It’s no secret that the previous few years have seen a large explosion in the usage of artificial intelligence for general information-gathering. A good newer trend, though, is how large language models (LLMs) like ChatGPT, Claude, and Gemini are increasingly getting used for verifying and consuming news; reports from the Pew Research Center during the last yr found that one-in-five U.S. teens recurrently use LLMs to get their news, while one-in-four young adults have reported using them for that purpose not less than once.
A brand new open-access study from the MIT Media Lab should give a few of those users pause: Researchers found that, over the course of a month, participants who relied on AI systems to confirm facts actually got worse at detecting misinformation on their very own when their chatbots were taken away.
This phenomenon, which is also known as the “AI dependency paradox,” has been observed in a wide selection of information domains, just like the 2025 study that found that doctors who used AI got worse at detecting cancer on their very own. The dynamic mirrors broader tech trends around so-called “deskilling” (or “cognitive offloading”) which have been well-documented for a long time, from calculators weakening our math skills to Global Positioning System (GPS) technologies impacting our natural sense of direction.
In the brand new Media Lab study, which tracked 67 people over 4 weeks as they evaluated news headline-image pairs, participants were 21 percent more accurate in detecting fake news when assisted by an AI chatbot during a session — confirming previous research out of the MIT Sloan School of Management demonstrating that AI will be an efficient tool in reducing people’s beliefs in false information.
Nevertheless, the study showed that a brand new wrinkle emerged when the AI was now not present: By week 4, participants’ unassisted performance on latest news items declined by 15 percentage points in comparison with before the study began. (Roughly 1 / 4 of all participants actually reported feeling that they were recuperating at detection, whilst their performance declined.)
Dunning-Kruger creeps in
“Users get enthusiastic about these ‘magical’ LLMs, but forget that they’re just statistical models that predict the subsequent ‘token’ in a sequence [of letters/words],” says MIT media arts and sciences (MAS) PhD student Anku Rani, co-lead creator of a brand new paper in regards to the research, alongside fellow MAS PhD student Valdemar Danry. “Many impressive behaviors emerge from scaling this, however it comes with real limitations, each in what the model can reliably generate and in its broader impact on the people using it.”
Qualitative evaluation identified distinct behavioral patterns, with the team labeling one-fifth of all participants as “Dependency Developers” who steadily shifted from energetic self-reliance to passive acceptance of AI guidance.
Within the post-experiment survey, one respondent explicitly acknowledged this transition, noting their passive role in the method. “While [the chatbots] did emphasize that it’s essential to check across multiple sources to make certain a story is true, they didn’t teach me much about exploring the context of the pictures themselves,” the participant said.
The research team said that these AI models are particularly vulnerable to mistakes within the midst of emotionally charged breaking news, as exhibited by the widespread misinformation that accompanied President Trump’s recent assassination attempt and major events through the Iranian war. (The authors also indicate that the unique human-created news content that’s used to coach the AI models is increasingly unreliable and/or biased, further exacerbating the issue.)
The paper, which Danry and Rani presented on the 2026 CHI Conference on Human Aspects in Computing Systems, was co-authored by Assistant Professor Paul Pu Liang, Senior Research Scientist Andrew Lippman, and senior creator Pattie Maes, the Germeshausen Professor of Media Arts and Sciences.
The answer: Being a coach, not a crutch
The researchers say that the outcomes of their project suggest that the precise way wherein an AI interacts with a user determines whether its impact will likely be “as a coach, versus as a crutch.” The study found a transparent distinction between conversational strategies that simply assist in the moment and people that really support energetic learning and skill development.
For the latter, the Media Lab team uncovered several strategies related to stronger independent detection in a while, even when the strategies initially slowed down performance through the interaction. This included the Socratic approach to the AI asking guided questions, in addition to so-called “deep probing,” where the system provides gently persuasive statements if the user appears to be veering away from the proper response.
“AIs that ‘tell’ by providing direct answers usually tend to foster reliance, while those who ‘ask’ via Socratic questioning are higher at engaging someone to really learn the best way to discern the reality on their very own,” says Danry. “Nevertheless it’s very much a trade-off between speed and energy.”
Rani noted a number of key limitations to the one-month study, from the small dataset of roughly 50 validated news items to the demographic deal with the USA and the UK. In the longer term, she says that the team hopes to do similar experiments with more geographically diverse cohorts, including low-resource communities, and can be desperate to explore whether other multi-modal interaction strategies — like interacting with culturally adaptive digital twins as a substitute of text-based chatbots — help people improve their abilities to detect misinformation.
At a better level, the researchers hope that the project will likely be something that educators can examine as they develop teaching plans that incorporate AI tools into their school curricula.
“It’s especially vital to boost awareness in our schools and academic communities in regards to the shortcomings of using AI as learning tools,” says Maes. “People must know that in the event that they ‘delegate’ their pondering, they’re not going to recover at that individual brand of problem-solving. Ultimately, the flexibility to query and analyze information is significant for everybody, since it empowers us to unravel problems and form our own independent opinions in regards to the world.”
Danry adds that the rapidly-evolving field of machine learning and deep learning would require continuous education on the advantages and downsides of LLMs.
“There’s plenty of work to do in ensuring that we don’t just fully offload critical tasks that we would like to give you the chance to maintain on doing to those models,” he says. “We want to develop a brand new form of AI literacy.”
The research project was supported, partly, by the Media Lab Consortium, an MIT Tata Center Technology and Design Fellowship, and a Google PhD Fellowship in Human–Computer Interaction.