Reliability and control · Layer 6
AI evaluation
Measuring, on cases prepared in advance, whether an AI gets things right, before go-live and after every change.
A way to picture itAn exam with a marking scheme: you ask every new version the same questions and compare the papers. An automatic marker speeds things up, but you need to check its marks from time to time.
A legal department wants an assistant that answers operational staff’s questions about standard contracts. It writes 150 questions with reference answers, including 20 where the right answer is “it depends, ask a lawyer”. A judge model scores every answer, and each week a lawyer reviews a sample of the scores to check that the judge is still reliable. When the model version changes, the score drops on termination clauses: the update is put on hold.
For high-risk systems, the provider must achieve an appropriate level of accuracy and declare it, with its metrics, in the instructions for use given to the deployer (EU · AI Act, art. 13 and 15). The GDPR requires accurate data (EU · GDPR, art. 5): an AI that produces information about people must be tested on that point. Keep test sets, results and go-live decisions: they are proof of control, useful in an audit as well as after an incident.
“On which test set, approved by whom, do you measure your solution’s quality, and can we rerun it ourselves with every new version?”
A second language model scores the first one’s answers against a grid. It is fast and cheap, but the judge can be wrong or biased: compare its scores with an expert’s on a regular basis.
Not on its own. You need to know which cases it was measured on, what the 5% of errors cost and who catches them: a rare error on a high-stakes decision matters more than a good average.
After every change of model, instructions or document base, and regularly in production, because real-world questions change over time.
Educational explainer, not legal advice. Up to date as of 29 October 2026.