An AI system routes decisions to human review. The override rate is 0.3%. Is the oversight working?

What the interviewer is testing

Whether you recognise automation bias and can design review that actually reviews.

Why 0.3% is a warning

An override rate near zero has two possible explanations: the model is right 99.7% of the time, or the reviewers are approving without genuinely assessing.

The second is far more likely, and the mechanism is well documented. A reviewer presented with a stream of mostly-correct decisions, under throughput pressure, with an interface that makes approval one click, becomes a rubber stamp. The oversight becomes a liability shield rather than a control — which is worse than no review, because it creates the appearance of a safeguard.

The way to distinguish the two is to measure independently: have a separate expert review a sample of the approved decisions and compare. If they find errors the reviewers approved, the review is not functioning.

Designing review that works

Route only cases that need judgement — low confidence, high value, unusual patterns, plus a random sample for quality measurement. A reviewer seeing only hard cases stays engaged; one seeing mostly easy cases stops looking.

Show the evidence, not the verdict. The retrieved sources, the specific factors, the reasoning — so the reviewer is forming a judgement rather than confirming one.

Make disagreement as easy as agreement. An interface with a prominent Approve and a buried override produces the outcome you would expect.

Manage throughput. A reviewer with 90 seconds per case cannot assess anything.

Track the override rate as a control metric, with both extremes as alerts: near zero means the review is not functioning; very high means the model should not be deployed here.

What a strong answer adds

The placement distinction: review before an irreversible action is oversight worth its latency; review after the fact is monitoring — legitimate, but it should not be described as oversight in a regulatory submission.

Common weak answers

Interpreting the low rate as model quality. Adding reviewers without changing the interface or the routing.