Responsible AI & Fairness

The Obligation Has Not Changed -- The Technology Has

Fair lending is not a new concept in banking. The Equal Credit Opportunity Act (ECOA) of 1974 and the Fair Housing Act (FHA) of 1968 have governed lending decisions for decades. Every banker understands that credit decisions cannot discriminate on the basis of race, color, religion, national origin, sex, marital status, age, or receipt of public assistance.

What has changed is the technology making those decisions -- or influencing them. When an AI system helps evaluate loan applications, generate customer communications, or prioritize collections outreach, it must comply with the same fair lending obligations that apply to human decision-makers. The technology is new, but the legal framework is not.

Types of Bias in AI Systems

Bias in AI is not always intentional -- in fact, the most dangerous forms of bias are the ones no one intended to create. Understanding the sources of bias is the first step toward detecting and mitigating it.

Training Data Bias

Foundation modelsFoundation ModelA large AI model trained on broad data that can be adapted to many tasks. Examples include GPT-4, Claude, and Gemini. Banks evaluate these for capabilities, safety, and regulatory fit.See glossary are trained on internet-scale data that reflects the biases present in human-generated text. If the training data contains stereotypical associations -- and it almost certainly does -- those associations become embedded in the model's behavior. An LLM may generate different language, tone, or recommendations when processing requests that vary only by names, locations, or other proxies for protected characteristics.

Historical Decision Bias

When AI systems are fine-tunedFine-TuningThe process of further training a pre-trained model on a specific dataset to specialize its behavior for a particular domain or task, such as banking compliance language.See glossary on historical banking data, they can learn and perpetuate patterns of past discrimination. If an institution's historical lending data reflects disparities -- even disparities that were legal at the time -- an AI trained on that data will reproduce those patterns. The model is not "biased" in the human sense; it is faithfully reflecting the statistical patterns in its training data, including patterns that reflect systemic inequality.

Proxy Discrimination

Even when protected characteristics are excluded from AI inputs, the model may use correlated variables as proxies. ZIP codes correlate with race. Names correlate with ethnicity. Employment patterns correlate with gender. An AI system that appears race-neutral in its inputs can still produce racially disparate outcomes through these proxy variables.

Representation Bias

If certain customer segments are underrepresented in training data, the AI will perform less accurately for those segments. A customer service AI trained primarily on interactions with affluent customers may provide lower-quality assistance to customers in underserved communities -- precisely the populations that fair lending laws are designed to protect.

Regulatory Expectations for AI Fairness

Banking regulators have made clear that existing supervisory expectations apply to AI and are issuing increasingly specific guidance.

SR 11-7: Model Risk Management

The Federal Reserve's SR 11-7 guidance applies to any model used in banking decisions, including AI models. Key requirements relevant to fairness include:

• Conceptual soundness: The model's design must be theoretically sound. For AI used in lending, this includes demonstrating that the model does not systematically disadvantage protected groups
• Outcomes analysis: Model validationModel Risk ManagementThe regulatory framework (OCC SR 11-7) governing how banks validate, monitor, and control AI models. Ensures models perform as expected and risks are identified and mitigated.See glossary must include analysis of outcomes across demographic groups to detect disparate impact
• Ongoing monitoring: Models must be monitored continuously for performance degradation and emerging bias

OCC Bulletin 2011-12: Sound Practices for Model Risk Management

The OCC's companion guidance reinforces that model risk management must address:

• Comprehensive testing across customer segments
• Documentation of model limitations and known risks
• Independent validation by parties not involved in model development

Fair Lending Examination Procedures

Federal banking examiners evaluate AI systems through the lens of established fair lending examination procedures:

• Comparative analysis: Do similarly situated applicants from different demographic groups receive similar outcomes?
• Regression analysis: After controlling for legitimate credit factors, do protected characteristics (or their proxies) predict different outcomes?
• Matched-pair testing: Do AI systems respond differently to otherwise identical scenarios when only demographic indicators change?

Emerging Regulatory Direction

The regulatory trajectory points toward more, not less, scrutiny of AI in banking:

• The CFPB has signaled that adverse action notices must explain AI-driven decisions in terms consumers can understand
• Multiple state attorneys general have initiated investigations into AI bias in lending and insurance
• The EU AI Act classifies credit scoring and lending AI as "high risk," requiring conformity assessments, human oversight, and bias testing before deployment

Implementing Fairness in Practice

Pre-Deployment Bias Testing

Before any AI system influences customer-facing decisions, the institution must conduct comprehensive bias testing:

1. Define protected groups: Identify all protected characteristics relevant to your use case (race, ethnicity, gender, age, disability status, etc.)
2. Establish baseline metrics: What are the current approval rates, pricing outcomes, and service quality metrics across protected groups?
3. Test AI outputs: Run the AI system against a representative dataset and analyze outcomes across protected groups
4. Apply statistical tests: Use standard disparate impact ratios (the four-fifths rule), regression analysis, and matched-pair comparisons
5. Document findings: Maintain comprehensive documentation of testing methodology, results, and any identified disparities

Ongoing Monitoring

Bias testing is not a one-time activity. AI models can develop new biases over time as:

• Input data distributions shift
• The underlying model is updated by the provider
• User behavior patterns change
• New products or policies interact with the AI differently

Establish automated monitoring that continuously evaluates AI outcomes across protected groups and alerts the model risk team when disparities emerge or worsen.

Explainability Requirements

When AI influences decisions that affect consumers, the institution must be able to explain those decisions. For traditional models, explainability techniques (SHAP values, partial dependence plots) are well-established. For LLMs, explainability is more challenging but no less required:

• Retrieval-based explanations: If the AI uses RAG, document which sources informed the response
• Prompt logging: Maintain complete records of prompts and responses for audit purposes
• Decision decomposition: For complex AI workflows, document each step where the AI contributed to the decision
• Plain language explanations: Adverse action notices must explain AI-driven decisions in terms that consumers can understand -- "the model decided" is not an acceptable explanation

The Responsible AI Framework

A practical responsible AI framework for banking institutions should include: