'Human in the loop' is not an operating model.

The claim

"Human in the loop" is a reassurance, not a mechanism. Until someone writes down the trigger, the authority, the response time, and the default, no human boundary exists.

How it fails

The phrase covers four specifications that were never made: what triggers a human review, what that human is allowed to decide, how fast they must respond, and what the system does when they do not. Without them, the "loop" degrades into one of two failure shapes: a reviewer approving everything at seconds per item, or a queue that silently blocks the system. Both are usually discovered in production, by an incident.

Warning signs

• Nobody can name the escalation trigger in one sentence.
• The reviewer's authority is undefined - can they override, or only observe?
• There is no response-time expectation and no default action on timeout.
• Review quality was sampled once, at launch, and never again.

What it costs

In a consequential workflow, the first real incident exposes that the control everyone cited was decorative. The incident review then becomes the production decision - made badly, in public, under pressure.

What would change the answer

A written escalation spec - trigger, authority, response time, default - tested with real reviewers before exposure grows. If your reviewers can explain their trigger and their authority without checking a document, this note does not apply to you.

Where this maps

This failure is a Floor gate failure. The PROOF Scorecard tests this gate in about four minutes; the full review proves it with evidence.