Override

A process in which a human decision-maker rejects or modifies a model-generated recommendation, score, or decision, using their judgment or supplemental information.

Source: AIEOG AI Lexicon (Feb 2026), adapted from Model Risk Management, Comptroller’s Handbook

An override occurs when a human decision-maker chooses to deviate from what an AI model recommends. The model says one thing, and the human decides differently, either rejecting the recommendation entirely or modifying it based on information or context the model did not account for.

Overrides are a critical feature of responsible AI deployment. They represent the institution’s ability to exercise human judgment when the model’s output is not appropriate. A credit model might recommend denial, but a human reviewer might override based on compensating factors the model cannot assess.

Overrides become a governance concern when they are too frequent (suggesting the model is not performing well), too infrequent (suggesting automation bias), or systematically biased (suggesting human biases are overriding model objectivity).

Overrides are a regulatory focus area. Examiners assess override patterns as part of model risk management and fair lending examinations:

• Fair lending. Override patterns showing disparate treatment across protected classes create fair lending exposure.
• Model effectiveness. High override rates may indicate the model is not capturing relevant decision factors.
• Audit trail. Regulators expect institutions to document overrides, including rationale, approver, and outcome.
• Policy compliance. Override authority should be clearly defined in policy.

1. Establish override policies. Define who has override authority, what circumstances justify an override, what documentation is required, and what approval levels apply.
2. Track and analyze override patterns. Monitor override rates, reasons, approvers, and outcomes. Analyze for patterns that suggest bias or model weakness.
3. Document every override. Require written rationale for every override.
4. Test for fair lending impact. Analyze override patterns across demographic groups to identify potential disparate treatment.
5. Use override data for model improvement. Systematic override patterns can inform model recalibration.
6. Report on override metrics. Include override analysis in regular governance reporting.

Human-in-the-Loop (HITL), Guardrails, AI governance, Model risk, Explainability, Bias