Bias Audit

The audit's foundation is the explicit identification of protected attributes relevant to the system's context and jurisdiction (e.g., race, gender, age under GDPR, EU AI Act). This involves:

• Mapping applicable anti-discrimination laws (e.g., U.S. Equal Credit Opportunity Act).
• Identifying proxy variables (e.g., zip code, surname) that may serve as indirect substitutes.
• Documenting how attributes are constructed (self-identification, inference) and their granularity.

Aggregate performance metrics mask disparities. This component involves slicing evaluation data by protected groups to calculate fairness metrics. Critical activities include:

• Computing subgroup-specific accuracy, precision, recall, F1, and false positive/negative rates.
• Conducting intersectional analysis across combinations of attributes (e.g., Black women aged 25-34).
• Comparing these metrics to identify disparate impact (statistical differences in outcome rates) and disparate treatment (differences in error rates).

Models amplify biases present in their training corpus. This phase audits the data pipeline for:

• Representation bias: Checking for under/over-representation of subgroups in datasets.
• Historical bias: Analyzing whether labels reflect past societal inequities.
• Embedding bias: Using tests like the Word Embedding Association Test (WEAT) to quantify stereotypes in model representations.
• Label bias: Assessing if human annotators introduced subjective skew.

This advanced component tests for individual fairness by examining if a model's decision changes for a hypothetical individual when only a protected attribute is altered. It involves:

• Generating counterfactual examples (e.g., "Would this loan application be denied if the applicant's gender were male instead of female?").
• Evaluating adherence to counterfactual fairness, a causal criterion.
• Using techniques like causal graphs to model relationships between attributes, proxies, and outcomes.

Audits must assess the efficacy and cost of proposed bias mitigation techniques. This includes:

• Testing pre-processing (data reweighting, transformation), in-processing (adversarial debiasing, fairness constraints), and post-processing (threshold adjustment) methods.
• Quantifying the accuracy-fairness trade-off: documenting any reduction in overall model performance when fairness is improved.
• Validating that mitigation does not inadvertently harm other subgroups (fairness gerrymandering).

The final, critical component is producing auditable records for stakeholders and regulators. Outputs typically include:

• A detailed bias audit report with methodology, findings, and actionable recommendations.
• Model Cards or System Cards that transparently disclose performance across subgroups and known limitations.
• Evidence for an Algorithmic Impact Assessment (AIA).
• Documentation of fairness metric choices and their justification relative to the use case.

The overarching field of study and engineering practice focused on ensuring automated systems do not create unjust outcomes. It provides the ethical principles (e.g., equity, justice) that a bias audit operationalizes into quantitative metrics. Unlike a one-time audit, algorithmic fairness is a continuous design philosophy applied throughout the ML lifecycle.

A critical finding in a bias audit. It occurs when a model's outputs, while facially neutral in code, create a disproportionately adverse effect on a protected group. For example, a resume screening model that rejects 70% of applicants from Group A but only 30% from Group B exhibits disparate impact, even if 'group' is not an input feature. Audits measure this via metrics like demographic parity.

The quantitative tool of a bias audit. These are statistical measures used to assess equity across groups. Key types include:

• Group Fairness: Compares statistical rates (e.g., selection rate, error rate) between groups. Examples: Demographic Parity, Equal Opportunity.
• Individual Fairness: Requires similar individuals receive similar predictions.
• Counterfactual Fairness: A causal approach asking if the prediction would change if the protected attribute were different. Choosing the correct metric is a core audit design challenge, as they are often mutually exclusive.

The technical interventions applied after an audit identifies bias. Techniques are categorized by when they are applied in the ML pipeline:

• Pre-processing: Modifies the training data (e.g., reweighting samples, transforming features) to remove underlying skews.
• In-processing: Alters the training algorithm itself, often by adding a fairness constraint or using adversarial debiasing to penalize the model for learning biased representations.
• Post-processing: Adjusts the model's predictions after deployment (e.g., applying different decision thresholds per group) without retraining.

The core methodological practice of a rigorous audit. It involves slicing evaluation data and calculating performance metrics separately for each demographic subgroup (e.g., by gender, race, age).

• Subgroup Analysis reveals disparities masked by aggregate metrics (e.g., high overall accuracy but very low recall for one group).
• Intersectional Analysis examines subgroups at the intersection of multiple attributes (e.g., Black women over 50). This is crucial because bias is often compounded, not additive. Audits that fail to do this provide an incomplete picture.

A broader, often policy-driven governance process that incorporates a technical bias audit. An AIA is a structured evaluation of a system's potential risks, benefits, and societal impacts before deployment. It includes:

• Scoping the system's use and affected parties.
• Conducting the technical bias audit.
• Assessing broader implications (transparency, accountability, privacy).
• Documenting findings in artifacts like Model Cards. While a bias audit focuses on statistical discrimination, an AIA places those findings within a full risk-management framework.