Glossary · Document and evidence governance
AI drift
In AI governance, AI drift is a change in an AI system’s outputs or interpretation such that previously valid assumptions, controls or evidence may no longer hold. It can result from changing input data, model updates, new prompts or retrieval sources, or changes in the operating context.
- AI
- Technical documentation
In one sentence
AI drift is a shift in an AI system’s outputs or interpretation that can silently invalidate earlier assumptions, controls and evidence.
Example
A vision system that checks guard positions on a packaging line was validated in summer; under winter lighting its detection rate drops, so the validation evidence no longer describes the system in use.
How it applies
- Evidence: Test results and accuracy figures are valid for a model version, data distribution and context. Record these conditions with the evidence so that drift can be recognized as a reason to re-validate.
- Monitoring: Detect drift with post-deployment monitoring: output statistics, error rates, user overrides and complaints, compared against defined thresholds.
- Technical documentation: Instructions for AI-supported functions should state the conditions under which performance was measured and what users should do when results look implausible.
- AI and retrieval: In documentation assistants, drift also appears when the retrieval index, the prompt or the underlying model changes. Treat each as a change that needs review.
AI drift vs. semantic drift
AI drift concerns the behavior of a model or AI system. Semantic drift is a change in the meaning of terms or data across systems and time. The two often interact: a model trained on one meaning of a field may misread data after the meaning has shifted. For high-risk AI under the EU AI Act, drift is a risk to be addressed in the provider's risk management and monitoring.