Studying human-in-the-loop AI across decision settings, we find that oversight changes outcomes only when the human has the context, the time, and the authority to dissent. Where any of the three is missing, human approval rates exceeded 98% and oversight added latency without changing a single outcome — oversight in name only.
- Oversight changed outcomes only when the human had context, time, and authority to dissent.
- Missing any one of the three drove human approval rates above 98% — rubber-stamping.
- Effective oversight surfaces the reasoning and protects the time to act on it.
When oversight is theatre
A human in the loop with no time to think, no context to judge, and no real authority to say no is not oversight — it is a signature. Our data shows these arrangements approve almost everything, add latency, and create a false record of human control. The presence of a human is not the same as the presence of oversight.
Method
We observed human-in-the-loop arrangements across decision settings, measuring approval rates, time-to-decide, and whether the human had access to the model’s reasoning and genuine authority to override. We correlated these against whether oversight changed any outcomes.
- Context — the reasoning and evidence behind the recommendation, in the moment.
- Time — enough of it to actually evaluate, not just acknowledge.
- Authority — a real, supported path to dissent and override.