The principle of maintaining human oversight or validation within an automated decision, whose real value depends on the quality of the control exercised.
The human-in-the-loop principle consists of retaining a human intervention, in the form of oversight, validation or the possibility of taking back control, within an automated decision process. It is often presented as the ultimate guarantee of safety and responsibility in the face of the growing autonomy of artificial intelligence systems, the idea being that a human watches over and can correct or block an inappropriate decision. Several degrees of it are distinguished, the human in the loop who validates each decision, the human on the loop who supervises and can intervene without systematically validating, and the human out of the loop for fully autonomous systems. The difficulty is that the real value of this principle depends entirely on the quality of the control exercised. When the human, out of convenience, overconfidence or overload, merely rubber-stamps the system's outputs without real scrutiny, which is automation bias, oversight becomes an empty formality that gives the illusion of control without offering its substance. For insurance and law, this question is central, because the formal presence of a human in the loop is often invoked to found liability or justify a deployment, even though it may no longer guarantee anything in practice.
A system requires a human to validate each automated approval decision. In practice, the overwhelmed operator approves hundreds of cases a day without examining them, so that the human in the loop exists on paper but no longer controls anything.
humain dans la boucle, human in the loop, supervision humaine, human oversight