Zum Hauptinhalt springen
Glossary

Model drift

Updated on 1 min

Model drift means that an AI model performs worse in operation although nobody has changed it. The cause is that reality moves away from the state captured in the training data. With data drift, the inputs change, for example through new products, customer groups or sensors. With concept drift, the relationship itself changes, for example when fraud patterns evolve or a new law shifts the basis of an assessment.

Drift rarely shows itself, because the model keeps producing plausible results. Continuous performance measurement against a reference, thresholds that trigger retraining and documented approvals after every adjustment help. Providers of high-risk AI systems record such observations in their Post-Market Monitoring Plan (Art. 72). The threat T07 Immature technology / model drift puts the topic in context.

No Cookies!

This wiki collects nothing, bakes nothing and leaves nothing behind. There's nothing to consent to. Privacy doesn't get better than this.