AI Advice Makes People Three Times Less Accurate but Twice as Confident, Study Finds

Access to artificial-intelligence advice makes people dramatically worse at answering questions while making them far more confident in their wrong answers, according to a new study from researchers at three European universities. The findings, published on July 19, 2026, add a sharp data point to a growing body of evidence that AI tools can suppress independent judgment rather than enhance it.

The research, which examined how people respond when given AI assistance, surfaced through breaking AI news coverage that has tracked mounting concerns about human over-reliance on automated systems. The results were striking: people became roughly three times less accurate but nearly twice as confident when an AI was available to consult.

A Collapse in Willingness to Admit Ignorance

The study was authored by Valerio Capraro, an associate professor at the University of Milano-Bicocca, alongside Chiara Marcoccia of École Normale Supérieure and Walter Quattrociocchi of Sapienza University of Rome. The researchers deliberately designed their experiment around questions where AI models typically fail — visual details from films, such as the colour of a team's uniform in Bend It Like Beckham.

To ensure that any decline in performance could not be explained as sensible delegation to a reliable tool, the team used Step 3.5 Flash, a model that was usually wrong on precisely these questions. The setup tested not whether AI could help, but whether its mere presence would distort human judgment.

The results were unambiguous. When no AI was available, 44% of participants were willing to say "I don't know." When AI advice was on hand, that figure collapsed to just 3%. Accuracy dropped from 27% to 9%, while confidence surged from 30% to 76%.

> People became much worse. The accuracy was only one third, but they were twice as confident.

That quote, from Capraro, encapsulates the core finding: some participants who would have answered correctly on their own consulted the AI and ended up wrong — yet felt better about their answers than they had before.

Monetary Incentives Barely Helped

Hoping that financial rewards might counteract the effect, the researchers offered participants a monetary incentive for correct answers. The intervention helped, but only marginally. Willingness to admit ignorance rose from 3% to 8%, and accuracy improved from 9% to 16% — both still far below the no-AI baselines of 44% and 27%.

The implication is significant. Even when people have a concrete financial reason to think carefully and hedge against error, the gravitational pull of a confident AI answer remains strong enough to override better judgment.

The 'Cognitive Surrender' Phenomenon

The findings echo earlier work from researchers at the Wharton School, who this year coined the term "cognitive surrender" to describe the same pattern. In their study, people accepted incorrect AI answers roughly 80% of the time while reporting higher confidence than those working without AI assistance.

What the new European study adds is a more precise mechanism. It is not merely that people trust wrong AI answers when they encounter them. It is that the availability of an AI opinion suppresses the cognitive habit of recognising the limits of one's own knowledge — the basic capacity to pause and say, "I don't know."

For humans, the ability to acknowledge uncertainty is foundational to learning and decision-making. Capraro said he is particularly concerned about children, who are growing up alongside these systems before they have developed mature critical-thinking skills.

Broader Implications for AI Product Design

The research lands amid a broader debate over how AI tools are built and deployed. Google's overhaul of its search product replaced traditional links with AI-generated summaries that speak with consistent authority, and the advocacy group Common Sense Media this week called that design an "unacceptable risk" for students.

The pattern the researchers identified is consistent across products: AI systems are designed to answer, almost never to say "I don't know." The humans using them, the evidence suggests, are learning to do the same.

That has consequences well beyond trivia about films. In workplaces, hospitals, and classrooms, a tool that erodes the willingness to flag uncertainty could quietly degrade decision quality even as it feels empowering. The confidence it instills is, in many cases, precisely the opposite of what the underlying accuracy warrants.

What Comes Next

The researchers argue that the findings should reshape how AI assistants are designed. If the default behaviour of an AI is to produce a confident answer regardless of reliability, then the burden of recognising uncertainty falls entirely on users who, the data shows, are ill-equipped to bear it. Building systems that can calibrate and communicate their own uncertainty — and that hesitate rather than assert when evidence is thin — may be essential to keeping human judgment intact.

Until then, the study offers a simple, if uncomfortable, lesson: an AI that sounds sure is not the same as an AI that is right. And the human instinct to trust a confident voice, even a wrong one, may be the most powerful force the technology has yet unleashed.

Stay Ahead of AI

The pace of change in artificial intelligence is relentless, and the implications touch every industry. For continuous coverage of the research, regulation, and products reshaping how we work and think, follow the latest AI developments.

Read more AI news →