People who leaned on chatbots to check facts grew worse at detecting misinformation on their own once the tools were removed, researchers found, a pattern with direct implications for schools.
People who rely on artificial intelligence to tell real news from fake get better at it while the chatbot is helping, then worse at doing it themselves once the tool is gone, according to a new study from the MIT Media Lab.
Researchers tracked 67 people over four weeks as they judged whether news headlines and images were genuine or fabricated, sometimes with an AI chatbot’s help and sometimes without. With a chatbot assisting, participants were 21 percent more accurate at catching false stories. The cost showed up later. By the fourth week, their unassisted accuracy on new items had slipped about 15 percentage points below where it had been before the study began.
The tool meant to help people spot lies, in short, left them less able to spot lies without it.
Researchers call the effect the “AI dependency paradox,” and they note it is not confined to news. A 2025 study found that doctors who relied on AI to read scans became worse at detecting cancer unaided. The pattern follows a long line of what scientists call deskilling, from calculators weakening mental arithmetic to satellite navigation dulling people’s sense of direction.
Confidence climbs as skill drops
One of the more striking results concerns how participants saw themselves. Roughly a quarter said they felt they were improving at detection even as their accuracy fell, a split between confidence and competence the researchers tied to the Dunning-Kruger effect.
The team singled out a subset, about one in five, who drifted over the month from checking claims themselves to passively accepting whatever the AI offered. The danger, the researchers argue, is not that AI provides answers but that leaning on it can train people to stop interrogating those answers, especially when the system sounds sure of itself. They add that the models tend to fail when it matters most, during fast-moving breaking news, when false claims travel fastest and the data used to train the systems is often unreliable to begin with.
Coach, not crutch
The study does not argue that AI has no place in learning. Its sharper claim is that the way a tool talks to a user largely decides whether it builds skill or eats away at it. The researchers cast the difference as AI working “as a coach, versus as a crutch.”
Systems that simply delivered a finished answer tended to breed dependence. Those that pushed back, posing guided questions in the manner of the Socratic method or nudging users who appeared to be drifting toward the wrong conclusion, did more to help people learn to judge information for themselves. That approach usually slowed users down in the moment, a trade-off between speed and real understanding. The field, co-lead author Valdemar Danry argued, needs to build “a new kind of AI literacy.”
Why it reaches the classroom
Although the participants were adults in the United States and Britain, the authors point straight to schools. Pew Research Center surveys over the past year found that one in five American teenagers regularly use chatbots for news, and one in four young adults have done so at least once. Children and teenagers are still forming the reading, reasoning and judgment the study suggests AI can quietly undercut.
Pattie Maes, the study’s senior author, said the results should shape how schools bring AI into teaching. People who “delegate” their thinking, she warned, do not get better at that kind of problem-solving, a skill she said matters because it lets people form their own conclusions about the world.
For parents, the message is less an argument for banning the technology than for paying attention to how it is used. A child who asks a chatbot to check their reasoning may gain from it. A child who uses it to avoid thinking may lose practice in the skills school exists to build.
An early and limited picture
The researchers were frank about the study’s bounds. The sample of 67 is small, the pool was only about 50 validated news items, and every participant came from the United States or the United Kingdom. Larger and more varied studies will be needed before the figures can be called settled. The paper was presented at the 2026 CHI Conference on Human Factors in Computing Systems.
Even so, the finding aligns with a widening body of research on cognitive offloading, and it arrives as chatbots push further into homework, search and everyday reading. The point the authors leave is not that AI makes people incapable, but that judgment, like any skill, weakens without use.
