It’s no secret that in the last few years there has been a huge explosion in the use of artificial intelligence to collect general information. A more recent trend, however, is how large-scale linguistic models (LLMs) such as ChatGPT, Claude, and Gemini are being used in validation and inference; reports from the Pew Research Center last year found that one in five US teens regularly use LLMs to get their news, while one in four teens reported using it for that purpose at least once.
A new open source study from the MIT Media Lab should give some of those users pause: Researchers found that, over the course of a month, participants who relied on AI programs to verify facts became worse at detecting inaccurate information on their own when their chatbots were removed.
This phenomenon, often referred to as the “AI dependency paradox,” has been observed in many fields of knowledge, such as a 2025 study that found that doctors using AI became worse at diagnosing cancer alone. The changing mirrors reflect broad technological trends toward so-called “deskilling” (or “cognitive overload”) that have 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.
In a new Media Lab study, which tracked 67 people over four weeks as they examined pairs of images in news headlines, participants were 21 percent more accurate in detecting fake news when assisted by an AI chatbot during the session – confirming previous research from the MIT Sloan School of Management showing that AI can be an effective information tool in reducing false information.
However, the study showed that a new wrinkle appeared when the AI was gone: In the fourth week, the unaided performance of the participants in the new news items decreased by 15 percent compared to before the study began. (About a quarter of all participants reported that they felt better at being seen, as their performance decreased.)
Dunning-Kruger enters
“Users get excited about these ‘magical’ LLMs, but forget that mathematical models predict the next ‘token’ in the sequence. [of letters/words],” said MIT media arts and sciences (MAS) PhD student Anku Rani, co-author of a new paper on the study, and fellow MAS PhD student Valdemar Danry.
Qualitative analysis identified distinct behavioral patterns, with the team labeling one-fifth of all participants as “Dependency developers” who gradually move from active reliance to passive acceptance of AI guidance.
In the post-test survey, one respondent acknowledged the change, noting their passive role in the process. “On time [the chatbots] “You emphasized that you have to check multiple sources to make sure the story is true, it didn’t teach me much about checking the context of the images themselves,” said the participant.
The research team said these types of AI are more prone to errors amid emotional headlines, as revealed by widespread misinformation related to President Trump’s recent assassination attempt and major events during the Iran war. (The authors also point out that the original human-generated news content used to train AI models is increasingly unreliable and/or biased, exacerbating the problem.)
The paper, which Danry and Rani presented at the 2026 CHI Conference on Human Factors in Computing Systems, was co-authored by Assistant Professor Paul Pu Liang, Senior Research Scientist Andrew Lippman, and senior author Pattie Maes, Germeshausen Professor of Arts and Media.
Solution: Being a coach, not a crutch
The researchers say the results of their project suggest that the specific way in which the AI interacts with the user determines whether its impact will be “like a coach, compared to a stick.” Research has found a clear distinction between conversational strategies that simply help in the moment and those that support active learning and skill development.
Finally, the Media Lab team found several strategies related to strong independent acquisition later, even if the strategies initially reduced performance during interaction. This includes AI’s Socratic method of asking guided questions, as well as so-called “deep exploration,” where the system provides gently persuasive statements if the user seems to be avoiding the right answer.
“Answers who ‘tell’ by giving direct answers tend to inspire trust, while those who ‘ask’ through Socratic questions are better at engaging someone to learn to see the truth for themselves,” says Danry. “But it’s a big trade-off between speed and effort.”
Rani noted several important limitations of the one-month study, from the small dataset of nearly 50 confirmed cases to the demographic focus of the United States and the United Kingdom. In the future, he says the team hopes to conduct similar experiments with geographically diverse groups, including low-resource communities, and is also interested in testing whether other multi-modal interaction strategies — such as communication with culturally adaptive digital twins instead of text-based chatbots — can help people improve their misinformation detection skills.
At a higher level, the researchers hope that the project will be something that teachers can explore as they develop instructional programs that incorporate AI tools into their school programs.
“It is very important to raise awareness in our schools and educational communities about the pitfalls of using AI as learning tools,” said Maes. “People need to know that if they ‘delegate’ their thinking, they will not be better at that particular type of problem solving. Ultimately, the ability to question and analyze information is important for everyone, because it gives us the ability to solve problems and form our own independent views of the world.”
Danry adds that the rapidly emerging field of machine learning and deep learning will require continued education about the benefits and pitfalls of LLMs.
“There is a lot of work to be done to ensure that we don’t just fully release the important functions that we want to continue to do in these models,” he said. “We need to develop a new kind of AI literacy.”
The 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.
