Experiment: AI doesn’t make people smarter, but people who “admit they don’t understand” are almost zero

French and Italian researchers conducted five experiments with 3,132 participants. They found that as long as AI recommendations are present, the proportion of people admitting “I don’t know” collapses from 44% to 3%, the correct-answer rate drops from 27% to 9%, but confidence in their own answers soars from 30% to 76%. The questions were intentionally designed so that AI would definitely get them wrong; even when incentives for answering correctly were added, delaying judgment only returned to 8%. Valerio Capraro, who led the research, believes this has to be addressed through AI literacy and education policy.
(Background: AI chatbots don’t just echo users—Nature research reveals the “amplification spiral” that leads to users’ incorrect hallucinations)
(Additional background: Humans develop the AI illness of “brain outsourcing,” and it gets extremely worse! iKala founder warns: seeking convenience destroys originality.)

Key takeaways

• Five experiments, 3,132 participants: when AI is present, the “I don’t know” rate drops from 44% to 3%
• Accuracy drops from 27% to 9%, while confidence in answers rises from 30% to 76%
• After adding incentives for getting answers right, the rate of delaying judgment returns to 8% and accuracy returns to 16%, still far below the baseline

In five experiments with 3,132 participants, the conclusion can be summed up in only one sentence. As reported by technology media The Register, French and Italian scholars found that whenever people have AI they can ask, the share of participants admitting “I don’t know” drops from 44% to 3%, and the correct-answer rate also drops from 27% to 9%, but confidence in their own answers instead jumps from 30% to 76%.

The study was jointly completed by Valerio Capraro, an associate professor at the University of Milano-Bicocca; Chiara Marcoccia of the École Normale Supérieure in Paris; and Walter Quattrociocchi of Sapienza University of Rome. The paper was posted on arXiv in mid-July, and four of the five experiments were pre-registered.

AI does not make people smarter or dumber—it dramatically lowers the threshold for people to “think they understand.”

The questions were intentionally designed so that AI would definitely get them wrong

The research team did not randomly select questions. They chose problems that large language models are especially prone to getting wrong, and all of them centered on visual details in movies—for example, what color the team uniforms in I Love Beckham are. These questions can only be answered by having seen the visuals; the model can only guess based on its language training data.

By fixing the AI recommendation to be wrong, it separates “whether AI is helpful” from “whether the AI is accurate.” Throughout the entire process, participants could choose not to answer at all, simply saying they did not know.

The result was that as long as AI was there, almost nobody selected the “don’t answer” option. Whether the suggestion was prompted by the participants themselves or the system directly displayed the answer on the screen, the suppressing effect was the same.

Even with money incentives, only a small portion is recovered

The team then added even more incentives: pay for correct answers and deduct for incorrect ones. The incentives did work—participants asked the AI less and copied less. The proportion of delayed judgment rose from 3% back to 8%, and the correct-answer rate increased from 9% to 16%.

But it was still far from the baseline. When there was no AI, delaying judgment was 44% and the correct-answer rate was 27%. In other words, even with money placed on the table, participants only recovered less than one-quarter of the gap.

“Being able to say ‘I don’t know’ is extremely important for humans; it means we recognize the boundaries of our own knowledge,” Valerio Capraro said. “But now that we have AI, almost any question can obtain ready-made answers. So we wanted to know whether this will interfere with humans’ ability to say ‘I don’t know’ and to delay judgment.”

Valerio Capraro’s real concern is children, because “adults have already learned critical thinking.” He believes this needs to be addressed through AI literacy and education policy, and that intervention at the education level has a better chance of working.

For people who use AI every day to read on-chain data and break down research reports, this paper is a mirror. What we should worry about is not that the model makes things up, but whether, after reading an AI summary, we ask ourselves: “Did I really understand?”

Frequently asked questions

Does using AI affect people’s critical thinking?

Five experiments by Valerio Capraro’s team (3,132 participants) show that when AI recommendations are present, the proportion of participants admitting they don’t know drops from 44% to 3%, the correct-answer rate drops from 27% to 9%, and confidence in their answers rises from 30% to 76%.

Why did this study design the questions so that AI would get them wrong?

The team deliberately selected movie visual-detail questions that large language models are prone to making mistakes on, so that the AI recommendation is fixed to be wrong. This allows them to separate “whether using AI” from “whether the AI is accurate,” proving that what suppresses delayed judgment is the presence of AI itself—not the quality of the answers.