Human-in-the-loop review
Also called: approval gate, review queue, human checkpoint in production
Building a required human approval into the system at the points where a wrong or irreversible action would be costly — the model proposes, a person confirms, before anything consequential happens.
Full automation is appropriate for reversible, low-stakes steps and dangerous for the rest. A checkpoint on the actions that matter — sending, paying, publishing, deleting — is the most dependable safeguard there is, because it doesn't depend on the model being right.
A person watching a dashboard. Real human-in-the-loop review blocks the action until someone decides; passive monitoring notices problems only after they've already happened.
- Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1U.S. National Institute of Standards and Technology · 2023-01-26
First-pass citations, limited to primary sources; a reviewer will broaden and verify these before this entry leaves draft.