Human-in-the-loop is a system design in which humans review, approve, or correct AI outputs at defined checkpoints: approving an agent's consequential actions,
A system design in which humans review, approve, or correct AI outputs at defined checkpoints, approving an agent's consequential actions, reviewing low-confidence classifications, or grading samples that feed back into evaluation.
Autonomy is granted per action class based on reversibility and stakes, with human gates on everything irreversible. Done well, the human effort compounds: corrections become training and evaluation data that improves the system.