Monitoring, maintenance and drift
Watch for deviations in accuracy and model drift - when the relationship between input data and output predictions changes over time. Model cards document features, data, versions and intended use. Two predictable post-deployment risks (a new purpose, new data) are mitigated by documentation and snapshots. Keep a 4-step underperformance response and a human shutdown.
Watch for deviations in accuracy, irregular decisions and drifts in data. New risks appear post-deployment when the AI meets new purposes or new data.
- Model drift, defined → the relationship between input data and output predictions changes over time → the conditions the model was trained under no longer apply and performance declines. Example → a spam detector failing on new types of spam.
- Model cards, in full → standardised records of key features, data used, number of versions, bias or explainability reports, intended use, performance metrics and benchmarked evaluation across cultures, demographics or race → document the original purpose and any new purposes.
- The two predictable risks → a new purpose the AI was not modelled for (mitigate with documentation/model cards) and new data entering the algorithm (mitigate by keeping snapshots of the algorithm and its outputs so you can roll back).
- Best practices → define a baseline, retrain with new data plus human input and feedback, prioritise risk levels and responses, red teaming internally or externally, including pre-deployment, and bug bashing and bug bounties for engagement and feedback.
Keep a procedure to deactivate or localise the system → a human must be able to shut down the algorithm remotely or without direct access. When the AI underperforms → 1) treat it as an incident and use the response plan · 2) identify the issue, who must be told, document the mitigation · 3) notify groups using integrated third-party tools · 4) enable the human shutdown.