For Heads-of · Practitioner

AI model drift and degradation

A deployed model's performance silently degrades over time as the input distribution shifts away from its training and validation data.

  • high
  • operational-resilience
  • model-drift
  • monitoring

How it happens

The real-world data a model sees gradually diverges from what it was trained and validated on, customer behaviour shifts, a new product category appears, a market condition changes, and the model's accuracy degrades without any change to the model itself.

Why it matters

Drift is a slow failure, not a sudden one, so it tends to be caught late, by an external complaint or an audit, rather than by the system that's supposed to be monitoring it.

Mitigating controls

The controls that address this risk, ranked by effectiveness.

Framework and clause references

FrameworkClauseTitle
NIST AI Risk Management Framework (AI RMF 1.0)MeasureMeasure
ISO/IEC 42001:2023Clause 9Performance evaluation

Related resources

The external sources behind this risk, from the Resources library.