In recent years, the use of artificial intelligence for gathering information has surged significantly. A newer trend involves large language models (LLMs) such as ChatGPT, Claude, and Gemini being utilized for news verification and consumption. According to the Pew Research Center, one in five U.S. teenagers frequently uses LLMs for news, and one in four young adults have done so at least once.
A study from the MIT Media Lab suggests caution for these users. The research found that participants relying on AI for fact-checking became less adept at identifying misinformation independently once the AI was removed. This “AI dependency paradox” is akin to findings in other fields, such as a 2025 study where doctors relying on AI struggled more with detecting cancer unaided.
The Media Lab study involved 67 individuals evaluating news headlines and images over four weeks. When aided by AI, participants were 21 percent more accurate in spotting fake news. However, by the end of the study, their accuracy without AI decreased by 15 percentage points compared to the beginning, even though some participants felt their skills had improved.
MIT PhD student Anku Rani explains that people are often captivated by LLMs, yet overlook their limitations. The research identified behavioral shifts, with some participants becoming overly reliant on AI. One participant noted that while chatbots advised checking multiple sources, they didn’t aid in understanding image contexts.
The study highlighted AI models’ vulnerability to errors during emotionally charged events, such as misinformation surrounding President Trump’s assassination attempt and key moments in the Iranian war. The study was co-authored by MIT’s Paul Pu Liang, Andrew Lippman, and Pattie Maes.
Researchers advocate for AI acting as a “coach” rather than a “crutch.” The study identified strategies promoting independent learning, like the Socratic method, which initially slows performance but boosts long-term skill development. Systems that ask questions are more effective than those providing direct answers, according to co-lead author Valdemar Danry.
Rani acknowledged limitations in the study, such as its small dataset and focus on U.S. and U.K. demographics. Future research aims to include diverse groups and explore different interaction strategies. The team hopes educators will incorporate AI tools thoughtfully into curricula, as Maes emphasizes the importance of critical thinking skills.
Danry notes the need for ongoing education on LLMs’ pros and cons, advocating for a new kind of AI literacy. The research was partially funded by the Media Lab Consortium, MIT Tata Center, and a Google PhD Fellowship.
Original Source: news.mit.edu
