HITL is not general supervision but a set of designated stops: the system halts before the irreversible steps and asks for a human decision. Money moving, outside communication, deleting data, signing a contract — these are the typical places where the approval is worth the time it costs.
The most important thing to know about it is that the approving human is not a neutral safety layer. In a 2026 randomised clinical trial, 44 doctors performed 14 percentage points worse when the AI gave a wrong recommendation, even after twenty hours of AI literacy training. The training did not protect them.
The most common design mistake follows from that: the more approval points there are, the less each one is worth. Someone who presses the same button fifty times a day is not checking anything, they are nodding mechanically — the literature calls this user fatigue, and the best technical defences end there too.
Where the approval belongs and where it does not is the subject of the human inside the process, a separate article. The directly related terms: least privilege, AI agent and guardrail.