Relying on AI to check the truth of news stories may help people catch misinformation in the short term, but it can also leave them less able to do the job on their own later, according to new research from MIT’s Media Lab.
In a study published as open-access research, MIT researchers tracked 67 participants over four weeks as they judged headline-image combinations. When people used an AI chatbot during the task, they became more accurate at identifying fake news. But once the chatbot was removed, their independent performance declined over time.
The findings add to growing concern about what researchers describe as an AI dependency paradox. The idea is familiar from earlier technologies that make tasks easier while slowly eroding the skill needed to do them unaided. The study’s authors compare the effect to how GPS may weaken a person’s sense of direction or how calculators can dull mental math.
According to the paper, participants were about 21 percent more accurate at detecting false news when an AI assistant was available. That result aligns with earlier research suggesting that AI can reduce belief in misleading information during an interaction. But the picture changed when users had to rely on themselves. By the end of the month, their unassisted accuracy on new stories had fallen by 15 percentage points compared with where they started.
Some participants seemed unaware that their skills were slipping. About a quarter said they believed they were improving, even as the data showed the opposite. The researchers also identified a group they called "Dependency Developers," people who gradually moved from actively evaluating content to accepting the chatbot’s guidance with less scrutiny.
The study’s authors say the risk is not that AI always makes people worse at evaluating news, but that the design of the interaction matters. Systems that simply provide the answer may encourage users to lean on them. In contrast, chatbots that ask guided questions or prompt people to explain their reasoning appear more likely to support learning.
One of the researchers, MIT media arts and sciences doctoral student Anku Rani, said users may be impressed by large language models but overlook their limits, including the fact that they are statistical systems rather than truth engines. Fellow co-lead author Valdemar Danry said the goal should be to make AI act more like a coach than a crutch.
The team pointed to Socratic questioning and a method they call deep probing as approaches that may better support long-term skill building, even if they slow users down in the moment. The trade-off, the researchers said, is between speed and effort.
The study also highlights a particular challenge in fast-moving and emotionally charged news environments. The authors argue that AI systems can be vulnerable to errors when information is developing quickly, and that the human-made content used to train them can itself be flawed or biased.
The work was presented at the 2026 CHI Conference on Human Factors in Computing Systems and co-authored by MIT researchers Paul Pu Liang, Andrew Lippman and Pattie Maes, along with Danry and Rani.
The authors acknowledge limits to the project, including its relatively small pool of validated news items and its focus on participants in the United States and the United Kingdom. They say future studies should test more diverse groups and explore whether other kinds of interactive tools can better help people resist misinformation.
For now, the researchers say schools and other educators should think carefully about how AI is introduced in learning settings. Their broader message is that if people delegate critical thinking to chatbots, they may not build the independent judgment needed to question information for themselves.