How was the experiment set up?

More than 1,300 practicing teachers in Greece reviewed identical pieces of student work, each paired with a checklist showing the correct answer and a deliberately wrong score. The wrong score was labeled as coming either from an AI system or from a human colleague, and researchers watched whether teachers corrected it.

What was the main finding?

When the wrong score was too harsh, teachers were markedly less likely to fix it if they believed an AI had produced it. The fairness gap was about 22 percent larger under the AI label than the human label. When the wrong score was too lenient, the source made no reliable difference.

Who was most likely to defer to the AI?

The effect was strongest among younger, more educated, and more tech-confident teachers. That is a notable twist, because those are often the people assumed to be best equipped to supervise automated tools rather than the most likely to be swayed by them.

Teachers Let AI's Harsh Grading Errors Slide, Study Finds

Why Experts Wave Through an AI's Harsh Call

We keep telling ourselves that the safeguard against automated mistakes is a human in the loop. A person checks the machine's work, catches the errors, overrides the bad calls. It is a reassuring picture. A new experiment suggests that the human in the loop may quietly go along with the machine at exactly the moment they are supposed to push back, and that the direction of the machine's error matters more than we would like.

The study, run by Sofoklis Goulas, Rigissa Megalokonomou of Monash University, and Panagiotis Sotirakopoulos, recruited more than 1,300 working teachers in Greece and gave them a task they do every week: judge whether a score on a piece of student work is fair [1]. The twist was hidden in the setup. Each score the teachers reviewed was wrong on purpose, and a checklist made the correct answer obvious. The only thing that changed between groups was a label. Some teachers were told the score came from an AI system; others were told it came from a human colleague.

The asymmetry that jumps out

If people treated AI and human recommendations the same, the labels would wash out. They did not. When the planted score was too harsh, teachers were much less likely to correct it if they thought an AI had assigned it. The researchers put the gap at roughly 22 percent wider under the AI label than under the human one [1][2]. In other words, an unfairly low grade got a pass more often precisely because a machine's name was attached to it.

The mirror-image case is just as revealing. When the wrong score was too lenient, the source of the recommendation made no dependable difference. Teachers were about equally willing to fix an over-generous grade whether a human or an AI had suggested it. So this is not a blanket case of people trusting machines more. It is something narrower and stranger: a reluctance to overturn a machine specifically when the machine is being tough.

Why would that be? The researchers found that a teacher's perception of the AI's competence and its responsibility accounted for more than half of the deference effect in the harsh scenarios [1]. Read that way, the behavior starts to make sense. A strict judgment can feel rigorous, even principled, and if you also assume the system is capable and that accountability sits with the system rather than with you, correcting it feels like second-guessing an authority for the sake of being nice. Leniency, by contrast, carries no such cover. Nobody worries that they were too demanding when they wave through a generous grade.

The people you would least expect

Here is the detail that unsettles the standard oversight story. The deference was most pronounced among younger, more educated, and more technologically confident teachers [1]. These are the very people we tend to nominate as the responsible supervisors of new tools, the ones who understand the technology well enough to keep it honest. Familiarity, it turns out, may make you more inclined to defer, not less.

That has an obvious parallel elsewhere in how we handle machine judgment. People do not respond to an AI's verdict as a flat fact; they respond to what they imagine the system is and what they think it is for. When an accusatory or severe output arrives, some of us dig in and resist it, and some of us stand down. Research on why people reject an AI's confident accusations of lying shows the opposite reflex in a different setting, which is a useful reminder that "trust in AI" is not one dial but many, pulling in different directions depending on the stakes and the framing.

What this does and does not prove

A few limits are worth stating plainly. The work was done in one country with one profession, so Greek teachers grading with an explicit checklist may not stand in for doctors, loan officers, or hiring managers using messier tools under real pressure. The task was deliberately clean, with an obvious right answer, which is almost never how grading feels in a crowded term. And because the study manipulated a label rather than following teachers over months, it captures a snapshot of behavior, not the long arc of how people learn to rely on or resist these systems.

Even with those caveats, the shape of the result is hard to shrug off. The comforting assumption behind "keep a human in the loop" is that the human will catch what the machine gets wrong. This study says the human's willingness to catch it depends on which way the error cuts and on a quiet story they are telling themselves about the machine's authority. Oversight is not automatic. It is a psychological act, and like all psychological acts it has blind spots. For more on how machine judgment reshapes human decisions, see our broader artificial-intelligence coverage.

Sources

Goulas, S., Megalokonomou, R., & Sotirakopoulos, P. (2026). Why do experts miss AI's errors? Evidence from a randomized labeling experiment. PNAS Nexus, 5(6), pgag146. https://academic.oup.com/pnasnexus/article/5/6/pgag146/8703788
EurekAlert / PNAS Nexus. (2026). Human limits when catching AI errors. https://www.eurekalert.org/news-releases/1130956

This article summarizes published research for general informational purposes only and does not constitute professional advice.

Frequently asked questions

How was the experiment set up?: More than 1,300 practicing teachers in Greece reviewed identical pieces of student work, each paired with a checklist showing the correct answer and a deliberately wrong score. The wrong score was labeled as coming either from an AI system or from a human colleague, and researchers watched whether teachers corrected it.
What was the main finding?: When the wrong score was too harsh, teachers were markedly less likely to fix it if they believed an AI had produced it. The fairness gap was about 22 percent larger under the AI label than the human label. When the wrong score was too lenient, the source made no reliable difference.
Who was most likely to defer to the AI?: The effect was strongest among younger, more educated, and more tech-confident teachers. That is a notable twist, because those are often the people assumed to be best equipped to supervise automated tools rather than the most likely to be swayed by them.