It’s no secret that the past few years have seen a massive surge in the use of artificial intelligence in general intelligence gathering. However, a more recent trend is that large-scale language models (LLMs) such as ChatGPT, Claude, and Gemini are increasingly being used for news validation and consumption. A Pew Research Center report last year found that 1 in 5 U.S. teens regularly use LLM to get news, and 1 in 4 young adults reported having used LLM for that purpose at least once.
A new open access study from the MIT Media Lab should give some of these users pause. The researchers found that participants who relied on AI systems for fact-checking actually became worse at detecting misinformation on their own over a one-month period after the chatbot was removed.
This phenomenon, often referred to as the “AI-dependence paradox,” has been observed across a wide range of knowledge areas, such as a 2025 study that found that doctors using AI were less able to detect cancer on their own. The move reflects a broader technology trend around so-called “de-skilling” (or “cognitive offloading”) that has been well-documented for decades, from calculators that weaken our math skills to Global Positioning System (GPS) technology that affects our natural sense of direction.
A new Media Lab study that tracked 67 people over four weeks as they evaluated news headline and image pairs found that participants who were assisted by an AI chatbot during their sessions were 21 percent more accurate at detecting fake news. This supports previous research from the MIT Sloan School of Management that demonstrated that AI can be an effective tool in reducing people’s beliefs about misinformation.
However, this study showed that a new wrinkle emerged when AI was no longer present. By the fourth week, participants’ unaided performance on new news items was 15 percentage points lower than before the study began. (About a quarter of all participants actually reported that they felt their detection ability had improved, despite the decrease in performance.)
Dunning Kruger sneaks up
“Users get excited about these “magical” LLMs, but forget that they are just statistical models that predict the next “token” in a sequence. [of letters/words]” says MIT Media Arts & Sciences (MAS) doctoral student Ankh Rani, co-lead author of a new paper on the study, along with fellow MAS doctoral student Valdemar Danry. “Scaling this up reveals a lot of impressive behavior, but there are real limits, both in what the model can reliably produce and in the broader impact on the people who use the model.”
Qualitative analysis identified a clear pattern of behavior, with the team classifying one-fifth of all participants as “dependent developers,” gradually moving from active independence to passive acceptance of AI guidance.
In the post-experiment survey, one of the respondents explicitly acknowledged this transition and referred to their passive role in the process. “meanwhile [the chatbots] “It emphasized the need to check multiple sources to confirm a story is true, but it didn’t tell us much about exploring the context of the image itself,” the participant said.
The researchers said these AI models are especially prone to mistakes during emotionally charged breaking news events, as demonstrated by the proliferation of misinformation surrounding President Trump’s recent assassination attempt and major events during the Iran war. (The authors also note that the original human-generated news content used to train AI models is becoming increasingly unreliable and biased, further exacerbating the problem.)
The paper, presented by Danley and Lani at the 2026 CHI Conference on Human Factors in Computing Systems, was co-authored by Assistant Professor Paul Phu Lian, Senior Researcher Andrew Lipman, and senior author Patty Mayes, Germeshausen Professor of Media Arts and Sciences.
Solution: Become a coach, not a crutch.
The researchers say the project’s results suggest that the specific way AI interacts with users will determine whether its impact is achieved “as a coach or as a crutch.” This study found a clear difference between conversation strategies that are only useful in the moment and those that actually support active learning and skill development.
In the latter case, the Media Lab team discovered several strategies that were associated with stronger independent detection later on, even if they initially resulted in slower performance during interactions. This includes the AI’s Socratic approach of asking guided questions, and so-called “deep probing,” where the system provides gentle persuasive remarks if the user is about to deviate from the correct response.
“AI that ‘tells’ by providing direct answers is more likely to foster trust, while AI that ‘asks’ through Socratic questions is better at engaging people in actually learning how to discern the truth for themselves,” Danley says. “But it’s a trade-off between speed and effort.”
Lani pointed out that the one-month survey had several important limitations, ranging from a small dataset of about 50 verified news items to demographics focused on the United States and United Kingdom. In the future, she said, the team would like to conduct similar experiments with more geographically diverse cohorts, including in low-resource communities, and is also keen to explore whether other multimodal interaction strategies, such as interacting with culturally adapted digital twins instead of text-based chatbots, can help improve people’s ability to detect misinformation.
At a higher level, the researchers hope this project will be something that educators can consider when creating instructional plans that incorporate AI tools into school curricula.
“It is especially important to raise awareness in schools and academic communities about the drawbacks of using AI as a learning tool,” Mace says. “People need to know that ‘delegating’ their thinking does not improve their specific problem-solving abilities. After all, the ability to question and analyze information is important for everyone, because it gives us the power to solve problems and form our own independent opinions about the world.”
Danry added that the rapidly evolving field of machine learning and deep learning requires continued education on the pros and cons of an LLM.
“There’s a lot of work to be done to avoid completely offloading the critical tasks that we want these models to continue to perform,” he says. “We need to develop a new kind of AI literacy.”
This research project was supported in part by the Media Lab Consortium, an MIT Tata Center Technology and Design Fellowship, and a Google PhD Fellowship in Human-Computer Interaction.
