‘A human in the loop is enough to stay in control.’

Putting a human in the loop has become the standard response to the risks of AI: the tool proposes, a person approves, and control is ensured. The formula is reassuring, and often required.

But a check that is not designed is not a check. A person approving hundreds of proposals a day ends up approving without reading. If they see only the proposal, without the elements that justify it, they cannot judge. And if the tool is right in almost every case, attention slackens precisely on the rare errors. Control has to be designed: who checks what, at what point, with what information and how much time.

• Approval becomes a reflex.Beyond a certain volume, approving is no longer examining.
• The checker often lacks information.They see the tool’s conclusion, rarely what produced it.
• Checking everything amounts to checking nothing.It is better to target uncertain or high-stakes cases than to skim everything.

1. 01Measure how many proposals each person approves per day, and how many they reject.
2. 02Have the elements justifying each proposal displayed, and the cases where the tool is unsure flagged.
3. 03Focus human checking on high-stakes cases, and check the others by sampling.

• How far should AI be allowed to decide on its own?→
• What do we do when AI gets it wrong?→
• How do we keep human expertise when AI does the work?→
• Who is responsible when an AI makes a mistake?→