Model integrity

The assurance that a model performs as intended, free from unauthorized modifications, tampering, or degradation, throughout its lifecycle.

Source: AIEOG AI Lexicon (Feb 2026), adapted from NIST AI 100-2e2025 and Model Risk Management, Comptroller’s Handbook

Model integrity is the confidence that an AI model is doing exactly what it was designed, validated, and approved to do, and that nothing has compromised its intended behavior. It encompasses both the absence of unauthorized changes (no one has tampered with the model) and the absence of unintended degradation (the model has not drifted from its validated performance).

Model integrity has three dimensions: accuracy integrity (the model produces correct outputs), security integrity (the model has not been tampered with or compromised), and behavioral integrity (the model behaves consistently with its documented design).

Maintaining model integrity requires controls across the full lifecycle: secure development environments, access controls on model artifacts, change management procedures, version control, drift monitoring, and regular validation.

Model integrity failures can have severe consequences. A compromised fraud model may fail to detect crime. A tampered credit model may produce discriminatory outcomes. A degraded monitoring model may miss suspicious activity. Regulators expect institutions to maintain model integrity as a core component of model risk management.

1. Implement access controls. Restrict who can modify model artifacts, configurations, and deployment settings.
2. Use version control. Track all changes to models with documented rationale, approvals, and rollback capability.
3. Monitor for unauthorized changes. Implement integrity checks that detect unauthorized modifications to deployed models.
4. Track performance continuously. Monitor model outputs for unexpected changes that could indicate integrity issues.
5. Conduct regular validation. Periodic revalidation confirms that models continue to perform as intended.
6. Include integrity in incident response. Procedures for responding to model integrity violations should be part of your incident response plan.

Model risk, Validation, AI drift/decay, Data poisoning, Adversarial AI, Version control, Performance monitoring