The Human Still Signs

The ninety-ninth recommendation of the afternoon says decline the claim. The reviewer has four minutes for it, a queue that does not shrink, and a dashboard that counts her throughput in the corner of the screen. She has disagreed with the system twice this month. Both times, she was the one asked to write up why.

She signs the ninety-ninth. She signs nearly all of them. And every argument about AI in her industry ends with the phrase her signature exists to prove: a human is still in the loop.

She is the loop.

Swap the details and she is everywhere — the contract drafted, the candidates ranked, the scan flagged, the military target recommended, each with a person at the end, approving. The phrase reassures because it quietly joins two different claims. The first is that a human remains formally accountable: there is a name at the bottom. The second is that the human still has the understanding, time and authority to change the outcome. The signature is treated as proof of both. It proves only the first.

A signature can do two jobs

A real sign-off is a control surface. The reviewer can inspect the evidence, reconstruct the reasoning, disagree without penalty, pause the process and make the refusal stick.

A ceremonial sign-off does something else. It gives responsibility somewhere to land.

Both versions look identical on an organisation chart. The difference only becomes visible when the human tries to say no — our reviewer's two write-ups this month are that difference, working on her.

The gap grows mechanically, because machine output and human review scale differently. A system can generate another hundred decisions at close to zero marginal effort; the supply of people who can independently check those decisions does not rise with it. That arithmetic is where her four minutes come from. And it rarely looks like a dramatic handover to the machines. It looks like the same approval workflow, the same professional signature, the same liability policy — now attached to a volume of work no named reviewer could have produced or checked unaided. Delegation arrives without the organisation ever admitting that it delegated.

When advice becomes authority

The line between an advisory system and an autonomous one is usually drawn at the final click. Did the machine act, or did a person approve its recommendation?

That is the wrong place to look. An adviser controls more than the final answer: which evidence is retrieved, which options are compared, how uncertainty is expressed, what never reaches the screen. A person choosing from a machine-framed menu has more agency than a person with no choice, but less than the signature suggests.

Then repetition does its work. When the recommendation has been right ninety-nine times, disagreement on the hundredth requires not only insight but nerve — a delay to explain, a target to miss, and usually no second reader to test the objection against.

Whether that pressure hardens into authority is decided by the surrounding workflow, and the levers are unglamorous. Random audit sampling keeps ordinary cases visible. Rotating reviewers stops one person's deference becoming institutional precedent. A protected second opinion makes disagreement cheap. The opposite arrangements — review measured by speed, overrides treated as defects, the same person watching the same system for years — turn the final click into an expensive way to agree. At some point advice becomes delegation by deference. No policy changes. Nobody removes the human. The human simply stops being the most causally important part of the decision.

The last name in the chain

None of this means human review is doomed. It means review is infrastructure, not ceremony: a reviewer with time to actually inspect, access to the underlying evidence rather than the system's summary of it, and the standing to pause the process without becoming the problem. And because a control that is never used may not be a control, disagreement and override rates are worth measuring — not to punish deviation, but to notice when a veto has quietly died.

The uncomfortable part is liability. If the vendor frames the options, the organisation sets an impossible review rate, and the reviewer rubber-stamps under time pressure, then assigning all responsibility to her signature does not create accountability. It launders the system's design choices into an individual's failure.

The better question is not was a human present? It is where could the outcome still have been changed, by whom, and at what cost?

A scenario, not a diagnosis

This essay, from the AI Scenario Explorer, describes one possible attractor — not a claim that every AI-assisted profession has already reached this state, and no claim about probability. The attractor: formal human responsibility surviving after practical human control has thinned out. Every row below is an observable quantity. A policy saying human reviewed is not.

Signs it is arriving	Signs it is not
Reviewer time per decision shrinking as machine output rises	Reviewer time per decision holding steady as volume grows
The share of decisions materially changed after review collapsing as systems mature	Override and disagreement rates staying alive and measured
Liability settling automatically on the last human name in the chain	Disputed cases placing liability at documented control points

The human signature could remain a genuine veto. But it will not remain one because the box is still on the form. It will remain one only if the person behind the signature keeps the time, knowledge and institutional power to make no mean no — only if, when the hundredth recommendation is wrong, her disagreement costs her an argument and not a career.