Human in the loop

Human in the loop describes a workflow deliberately split between AI and a person — some steps handled by the machine, some by human judgment. The phrase is routinely deployed as a five-word safety guarantee, which it isn’t: it means nothing until someone says where the human sits, which calls they make, which outputs they genuinely check, and which actions need their yes before firing.

Why it matters: The most common design — a person “reviewing the output,” watching a mostly-right machine for the moment it goes wrong — is oversight in name only, and it’s been failing on the record for forty years, from autopilots onward. A human can’t stay sharp watching correct output scroll by; by the two-hundredth item the eyes keep moving and the judgment has left the building. The design that works gives the person a real job: decision-maker on a defined slice of cases — the ambiguous ones, the expensive ones — while the machine runs the rest unsupervised. When a pitch says “there’s a human in the loop,” the only question that matters is: doing what, exactly?