Agents
Human-in-the-Loop
A human stays in the workflow and decides, confirms, or intervenes. The formula appears in almost every concept paper and describes two very different things. In one case someone actually checks: they have the information, the time, and the occasion to reject a result, and occasionally do. In the other they confirm what is there anyway, because it has always been right so far and because a hundred approvals a day don't allow a hundred checks. Formally the control is present in both cases; it is effective only in the first. The transition from one to the other is gradual and depends on volume: as soon as checking noticeably costs time and waving things through remains without consequences, it tips. Whoever wants to preserve the control therefore has to manage the volume, not tighten the rule.
An example
A clerk gets a hundred credit notes prepared by the system for approval each day. In the first week he checks each one carefully. By the fourth he clicks through them, because all of them have been right so far. Formally the human is still in the loop. In fact he has stopped being a control and has become a confirmation click, at exactly the moment the number of cases became large enough that checking costs him time.
A closer look
The human sits at defined points in the workflow: before consequential actions, at low model confidence, for exceptions. For this to hold, what is presented must be decidable, with context, justification, and a real possibility of saying no.
Why it matters
Human-in-the-loop appears in concepts more often than it works. For the control to remain real, it needs three things: a selection of what is presented at all, not everything but the unusual; enough information for a judgment in the time available; and feedback showing that checking changes something. If one of these is missing, automation bias results: the human adopts what the system proposes and formally bears responsibility for a decision he no longer makes.
Commonly misunderstood
An approval that arrives every minute gets waved through. Too many checkpoints produce exactly the sham control they were meant to prevent.
Where this term does its work in the magazine.
5 articles use the term or take up its principle.
On Stage
The Demo and the Tuesday
A good demo makes a claim memorable. It doesn’t make it true.Read the articleThe Long View
Who's Actually Leading Whom?
The more autonomous a system appears, the earlier responsibility has to be designed.Read the articleFrom Science
Benchmark Performance ≠ Bedside Performance
A good average can hide precisely the cases in which a system becomes dangerous.Read the articleThe Long View
Deskilling: What We Unlearn When AI Always Has an Answer
Anyone who never practices the fundamentals will later struggle to judge whether an AI result holds up.Read the articleGlass Lab