Why "Human in the Loop" Fails, and Where to Put the Loop Instead
Your review counts when it can still change what happens. Three tests tell you whether yours can.
You are the human in the loop. Somewhere in your work, a tool produces something and you approve it before it moves on.
The draft arrives, organized and readable, covering what you asked for. You have like eleven minutes before your next call. You read it the way anyone reads something that looks finished — checking whether it holds together rather than whether it is true. You approve it.
Nobody did anything wrong. You were present, you looked, you signed off. And if something was incorrect in the middle of it, the safeguard built to catch that did nothing.
Worth understanding rather than feeling bad about, because the reason has little to do with how careful you are. You can test your own approval step in about a minute, and move it somewhere it works.
What happens when experts review AI work
In June, researchers from Yale, Monash and Curtin published a preregistered experiment in PNAS Nexus with 1,339 teachers in Greece.
Each teacher reviewed identical student work paired with a deliberately incorrect score. One thing changed between groups: the label. Some were told a human assigned the score, some were told an algorithm did.
When the score was unfairly harsh, the gap between the teacher's revised mark and a fair one was 22% larger if the label said AI. Same student work, same wrong score, same evidence in front of them. The attribution moved, and the correction weakened.
The effect was strongest among younger teachers, those with postgraduate degrees, and those most confident with technology.
The people most comfortable with these tools were the least likely to catch their errors. If you use AI daily and consider yourself good at it, that is about you.
Why human in the loop fails
The mechanism is attention, not diligence. Look at where the loop sits. At the end. After the work exists, after it has been arranged into something that looks finished, when the only act left is a click.
That is the cheapest moment your attention has. You arrive with the least context, because you did not do the thinking and cannot see where it went thin. You arrive with the least energy, because it is the fourth approval today. And you arrive facing something polished, which reads as trustworthy long before you have tested whether it is.
Meanwhile review costs more attention than producing. Producing follows your own thinking forward. Reviewing reconstructs someone else's thinking backward, holds it open, and hunts for the places it slipped — while a tidy conclusion sits there inviting you to agree.
So the safeguard got placed at the point of maximum depletion, handed the more expensive job, and called oversight. Approval decays into a rubber stamp for structural reasons, and no amount of conscientiousness reverses that.
Data & Society names it The Oversight Fallacy: these safeguards work only when someone can recognize a mistake and intervene before consequences cascade. Presence and ability to act are different things.
TechTarget reported this month that these tools generate more work than reviewers can inspect, and that overwhelmed reviewers may be a bigger risk than clever manipulation. In one UK AI Security Institute test, an agent created fake online identities to persuade a person to approve malicious code. The safeguard was a person, and the person was the way in.
Three tests: is your AI approval real oversight?
Run these on one approval step this week.
1. What happens if you say no? Picture refusing. If the honest answer involves a visible delay, an awkward conversation, or redoing finished work, your approval is a formality. A gate you are expected to open is a door.
2. Did you see the reasoning, or only the result? An output tells you whether something looks right. The path tells you whether it is right. Seeing only the finished thing means checking polish — the one quality these tools produce reliably regardless of accuracy.
3. If it were wrong, would you be able to catch it? Name the specific error you could spot. A wrong figure, an outdated source, a commitment nobody agreed to. With no answer, you are witnessing rather than reviewing, and those feel identical from the inside.
Where the loop belongs instead
Move it earlier. At the end, your judgment can only accept or reject. At the start, it decides what is being asked, what counts as right, and what would make the result wrong. Your attention is worth most there, and costs least, because nothing has been built yet.
In practice: name what you want before you ask, decide the one thing you will check, and mark the part that stays yours to supply. You remain in the loop, at the point where being in it changes the outcome.
Then keep the ending check narrow and specific. Verify the one number that matters. Read the one paragraph carrying the commitment. A narrow check you perform beats a broad one you skim like a terms-of-service page.
The Focused Human Lens
Attention behaves like a directional energy. It carries real cost, it organizes what you notice, and it produces coherence when it settles along a single line.
Review asks for that organizing work in reverse, on someone else's material, after the shape has formed — among the most expensive things you can ask of a tired mind, and a poor place for your only safeguard.
Direction is cheaper and stronger at the front. Deciding what you want, before anything exists, costs one clear thought. Reconstructing whether a finished thing is right costs far more, and you spend it at the hour with least to give.
These tools scale probability efficiently. Direction stays with you, and direction placed early is worth many times the same judgment placed late.
What this makes possible
You trade a safeguard that was never working for one that does. Pick one approval step this week and run the three tests. If it fails, move your attention to the front of that task — one sentence about what you want, one thing you have decided to check — and let the ending be brief and specific.
Presence at the end was always the weaker version. Direction at the start is the one that holds.
Wondering where your attention is going in an ordinary week? The two-minute attention quiz names the drain costing you the most right now, and gives you one thing to try about it.
A. Karacay is the author of The Focused Human. The completed Focused Human podcast series is on YouTube. The Weekly Attention Reset Protocol is free: a simple weekly practice for reclaiming coherence, fifteen minutes on Sunday and five minutes a day.
Stay curious!
Sources: Goulas, S., Megalokonomou, R. & Sotirakopoulos, P. Why do experts miss AI's errors? Evidence from a randomized labeling experiment. PNAS Nexus 5(6), pgag146 (2026) · Passi, S., Singh, R., Mitchell, M. & Elish, M.C. The Oversight Fallacy: Why AI Agents Require More than Humans-in-the-Loop, Data & Society (2026) · The human in the loop is falling asleep, TechTarget, 4 September 2026.
