With an AI on hand, nobody says “I don't know” anymore

What color is the shirt in Bend It Like Beckham?

What animal sits on the prow of the pirate ship in Asterix and Obelix? What hairstyle does Agatha wear in The Grand Budapest Hotel? These are two of six questions three researchers put to 3,132 people this summer: annoying questions about visual details that almost nobody bothers to write down online.

That choice is the whole point of the design. The questions were picked because the language model handed to participants got them wrong almost every time. Handing someone that tool is like passing along a rigged calculator and watching who bothers to double-check the math.

Participants could always opt out by saying they didn't know. That's exactly what they stopped doing.

This isn't a study about getting the right answer

Chiara Marcoccia (École normale supérieure, Paris), Walter Quattrociocchi (Sapienza University of Rome), and Valerio Capraro (University of Milano-Bicocca) posted their paper on PsyArXiv on July 15. Five experiments, a panel recruited on Prolific in the United States, four preregistrations filed before the data came in. The text hasn't gone through peer review, and both the protocols and the data are public.

Their main measure carries an austere name: suspension of judgment. In practice, it's the share of the six answers where someone writes that they don't know, or leaves the box blank. Two of the authors hand-coded the responses, throwing out cautious guesses like “green, maybe?”; in this tally, dipping a toe counts as diving in.

One detail says a lot about the subject of the study itself: the team had planned to have GPT-5.5 and Claude Opus 4.8 validate that coding, then gave up on the idea. Both models turned out to be too unreliable for this kind of classification work.

Four abstentions out of ten, then one out of thirty

In the first two experiments, with no access to AI, participants skipped 36% and 44% of the questions. Give them an AI to consult, and those shares fall to 6% and 3%.

The starting point moves around depending on the setup: 17% in one experiment, 35% in another. The collapse, though, shows up in all five, and the with-AI condition never climbs above 8% or dips below 1%.

Here's the most unsettling part: confidence and accuracy started moving in opposite directions. Across every no-stakes condition, 27.5% of answers were correct without AI, versus 9.2% with it. Declared confidence, on a 0-to-100 scale, climbed from 29.6 to 75.9. People answered more, got it wrong nearly three times as often, and felt roughly two and a half times more certain of themselves.

Why the researchers rigged the AI to be wrong

This is the methodological call that makes the study hard to dismiss. Normally, following an AI's advice is the rational move: when the tool is right, trusting it looks like division of labor, not laziness. That makes any claim about eroded judgment impossible to tell apart from a simple efficiency gain.

By rigging the questions, the researchers closed off that escape hatch. The model on offer, Step 3.5 Flash, got almost everything wrong, and its name was never revealed to participants. GPT-5.5, Claude 4.6 Sonnet, and Gemini 3.5 Flash got most of the other questions right, all three stumbling only on the hardest one.

One last result drives the point home. With no money at stake, the group with AI access answered more questions than the group without, while producing fewer correct ones. The machine's opinion did more than fill in blanks: it turned into mistakes answers these people would otherwise have gotten right on their own.

And when nobody asked for anything

The fourth experiment is the one that looks most like an ordinary day. The AI's answer appeared automatically, without anyone requesting it, like a dish served before the menu is even opened. It's the situation anyone faces with a search engine that drops a summary above its links, or a word processor that suggests how to finish a sentence.

The effect barely moved. With no stakes, abstention fell from 35% without AI to 1% once the answer was shown. With money on the line, from 39% to 7%. It only takes a plausible-looking answer sitting there for doubt to stop forming at all.

Ten cents changes something, just not what they expected

The researchers were banking on a different outcome. They had preregistered the hypothesis that money would push abstention back up, and that having AI around would dull that effect. Across the three experiments where this was tested, that interaction wasn't significant, and they say so plainly.

What money does do can be measured. With ten cents earned per correct answer and ten cents lost per mistake, participants asked the AI less often: 4.53 times out of six, versus 5.27. Their accuracy rose, specifically in the conditions where AI was available. Their rate of saying they didn't know ticked back up a little, without ever approaching the level of a group with no AI at all.

What the study doesn't say, and the authors admit it first

The first limitation is the setting. Everything rests on visual movie details, chosen precisely because they invite made-up answers. A protocol built to catch the machine lying says nothing about what it produces when it's right, and the authors say so openly: nobody knows whether abstention collapses just as hard when the AI is usually correct.

Nothing here measures medicine, law, or finance. Testing domains where a mistake costs more than ten cents is, in their own words, a priority. They argue the mechanism doesn't depend on the subject matter, noting that the effect holds for blockbuster films as much as for obscure titles. That remains an argument, not a measurement.

The second limitation connects directly to something we wrote here back in March, when verification emerged as AI's hidden cost: the study never recorded whether participants who dismissed the machine's answer went and checked somewhere else. That cost stays out of frame.

The third is the nature of the measurement itself. No longer writing that you don't know isn't the same as believing you do, and what got counted here were statements, not beliefs. Confidence, which doubled, is the only clue to what's happening inside people's heads, and the authors leave the mechanism as an open question: trust in the tool, a lower perceived cost of committing to an answer, or a fluency effect that short-circuits the alarm bell.

The third paper in the same series

Back in April, we noted here that users' trust in AI is poorly calibrated. Together with March's paper on the cost of verification, that body of work sketched out a problem without ever putting a number on it. This study is the first experimental setup to measure the missing piece: the will to check in the first place.

There's one detail left that's as unsettling as it is oddly hopeful. Ten cents were enough to make people ask a machine less often and get a bit of their own judgment back. What AI has taken away looks less like a capacity than a habit: doubt hasn't vanished, it just stopped costing anything.