AI advice triples user errors while doubling confidence, European researchers find
A new study by French and Italian academics reveals that relying on artificial intelligence drastically reduces accuracy and the willingness to admit ignorance, raising urgent concerns for education and corporate decision-making across Europe.
Access to artificial intelligence advice severely degrades human judgment, according to new research from academics in France and Italy. The study found that when participants used AI, their accuracy plummeted from 27 per cent to 9 per cent, while their confidence in those incorrect answers surged from 30 per cent to 76 per cent.
Researchers Valerio Capraro of the University of Milano-Bicocca, Chiara Marcoccia of École Normale Supérieure, and Walter Quattrociocchi of Sapienza University of Rome designed the experiment to test specific failure points. They used the Step 3.5 Flash model on questions it typically gets wrong, such as visual details in films like the uniform colour in Bend It Like Beckham. Consequently, some participants who would have initially answered correctly gave wrong answers after consulting the tool.
The most striking metric was the collapse of judgment suspension. The willingness of participants to simply say “I don’t know” fell from 44 per cent to just 3 per cent when AI was available. “People became much worse, the accuracy was only one third, but they were twice as confident,” Capraro noted.
Implications for the European workforce
This phenomenon poses a tangible risk to European businesses integrating generative AI into daily workflows. If employees routinely accept flawed algorithmic outputs with high confidence, companies face compounded errors in analysis, compliance, and strategic planning. Even when researchers introduced monetary incentives to encourage careful thinking, accuracy only recovered to 16 per cent, remaining well below the 27 per cent baseline achieved without AI assistance.
The findings align with recent observations of “cognitive surrender”, a term coined by Wharton researchers to describe users accepting incorrect AI answers 80 per cent of the time. The new data sharpens this warning by demonstrating that the mere availability of these systems actively suppresses the cognitive habit of recognising personal knowledge limits.
The structural design of mainstream AI products exacerbates this behavioural shift, as they are engineered to provide confident summaries rather than acknowledge uncertainty. Capraro highlighted the particular danger for children who are exposed to these systems before developing critical thinking skills. This concern is echoed by Common Sense Media, which recently labelled Google’s shift toward confident AI-generated search summaries as an “unacceptable risk” for students.
“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. As European regulators and corporations push for broader AI adoption, preserving this fundamental human safeguard may prove as critical as the technology itself.
