of fairness used in advanced audits. A decision for an individual is fair if it would remain the same in a counterfactual world where they belonged to a different demographic group. This approach uses Structural Causal Models to distinguish discriminatory path-specific effects from legitimate, non-discriminatory influences.