AI reduces our ability to detect fake news

What happened?

A study from the MIT Media Lab, published in June 2026, has put numbers to a growing fear: reliance on artificial intelligence to verify news is weakening our own ability to detect false information. Over four weeks, 67 participants evaluated pairs of headlines and images, some with the help of an AI chatbot and others without. The results show that although AI improves momentary accuracy by 21% (confirming previous research from the MIT Sloan School of Management published in Science in 2025), when the assistant is removed, users' ability drops by 15% compared to their initial level. That is, they become worse at detecting fake news than before using the tool.

The open-access study identifies a behavioral pattern: approximately one in five participants becomes a 'dependency developer,' shifting from active critical thinking to passive acceptance of AI suggestions. In follow-up surveys, some users admitted feeling more confident in their ability to detect hoaxes, even though their objective results worsened—a clear Dunning-Kruger effect induced by technology. This phenomenon, dubbed the 'AI dependency paradox,' has been observed in other fields: a 2025 study in The Lancet Gastroenterology & Hepatology found that radiologists using AI to detect colorectal cancer performed worse without the tool. Similarly, GPS has atrophied our sense of direction for decades, and calculators have weakened basic math skills.

Why is it important?

This finding is not isolated. Misinformation is a national security and public health issue. According to the Pew Research Center, one in five American teenagers (20%) regularly uses LLMs like ChatGPT, Claude, or Gemini to get informed, and one in four young adults (25%) has done so at least once. If citizens delegate news verification to systems that, like LLMs, are prone to hallucinations and biases (e.g., inventing sources or favoring certain narratives), the democratic fabric becomes more vulnerable. The study underscores that AI not only fails to make us more resistant to fake news in the long run but makes us more dependent and less critical. In a context where the 2026 elections in several countries (Brazil, Mexico, United States) already face disinformation campaigns, the implications are massive.

Furthermore, the study reveals that about a quarter of participants (25%) showed a significant decrease in their verification ability without help, while a small group maintained or improved their performance, suggesting that the impact varies by user. Researchers also noted that participants who trusted AI the most were the ones who worsened the most, aggravating the problem: less critical users become even more vulnerable.

Consequences and outlook

In the short term, we can expect an increase in misinformation during crises, such as elections or natural disasters, where users blindly trust AI-generated responses. For example, during the 2025 hurricane in Florida, hoaxes about shelters were spread and amplified by poorly configured chatbots. In the medium term, tech companies may face pressure to redesign their assistants, incorporating 'critical thinking training' instead of just providing answers. Some startups are already experimenting with modes that force users to verify sources before displaying a response. In the long term, media literacy will need to integrate conscious use of AI as a basic competency, similar to how evaluating sources on the internet is taught.

The study also opens the debate on platform responsibility: if their tools degrade human skills, should they be regulated? The European Union, in its AI Act, already classifies news verification systems as 'limited risk,' but some experts are calling for warning labels similar to those on medications: 'Prolonged use of this assistant may reduce your ability to detect fake news.' In the United States, the FTC has shown interest in such warnings, though no concrete legislation exists yet.

What readers should know

• Don't trust blindly: LLMs are statistical models, not oracles. Always verify with primary sources and cross-check with multiple sources. A Stanford University study (2025) showed that ChatGPT has a 19% hallucination rate on current events.
• Train your judgment: Use AI as a support tool, not a substitute. Practice manual hoax detection with sites like Snopes or FactCheck.org. Researchers recommend weekly verification exercises without help.
• Beware of overconfidence: If you feel AI has made you an expert, you're likely in the Dunning-Kruger phase. The study found that the most confident users were the ones who had worsened the most.
• Demand transparency: News apps with AI should report their error rates and offer modes that foster critical thinking, such as showing the sources used and allowing manual verification.

“Users get excited about these 'magical' models, but forget they only predict the next word,” warns Anku Rani, co-author of the study. “They have real limitations, both in what they generate and in their impact on those who use them.” The study concludes that AI can be a crutch that, when removed, leaves the user more lame than before.

In summary, the AI dependency paradox is not an academic curiosity: it is a tangible threat to the integrity of information in the digital age. Users, companies, and regulators must act to prevent the tool that promised to protect us from misinformation from becoming its best ally.