For a generation that already lets GPS remember the way home and lets calculators handle the arithmetic, a new study from the MIT Media Lab adds an unsettling entry to the ledger of skills we are outsourcing to machines: the ability to tell real news from fake.
In research published June 9, 2026, MIT Media Lab scientists tracked 67 people over four weeks as they judged the credibility of news headline-image pairs. The headline finding was encouraging at first glance. When an AI chatbot was at their side, participants were 21 percent more accurate at flagging misinformation, confirming earlier work from the MIT Sloan School of Management showing that conversational AI can genuinely shake people loose from false beliefs.
Then the researchers took the chatbot away. By week four, participants' unassisted accuracy on fresh news items had fallen 15 percentage points below where they started, before they had ever touched the AI. The crutch had not just failed to make them stronger; it appeared to have left them weaker than when they began.
The pattern has a name in the literature: the "AI dependency paradox." It is a close cousin of what cognitive scientists call cognitive offloading, or automation complacency, the well-documented tendency to let a reliable tool absorb a mental task until the underlying skill atrophies. The MIT team draws an explicit line from calculators dulling mental math to GPS dulling our sense of direction, and points to a 2025 study in which doctors leaning on AI got worse at detecting cancer on their own.
"Users get excited about these 'magical' LLMs, but forget that they're just statistical models that predict the next 'token' in a sequence," said Anku Rani, an MIT media arts and sciences PhD student and co-lead author of the paper, alongside fellow PhD student Valdemar Danry. "Many impressive behaviors emerge from scaling this, but it comes with real limitations, both in what the model can reliably generate and in its broader impact on the people using it."
Perhaps the most striking detail is psychological. Roughly a quarter of participants reported feeling they were getting better at spotting fakes even as their measured performance slid, a textbook flicker of the Dunning-Kruger effect. The team labeled one-fifth of participants "Dependency Developers," people who drifted from actively reasoning through a story to passively accepting whatever the AI told them. One respondent captured the slide in a post-study survey, noting that while the chatbots stressed checking multiple sources, "they didn't teach me much about exploring the context of the images themselves."
Notably, the decline was concentrated in spotting fake news; accuracy on genuine news held roughly steady. The researchers warn that these models are especially fragile around emotionally charged breaking news, and that the human-written content used to train them is itself increasingly unreliable and biased.
Why it matters
The timing is pointed. Pew Research Center reporting over the past year found that one in five U.S. teens now regularly turn to large language models for news, and one in four young adults have done so at least once. If the tools millions are adopting to navigate an information environment thick with deepfakes are themselves eroding the discernment needed to use them, the long-run cost to information literacy could be steep, and largely invisible to the people experiencing it.
The study does suggest a way out, and it hinges on how the AI talks. The researchers found a clear split between a chatbot that acts "as a coach, versus as a crutch." Systems that simply hand over answers tended to breed dependence. Systems that used the Socratic method, asking guided questions, or "deep probing," gently nudging a user who was drifting toward a wrong call, fostered durable independent skill, even though they slowed people down in the moment.
"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," Danry said. "But it's very much a trade-off between speed and effort."
The paper, titled "Dialogues with AI Reduce Beliefs in Misinformation but Build No Lasting Discernment Skills," was presented at the 2026 CHI Conference on Human Factors in Computing Systems. Beyond Rani and Danry, it was co-authored by Assistant Professor Paul Pu Liang, Senior Research Scientist Andrew Lippman, and senior author Pattie Maes, the Germeshausen Professor of Media Arts and Sciences.
For Maes, the lesson is one for the classroom. "It's especially important to raise awareness in our schools and academic communities about the shortcomings of using AI as learning tools," she said. "People need to know that if they 'delegate' their thinking, they're not going to get better at that particular brand of problem-solving."
A few caveats are worth holding. The sample was small, just 67 people, drawn only from the United States and the United Kingdom, and tested against a modest set of roughly 50 validated news items over a single month. Rani says future work aims for more geographically diverse cohorts, including low-resource communities.
What to watch
Watch whether AI developers begin building "coach-mode" features, Socratic prompts and friction-by-design, into consumer chatbots, and whether educators fold these findings into media-literacy curricula. Watch, too, for replication on larger, more global samples, and for whether regulators or platforms take up the team's broader plea. As Danry put it: "We need to develop a new kind of AI literacy."
“Users get excited about these 'magical' LLMs, but forget that they're just statistical models that predict the next 'token' in a sequence.”— Anku Rani, PhD student and co-lead author, MIT Media Lab