In a film-trivia experiment where large language models were known to stumble, “I don’t know” collapsed from 44% to 3% when people could ask AI for advice. Even more brutal: correct answers fell from 27% to 9%, while confidence surged from 30% to 76%.

The researchers behind the paper, including Valerio Capraro (associate professor at the University of Milano-Bicocca), Chiara Marcoccia (École Normale Supérieure), and Walter Quattrociocchi (Sapienza University of Rome), set out to test something that feels obvious in hindsight, and dangerous in practice: does AI make humans better thinkers, or just louder parrots? Their headline result was blunt, and it matches what their measurements showed. With AI advice, the normal human pause that says, “I don’t know,” is suppressed, and that suppression shows up in both accuracy and how certain people feel.

Here is what they did. Capraro and colleagues designed questions where large language models typically fail. They used visual details in films, such as the color of the team’s uniform in "Bend It Like Beckham" or the vehicle Monica drives in "Like a Cat on a Highway." The researchers expected these details would be missing from most model training data, which was true for the model used in the experiment (Step 3.5 Flash). They also tested frontier models (GPT-5.5, Claude Sonnet 4.6, Gemini 3.5 Flash). Those sometimes got other details right, and the vehicle question was missed even when the models were newer. That matters because it blocks the easiest excuse: “People just delegated wisely to a reliable tool.” The tool was not reliably right.

The baseline and the “AI advice” condition were built to isolate the effect. One group answered without AI advice. Another group could ask the AI for advice. In the baseline, 44% of participants responded that they didn’t know the answer, suspending judgment. With AI advice available, only 3% did. Capraro summarized it as judgment suspension collapsing. It is a small wording shift with huge consequences. “I don’t know” is a cognitive circuit breaker. If it gets removed, you get confident answers that may be wrong.

And then the study delivered the second hit. Accuracy collapsed when AI help was available. Capraro said that in the baseline, 27% of people gave the correct answer. With AI advice, only 9% did. The study interpretation is straightforward: some participants who might have answered correctly on their own asked the AI anyway, trusted its output, and became wrong. That is not just an error rate issue. It is a trust and calibration issue. People are not simply making mistakes, they are upgrading those mistakes into something that feels authoritative.

The confidence data is where executives should sit up. In the baseline, confidence was at 30%, but with AI help it rose to 76%. Participants believed the bots, despite the possibility of hallucinations. In Capraro’s words, people became much worse, because accuracy was only one third, but they were twice as confident. If you run teams that need quality decisions, this is the part that should keep you awake at night. Overconfidence is expensive. It drives bad escalation, bad approvals, and bad “alignment” with whatever source sounds most fluent.

The researchers also ran an incentive version, which makes the result even harder to dismiss. They conducted the experiment with monetary incentives. Incentives helped, but only partially. Willingness to suspend judgment rose from 3% to 8%, and accuracy rose from 9% to 16%. Still, both metrics were below the baseline levels of 44% for “I don’t know” and 27% for correct answers. So the mechanism is not only “people are careless.” It is that AI advice changes what people treat as sufficient justification, even under pressure to be right.

Why does this matter beyond trivia? The researchers used film trivia, but Capraro said they contend their findings can be generalized across other domains. That is a key point for board members and compliance leaders who are trying to translate research into operational risk. When AI becomes a default assistant, it can reshape how humans calibrate uncertainty in real work: customer support, legal review, medical information triage, engineering decisions, and any domain where “the bot said so” can override the human’s internal “stop, I need to verify.”

Capraro also flagged the policy and education angle, saying the issue needs to be dealt with at a societal level through AI literacy and education policy initiatives. He added that while model providers should try to help, incentives are likely misaligned. His specific worry centered on children: adults have learned critical thinking, but for children “born with these systems,” the risk is they do not even learn the basic critical skills. For decision-makers, the second-order implication is clear: you cannot treat AI adoption as purely a productivity story. You must treat it as a capability-shaping environment that can rewrite how uncertainty is expressed, and therefore how errors propagate through organizations.