# What is Model Drift?

Also called Concept Drift, Data Drift, Performance Degradation.

Model drift is the degradation of an AI system's performance over time as production conditions diverge from the conditions it was built for. It covers data drift, where input distributions shift, concept drift, where the relationship between input and correct answer changes, and version drift, where an underlying model or prompt is updated beneath a working system.

The three forms have different signatures and different fixes. Data drift shows as changing input characteristics, such as new topics, longer documents, or a new language appearing in traffic. Concept drift shows as unchanged inputs producing answers that used to be right and no longer are, typically after a policy, price, or product change. Version drift shows as a sudden behavior change with no corresponding change on the team's side.

Version drift deserves particular attention when building on hosted models. A provider updating a served model can shift tone, verbosity, formatting, refusal thresholds, and tool-calling behavior for a system whose own code did not change. Pinning a version where the provider allows it, and recording which model and prompt version produced every evaluation result, is what makes the cause identifiable afterward.

Detection needs signals that do not depend on knowing the right answer, because ground truth is usually absent in production. Practical proxies include monitoring input distributions against a baseline, tracking output characteristics such as length, refusal rate, and tool-call frequency, watching user behavior signals like retries, edits, escalations, and abandonment, and periodically rerunning a fixed evaluation set as a stable reference.

Response depends on which drift is occurring. Data drift often calls for expanding retrieval coverage or the evaluation set to include the new distribution. Concept drift calls for updating the sources of truth the system depends on, since the world changed rather than the model. Version drift calls for evaluating the new version deliberately and deciding whether to adapt the prompt or pin the previous one.

## Key points

- Three forms: data drift, concept drift, and version drift
- Hosted model updates change behavior with no change on your side
- Record model and prompt version with every evaluation result
- Detect via input distributions, output stats, and user behavior
- The right response depends on which form of drift it is

## In practice

A support assistant's escalation rate rises over two weeks with no deployment on the team's side. Input topics are unchanged, but replies have grown longer and hedge more, and the fixed evaluation set now scores lower on directness. The cause is traced to a provider updating the served model. Because every evaluation run recorded a model version, the comparison took an afternoon rather than a week.

## Related terms

- [Model Evaluation](/en/glossary/model-evaluation)
- [Hallucination](/en/glossary/hallucination)
- [AI Governance](/en/glossary/ai-governance)
- [Audit Trail](/en/glossary/audit-trail)
- [Bias](/en/glossary/bias)

[Back to the AI Glossary](/en/glossary)