In plain terms
An air-traffic controller does not fly the planes; she watches the whole picture and intervenes when something looks wrong. In human-on-the-loop the AI makes and carries out decisions without waiting, and a person supervises from above with the power to override.
Why it matters
It is the practical choice when volume or speed makes approving every item impossible: fraud screening, content moderation, routing thousands of tickets. It only counts as oversight if the person can really notice problems and really stop the system. A dashboard nobody watches is not human-on-the-loop.
Example
A fraud model blocks suspicious card transactions in milliseconds. Analysts watch the block rate and a sample of decisions; when false positives spike after a new release, they roll the model back within minutes.
Most often confused with
HOTL vs. Human-in-the-Loop (HITL)
The difference is what happens when nobody acts. In HITL nothing proceeds without approval. In HOTL everything proceeds unless someone intervenes. HITL catches errors before they happen and costs speed; HOTL keeps speed and catches errors after some have happened.
Under the hood
Requirements for HOTL to be real: telemetry that surfaces anomalies (drift, error rates, unusual actions), alert thresholds, sampling of decisions for review, a tested stop or rollback mechanism, and a named owner with time to watch. Many systems mix the modes by risk: low-impact actions run on the loop, high-impact or irreversible ones are escalated into the loop. Known weaknesses are automation bias and alert fatigue, both of which grow as the system gets more reliable.