What 'Human in the Loop' Really Requires

"Human in the loop" is everyone's favorite fairness safeguard. A person reviews the AI's decisions, catches the mistakes, keeps things fair. It sounds airtight — until you look at how it usually works.

In practice, the human often sees the AI's recommendation first, under time pressure, with a hundred more cases in the queue. Study after study shows what happens: people defer to the machine. The human becomes a rubber stamp, and the "oversight" is theater.

In plain English: a human who can't realistically disagree isn't oversight. They're decoration.

Real human oversight requires design, not just assignment. The reviewer needs the time to actually review, the information to judge independently (not just the AI's score), the training to spot the system's known weak spots, and — hardest of all — a culture where overturning the AI is normal, not insubordinate.

So if your fairness plan says "a human reviews every decision," ask the follow-up questions: which human, with what information, how much time, and when did they last overrule the system? If the answer to the last one is "never," you don't have a human in the loop. You have a human in the photo.

I explain frameworks; I don't give legal counsel.