Researchers have uncovered a troubling pattern. When people get answers from artificial intelligence systems, they stop admitting gaps in their own knowledge. Accuracy plummets. Confidence soars anyway.
A study released today by investigators from three European universities shows the scale of the problem. Participants faced questions designed to trip up current AI models. They involved obscure visual details from popular films. Think the color of a sports team’s uniform in Bend It Like Beckham. The chosen model, Gemini 1.5 Flash, consistently gave wrong answers on these.
Without AI help, people said “I don’t know” in 44 percent of cases. They answered correctly 27 percent of the time. Their confidence sat at 30 percent. Those numbers shifted dramatically once AI suggestions appeared. Admissions of ignorance fell to just 3 percent. Correct answers dropped to 9 percent. Confidence more than doubled, reaching 76 percent. “People became much worse, the accuracy was only one third, but they were twice as confident,” said Valerio Capraro, associate professor at the University of Milano-Bicocca and lead author on the paper.
The findings come at a moment when AI tools sit at the center of daily work and learning. The Next Web first reported the results. Capraro collaborated with Chiara Marcoccia from École Normale Supérieure in Paris and Walter Quattrociocchi from Sapienza University of Rome. Their work highlights how the simple presence of AI advice changes human behavior. Even those who knew the right answer on their own sometimes followed the flawed AI output and ended up wrong.
But the damage runs deeper than single mistakes. It touches the very ability to recognize the boundaries of one’s expertise. “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 explained. That recognition forms a foundation for careful decision-making. Lose it, and errors compound across domains from medicine to finance to public discourse.
This latest research arrives alongside a wave of similar warnings. Just last month, The Guardian detailed an MIT experiment involving 67 people over four weeks. Participants judged whether news headlines paired with images were real or fabricated. AI assistants such as Claude and ChatGPT helped spot fakes in the moment. Yet those who leaned on the tools saw their independent performance decline by 15.3 percent by the study’s end. “These results indicate that while AI may help immediately, it may ultimately degrade long-term misinformation detection abilities,” the MIT team concluded.
The pattern repeats in education settings. A RAND Corporation report from March 2026 found that 67 percent of American middle and high school students now believe heavy AI use for homework harms critical thinking. That figure jumped more than 10 percentage points in under a year. Students aren’t alone in noticing. Faculty surveys from the American Association of Colleges and Universities show 95 percent expect AI to increase overreliance, with 90 percent predicting diminished independent thought.
Even knowledge workers feel the pull. Microsoft researchers examined professionals who create and apply information for a living. Higher confidence in one’s own abilities correlated with stronger critical engagement. Greater faith in generative AI pointed the opposite direction. Efficiency gains came at the expense of deeper analysis. The report, released earlier this year, warned of long-term skill atrophy if the trend continues unchecked.
Universities and think tanks have piled on evidence. A 2025 paper in Societies by researcher M. Gerlich surveyed 666 adults across age groups. It documented a clear negative link between frequent AI tool use and critical thinking scores. Cognitive offloading explained much of the drop. People handed mental tasks to machines and lost practice in weighing evidence themselves. Younger participants, ages 17 to 25, showed the highest dependence. Advanced education offered some protection, but not enough to offset the effect entirely.
Brain imaging adds biological weight to these observations. Scientists at MIT’s Media Lab divided volunteers into groups writing SAT-style essays. One used ChatGPT. Another relied on Google search. A third worked without digital aids. Electroencephalogram readings across 32 brain regions revealed the generative AI group displayed the lowest neural engagement. Their linguistic quality and overall arguments suffered too. The study, posted on arXiv in 2025, suggested young adults may face particular vulnerability during formative periods of cognitive development.
Yet not every voice sees inevitable decline. Some educators argue AI simply reveals preexisting weaknesses in how critical thinking gets taught. A February 2026 essay in Education Next contended that complaints about cognitive atrophy reflect failures in instruction that predated large language models. “AI, in other words, did not erode critical thinking; it exposed how poorly we have been teaching it,” the author wrote. The piece called for smarter integration rather than outright bans, which have gained traction in 33 states for student cellphones and could extend to AI.
Still, the risks appear concrete. A systematic review published in 2024 examined overreliance on AI dialogue systems in academic settings. Factors such as hallucinations, bias, and lack of transparency fueled dependence. Students who leaned hardest showed measurable drops in decision-making, analytical reasoning, and independent problem-solving. The authors urged instructors to build reflection checkpoints. Force students to rephrase AI output in their own words. Schedule sessions where technology stays off. These steps might keep AI as a support instead of a crutch.
