Human in the loop is not a job description

The approval queue is a lonely place, and nobody designed it to be.

The system has done the first pass. The recommendation is already highlighted. A clock is running somewhere, even if it isn’t on your screen.

Your name is beside the button that says approve.

You’re expected to be the final check. You’re also expected to answer messages, clear the next case, remember yesterday’s exception, and leave on time.

By the time the case reaches you, the work of deciding has been compressed into the work of agreeing.

You’re not a weak reviewer because you can’t catch every error at machine speed.

Nobody can review an unlimited stream meaningfully.

You’re tired. You’ve got deadlines, children to collect, bills, and a job whose interruptions don’t become less real because a dashboard calls them workflow.

When the system makes an ordinary case look settled, accepting that is a human response to a day already full.

Nothing here wants your job.

The problem is quieter than that.

A person who glances at a model output isn’t necessarily a control. A person who clicks the last button isn’t necessarily the person who decided.

1994 — You type the numbers into the ledger. The total is yours because your hands made it.

2007 — The system validates the fields. You still see the number it refuses.

2023 — The model recommends. The explanation sounds like the conclusion.

2026 — You click approve, knowing the queue has already moved on.

Your title hides the unit

For decades people described computerisation as if it arrived for occupations.

Accountants.

Clerks. Factory workers. Designers.

The employment record says something more awkward and much more useful to you. Technology moves through tasks.

Your job is a bundle. Inputs, repeatable transformations, decisions, exceptions, hand-offs, quality checks, communication, accountability.

A system can take one part of that bundle and leave you holding the part that hurts to name.

Across the United States between 1970 and 1998, labour input moved toward nonroutine cognitive work and away from routine cognitive and manual work. Computerisation explains roughly 60% of the demand shift that favoured college-educated labour over that period.

The word doing the work there is relative. It doesn’t mean a credential keeps you safe.

It means the work around the machine didn’t distribute itself evenly.

Ask the typesetters. The travel agents. The translators.

Not because their work vanished in one clean historical scene. Because a title can survive while its routine centre is hollowed out, and what’s left becomes more exposed, more exceptional, more contingent.

Which is why “this will augment the team” isn’t yet a description. It’s a claim that needs a unit.

Which task gets faster? Which task is newly required? Who checks the result?

And what happens to your released time — better service, a bigger queue, fewer hires, fewer people?

Capacity and headcount are not the same event. Neither is a glowing demo and a complementary job.

Not a spare part

The modern version of that division has a flattering name. Human in the loop.

It sounds like a job description. It can mean a radiologist who sees the image, the clinical history, the model mark and the original study before signing a report.

It can also mean you, receiving a polished draft with a green confidence badge while the next ticket lands.

Those aren’t two versions of the same safeguard.

In one study of chest-trauma patients, AI-assisted radiologists reached 94.2% sensitivity for rib fractures against 69.2% for standard double reading.

The point isn’t that medical AI works. The point is that the human role is visible. Inspect the image. Consider the history. Turn the aid off. Record disagreement. Get another reader where the protocol demands one.

The machine supplies a mark. The person still has something to look at, and a way to disagree.

Now compare a review queue built around throughput.

An eight-minute case doesn’t become a one-minute case because the first draft took one minute. If you spend four minutes checking sources, correcting a policy exception and recording what happened, the model moved the work. It didn’t remove it.

Sent unchanged. Materially edited. Rejected. Escalated. Those are four different outcomes, not four moods.

That’s the correction. You’re not the human in the loop because your name is on the screen.

You’re one only when you can change what happens next.

Advice is not evidence

The trouble isn’t only that a model can be wrong.

It’s that advice changes a correct human judgment, in the exact person who was retained to prevent that.

Automation bias has two names for how it gets you. A commission error is following bad automated advice. An omission error is failing to act because the aid didn’t flag anything.

Across prospective clinical studies, consultations where the advice made a decision worse ran somewhere between 6% and 11%.

You don’t have to be gullible for that to happen to you.

Monitoring gets less likely under high workload and competing tasks. Give a reviewer several queues and they get less effective as output rises.

The volume promise quietly defeats the human promise. More outputs. Less attention. A signature at the end.

Confidence makes the trap look scientific.

In one study of people solving logic puzzles, well-calibrated confidence displays raised accuracy by twenty percentage points. Miscalibrated ones produced almost nothing, and increased both overreliance and underreliance.

A percentage beside an answer isn’t evidence because it has a percent sign. It’s evidence only if it’s calibrated for this task, this population, this setting. And only if it lets you act.

Order matters as much as content.

An independent first judgment gives you something to compare the recommendation against. A recommendation-first screen saves you time and anchors the judgment it claims to assist.

Across preregistered experiments, people leaned harder on algorithmic advice as the task got objectively harder. The cases where you most want help are the cases where independent verification is hardest.

Sending those to one rushed person isn’t a solution. It’s a transfer.

Make the complement real

There’s no general law that a human and an AI together beat either one alone.

A preregistered meta-analysis of 106 experimental studies found human-AI combinations performed worse, on average, than the better of human-alone or AI-alone.

They did better at content creation and worse at decisions.

Which makes sense once you say it plainly. A draft is an object you can push back against. A recommendation arrives already pretending to be the reason.

So configuration is the whole story.

In a mammography model, using AI as a second reader or a high-confidence filter improved sensitivity and cut human reads sharply. The same model also degraded performance through automation bias in multi-reader settings.

One model. Four positions — first, second, beside you, or between you and the cases you’ll never see. Four different interventions.

And the remaining work has to be counted honestly.

Concurrent assistance reduces your reading time on every case. Pre-screening reduces the number of cases and leaves you a smaller, harder residue.

Those are not interchangeable savings. Route 95% of the work automatically and your specialist gets the 5% most adverse, ambiguous and emotionally difficult.

Fewer cases is not necessarily less work.

Here’s where oversight becomes either a control or an alibi. You need time proportionate to the error class. Source excerpts, not a bare conclusion. A short dissent path. Authority to stop or alter the result. A fallback when the system is uncertain or absent.

Contradictory policy text should route to a specialist. It shouldn’t become an invitation for the fluent answer to choose.

And when the assistant fails, the approved knowledge base and the manual template still have to work. A service outage can’t be allowed to become a customer-service outage.

NIST’s AI RMF names four cross-cutting functions — Govern, Map, Measure, Manage — and treats governance as continuous rather than a signature at the start.

Its playbook wants human oversight documented through what actually happens downstream. Overrides. Reported errors. Response time. Adjudication.

That’s not paperwork bolted onto the real system. It’s the evidence that your part of the system exists at all.

If nobody can see the source, dissent, wait, intervene or recover, then your employer hasn’t kept a person in the loop.

It has kept a name at the end of one.

You may still be useful to the system. You may be indispensable to it.

But usefulness isn’t authority. Presence isn’t evidence. A click is not a decision.

The old computerisation record isn’t a prophecy that your job disappears. It’s a warning against taking your title as the unit of analysis.

The human-AI record adds the sharper one. Complementary work becomes a safeguard only when the complement can really see, really decide, really refuse, and really repair what the system has done.

Everything else is a job description for whoever carries the blame after the queue has moved on.