The right to a human

The phrase human in the loop has become a kind of talisman, invoked to reassure anyone worried about a system that decides consequential things. There is a human in the loop, we are told, as though the sentence resolved the worry. But the phrase hides the only question that matters, which is not whether a human is present but whether the human has authority — whether, when the machine is about to do something wrong, the person watching can actually make it stop. A human in the loop who cannot override the machine is not a safeguard. They are a witness. And a witness to your own powerlessness is not what anyone means when they ask for a human.

The two things get conflated because they look identical from the outside. In both cases there is a person, a screen, a decision passing before their eyes. But watch what happens when the person disagrees with the machine. In one arrangement, their disagreement changes the outcome: they can hold the action, reverse it, substitute their own judgment, and the system does what they say. In the other, their disagreement changes nothing: they can note it, escalate it, feel it — but the action proceeds, because the human was placed in the loop to observe and perhaps to absorb blame, not to command. The first is a human with authority. The second is a human as upholstery, positioned to make an automated decision feel supervised without being governable.

You can tell the difference by asking a single question of any human-in-the-loop arrangement: what happens when the human says no? If the honest answer is the decision goes through anyway, but now there is a person we can point to, then the human is not in the loop in any meaningful sense. They are attached to the outside of it, close enough to be blamed and too far to intervene. This is the arrangement that lets an institution claim human oversight while having engineered away the one property that would make oversight real — the capacity of the overseer to prevail.

The question is never whether a human is in the loop. It is whether, when that human says no, the machine actually stops.

Why supervision without authority is a trap

Placing a human beside an automated decision without giving them power to change it does something worse than nothing: it manufactures the appearance of accountability while removing its substance. The institution can now say a person reviewed the decision. The person's name is in the record. If the decision was wrong, there is someone who "should have caught it." But the person had no power to catch it — no ability to halt the action, no time to examine it, no standing to override the system's output. They have been handed the liability of a decision-maker and the authority of a bystander, and this combination is not oversight. It is a lightning rod, installed to draw blame away from the system and the people who built it and onto whoever happened to be sitting in the chair.

This trap is easy to build by accident. Give a reviewer a thousand decisions an hour and a single button that only says approve, and you have technically kept a human in the loop while making genuine review physically impossible. The person becomes a rubber stamp not because they are lazy but because the arrangement permits nothing else — no time to read, no real option to reject, no consequence-free path to disagreement. Their presence is now pure theater, and the theater is load-bearing: it is what lets everyone downstream say the decision was humanly supervised. The reviewer's helplessness is not a flaw in the system. On a certain reading, it is the system working exactly as designed.

Real authority has requirements that upholstery does not. The human must have the time to actually consider the decision, not merely the opportunity to wave it through. They must have a genuine option to decide otherwise, one that carries no punishment for its exercise. They must have the information to decide well — which means the system owes them not just its conclusion but the evidence and reasoning behind it, or their override is uninformed and therefore hollow. And crucially, their no must be dispositive: it must stop the action, not merely register a complaint that the action ignores. Strip away any one of these and you are back to a witness, however senior their title.

What a right to a human would actually require

The remedy is to stop measuring oversight by the presence of a person and start measuring it by the reachability of an override. A meaningful right to a human is a right to a human who can actually change the outcome — someone with the authority to countermand the machine, the time to exercise it, and the information to exercise it well. That last requirement is where the record does its work: a human cannot responsibly override a decision they cannot understand, so the system must hand the overseer a Decision Receipt — the evidence it relied on and the reasoning it followed — turning the human's authority from a rubber stamp into an informed judgment. Authority without understanding is just a different way to abdicate.

And the override itself must be accountable in the same way the automated decision is. When the human says no, that no should be captured with its reasons, so that the exercise of authority is itself on the record — contestable, reviewable, real. This is what closes the loop honestly: the machine can be held to account because its reasoning was kept, and the human can be held to account because their override was kept, and neither can hide behind the other. The point of a human in the loop is not to have a name to blame. It is to place, at the point of decision, a person who can be answered to and who can actually answer back.

So the next time an institution assures you there is a human in the loop, ask the only question that separates a safeguard from a decoration: when that human disagrees with the machine, whose judgment wins? If the machine wins, the human is there to absorb your objection, not to honor it — and what you have been given is not a right to a human at all, but the right to be told, by a person who was powerless to help you, that the decision has already been made.

— Dispatches · Summit Cognitive