Study: Relying on AI to Check Facts Makes You Worse at Spotting Misinformation
Researchers at MIT's Media Lab found that people who use AI chatbots to verify news become worse at detecting fake stories on their own. After a month of relying on ChatGPT and similar tools, participants' unassisted performance declined by 15 percentage points when the AI was removed.
The study tracked 67 people over four weeks as they evaluated news headlines and images. While AI assistance improved accuracy by 21 percent during the sessions, the gains evaporated once the tool disappeared.
This pattern mirrors a broader phenomenon. GPS weakened our navigation skills. Calculators diminished mental math ability. Now AI is doing the same for critical thinking about information.
How the dependency develops
Researchers identified a group they called "Dependency Developers"-about one-fifth of participants who gradually shifted from active self-reliance to passively accepting whatever the AI suggested. One participant explicitly acknowledged this shift after the study ended, saying the chatbot didn't teach them to examine image context, only to check multiple sources.
The problem compounds during breaking news. AI models struggle most with emotionally charged events, as seen during recent major news events where misinformation spread widely. The underlying training data itself is increasingly unreliable and biased, researchers noted.
Roughly a quarter of all participants reported feeling they were improving at detecting misinformation even as their actual performance declined-a classic case of overconfidence masking skill loss.
The difference between a coach and a crutch
Not all AI interactions produce the same outcome. How an AI engages with users determines whether it functions as a learning tool or a dependency.
AI systems that simply provide direct answers foster reliance. Those that ask questions-using the Socratic method-engage users in actual learning. Deep probing, where the system gently redirects users away from incorrect conclusions, also strengthened independent detection skills later on.
"AIs that 'tell' by providing direct answers are more likely to foster reliance, while those that 'ask' via Socratic questioning are better at engaging someone to actually learn how to discern the truth on their own," said Valdemar Danry, a co-author of the research.
The trade-off is real: questioning strategies initially slow down performance compared to direct answers, but they build lasting skills.
What this means for newsrooms and communicators
One in five U.S. teens and one in four young adults now use LLMs to consume news, according to the Pew Research Center. Those audiences are getting worse at detecting what's real.
For communications professionals, this matters. Misinformation spreads faster when people lack the skills to question it. The researchers emphasized that understanding how AI tools affect critical thinking should inform how schools and organizations incorporate these systems.
Pattie Maes, senior author of the study, said people need to understand the cost of delegating their thinking. "The ability to question and analyze information is important for everyone, because it empowers us to solve problems and form our own independent opinions about the world."
Danry added that the field needs "a new kind of AI literacy"-ongoing education about both the benefits and limits of language models, especially for tasks where human judgment remains essential.
Study limitations and next steps
The research had constraints. The dataset included roughly 50 validated news items, and participants came primarily from the United States and United Kingdom. Future studies will test whether findings hold across more geographically diverse populations and lower-resource communities.
Researchers also want to explore whether other interaction methods-such as culturally adaptive digital twins instead of text-based chatbots-produce different outcomes.
Understanding how to design AI systems that teach rather than replace human judgment will become increasingly urgent as these tools proliferate in newsrooms, schools, and workplaces. The difference between effective AI assistance and harmful dependency often comes down to design choices made by the people building these systems.
Your membership also unlocks:
