In 2026, AI still hallucinates. It still gives wrong answers a significant portion of the time. Yet a new study from researchers at three European universities has found something more troubling than the models’ errors: the humans using them are losing the ability to know when they don’t know.

The study, conducted by Valerio Capraro of the University of Milano-Bicocca, Chiara Marcoccia of École Normale Supérieure, and Walter Quattrociocchi of Sapienza University of Rome, produced a striking finding. When participants had access to AI advice on visual detail questions drawn from films, their accuracy dropped from 27% to 9%. Their willingness to say “I don’t know” collapsed from 44% to 3%. And their confidence in their own answers rose from 30% to 76%.

“People became much worse — the accuracy was only one third — but they were twice as confident,” Capraro said.

The design behind the finding

The researchers deliberately engineered the experiment to isolate a specific mechanism. They selected questions about visual details from films — the color of a team’s uniform in Bend It Like Beckham, the vehicle a character drives in Like a Cat on a Highway — details they expected would be absent from most model training data.

They used Step 3.5 Flash, a model they knew was usually wrong on these questions. This was intentional. Any drop in human judgment could not be explained as “sensible delegation to a reliable tool.” Some participants who would have answered correctly on their own asked the AI and became wrong.

The study involved five experiments with over 3,000 participants, most of them preregistered. The researchers also tested recent frontier models — GPT-5.5, Claude Sonnet 4.6, Gemini 3.5 Flash — which missed the vehicle question but often got other details correct. The pattern held.

Monetary incentives barely helped

Even offering money for accuracy only nudged the numbers. Willingness to admit ignorance rose from 3% to 8%, and accuracy from 9% to 16% — both still well below the no-AI baselines of 44% and 27%.

The researchers’ interpretation is metacognitive. “The mere availability of AI suppresses the cognitive habit of recognising what you do not know,” Capraro said. This is not about trusting wrong answers. It is about the erosion of the cognitive habit of uncertainty itself.

“For humans, the capacity to say ‘I don’t know’ is very important because it represents the recognition of the limits of our own knowledge,” Capraro said. “But now with AI, we can get an easy answer to virtually every question.”

A pattern, not an isolated study

Wharton researchers Gideon Nave and Steven Shaw reached a related conclusion earlier this year, coining the term “cognitive surrender” to describe the same phenomenon. Across three experiments and 1,372 participants, they found that simply having an AI available was enough for people to surrender their own judgment. Users accepted incorrect AI answers 80% of the time while reporting higher confidence than those working without AI.

Shaw described the dynamic plainly: “We saw that even when cognitive surrender is engaged, people adopt those answers and are more confident in those answers.”

The Capraro study adds sharper data to that conceptual framework. As AI Weekly noted, “the shape of the finding matches what Wharton saw independently.”

The design implication

Capraro expressed particular concern about children, who are growing up with these systems before they have developed critical thinking skills. Common Sense Media this week called Google’s AI search overhaul — which replaced links with confident AI-generated summaries — an “unacceptable risk” for students. The organization recommended that students stop using it entirely until schools can disable the AI features.

The pattern across these studies is consistent: AI products are designed to answer, never to say “I don’t know.” The humans using them are learning to do the same.

For anyone designing AI-assisted decision-making systems, the implication is not to abandon AI. It is to design for metacognitive preservation — to build interfaces that actively preserve the user’s capacity to recognize and express uncertainty, even when the model is confidently wrong.

The challenge is no longer how to make the model more helpful. It is how to stop it from making humans less capable.