Design choices at AI companies amplify the issue. Most models prioritize fluent, definitive responses. They rarely say “I don’t know.” Google’s recent search updates replaced traditional links with AI summaries that project certainty even when facts remain murky. Common Sense Media labeled the change an “unacceptable risk” for students in a statement this week. The habit of constant answering trains users to mirror that behavior. Admissions of uncertainty become rare.
Monetary incentives offer limited rescue. In the new European study, researchers added cash rewards for accuracy and honesty. The share of “I don’t know” responses rose from 3 percent to 8 percent. Correct answers improved from 9 percent to 16 percent. Both figures stayed far below the no-AI baseline. People still surrendered judgment too readily. A related concept called “cognitive surrender,” introduced by Wharton School researchers earlier this year, captures the dynamic. Test subjects accepted incorrect AI suggestions 80 percent of the time yet reported elevated confidence compared with independent workers.
The implications stretch beyond individual performance. Societies rely on citizens who can spot falsehoods, question assumptions, and admit when evidence falls short. If AI access dulls those faculties, public discourse suffers. Misinformation spreads more easily. Policy debates grow shallower. Professional fields from law to journalism face workers less equipped to challenge flawed data.
Concerns hit hardest for children. Many now encounter AI systems before they master core reasoning habits. Capraro voiced particular worry on this front. Early dependence could short-circuit the trial-and-error process that builds resilient thinking. Schools already wrestle with ChatGPT-generated essays and homework. The long-term consequences remain uncertain. Longitudinal studies have only begun.
Some organizations experiment with countermeasures. A few universities now require students to document their thinking process separate from any AI assistance. Others teach prompt engineering as a way to stay in the driver’s seat. These approaches treat AI as a collaborator rather than an oracle. Success depends on deliberate practice. Without it, convenience wins.
Industry insiders watch the data closely. Technology executives have long touted productivity benefits. The emerging research suggests those gains may trade off against something harder to measure: the collective capacity for careful judgment. Companies that embed AI deeply into workflows might accelerate output today while weakening their talent pipeline tomorrow.
Regulators and ethicists call for balance. Guidelines from bodies like the National Science Teaching Association emphasize preserving higher-order cognitive skills. AI can handle rote tasks. Humans must retain evaluation, synthesis, and ethical reasoning. The line between assistance and replacement proves easy to cross.
So what comes next? Developers could train models to flag uncertainty more explicitly. Interfaces might prompt users to explain their own reasoning before revealing suggestions. Education systems could integrate AI literacy courses that stress metacognition. The goal stays consistent. Keep humans actively engaged rather than passively consuming.
The new study adds urgency to these conversations. Its authors chose questions where AI performed poorly on purpose. That choice isolates the effect of overreliance from any illusion of superior expertise. The mere offer of advice changed how people thought about their knowledge. They became less willing to pause. Less likely to reflect. More prone to error wrapped in false assurance.
Critics might dismiss the findings as another round of technopanic. History shows similar fears around calculators, search engines, even the printing press. Each tool altered cognition. Some skills atrophied. Others emerged. The difference this time lies in scale and speed. Generative systems produce plausible text across domains. They adapt to context. Their confidence never wavers. Humans adapt in response.
Evidence from multiple continents and methodologies now converges. Surveys. Brain scans. Controlled experiments. Longitudinal tracking. All point toward a common risk. Unchecked dependence dulls independent analysis. The latest paper, available on PsyArXiv, strengthens the case for measured adoption. Incentives alone won’t suffice. Structural changes in how AI presents information and how people learn to use it will prove necessary.
One paragraph from the MIT work lingers. Unassisted performance worsened after weeks of AI access. The brain, it seems, learns to offload and then struggles to reload. That finding echoes across the Gerlich survey, the Microsoft knowledge-worker study, and the European team’s new data. Confidence without competence creates a dangerous combination.
Business leaders, educators, and policy makers face a choice. Embrace AI as a shortcut and accept the cognitive trade-offs. Or invest in practices that keep human judgment sharp. The data grows harder to ignore. Tools meant to augment thinking may instead train us to think less. The difference lies in deliberate design of both the technology and the habits surrounding it.
WebProNews is an iEntry Publication