noun · AI practice
Human in the loop
also called approval step
A person keeps a hand on the brake. The AI does the work, but you say yes before anything important happens.
A design where a person reviews or approves an AI system's actions at chosen points before they take effect.
The harness pauses before certain actions and waits for a person to approve, edit or reject them. The model keeps no say in this; the rule is code that sits between its request and the actual action.
Real-life example
You let a coding assistant run on your project. It reads files freely, but when it wants to delete a folder or push code, a prompt appears asking you to approve. After a hundred harmless approvals you start pressing yes without reading. The safeguard is still there, but it has stopped doing its job.
Where the analogy breaks
-
It is not automatically safe. People get tired of approvals and start clicking through, so the checks that matter must be few and clear.
-
It slows things down on purpose. Put the pause where a mistake is costly or hard to undo, and let low-risk steps run on their own.
-
A person can only judge what they can see. If the approval screen hides what the action really does, the check means little.