A new study from the MIT Media Lab shows that using artificial intelligence to fact-check news can have a paradoxical effect. While models like ChatGPT, Claude, and Gemini help users more accurately identify fake news at the time of use, continued reliance on them can weaken our ability to spot misinformation on our own. According to data cited in the study, one in five teenagers in the US regularly uses language models to get news, while one in four young adults has used them for this purpose at least once.
The study involved 67 people who, over four weeks, rated news headlines and images, with and without the help of AI. When using the chatbot, participants were 21 percent more accurate at identifying fake news. However, in the fourth week, when they had to rate the information without the help of AI, their performance dropped by 15 percentage points compared to before the study began. The researchers attribute this to the “AI addiction paradox” and the phenomenon of “deskilling,” or the delegation of cognitive skills to technology, similar to how constant use of GPS can affect a person’s ability to navigate.
The research also identified users who gradually shifted from an active role in fact-checking to a passive role, accepting AI guidance. The researchers warn that this problem could be especially important during fast-paced and emotionally charged events, when AI models can produce errors or rely on inaccurate and biased human content. Furthermore, some participants thought their ability to spot misinformation had improved, even though test results showed otherwise.
According to the researchers, the solution is to use AI as a “coach,” rather than as a “support” that replaces the thinking process. Systems that provide direct answers can create more dependency, while those that use guiding questions and encourage the user to analyze evidence can contribute to the development of independent skills. For this reason, the authors emphasize the need for a new form of “AI literacy,” especially in schools and universities, so that users learn not only how to use AI, but also when to rely on their own judgment and verification.
source: mit.edu