MIT Study: AI Reliance Can Weaken Human Fake News Detection Skills

A study conducted by the MIT Media Lab reveals potential drawbacks to relying on artificial intelligence (AI) for verifying news, suggesting that it may impair individuals' natural ability to detect misinformation.

The research observed that participants who depended on AI systems for fact-checking gradually became less adept at identifying fake news independently once AI assistance was removed. This effect is referred to as the “AI dependency paradox,” mirroring broader trends of “deskilling” or “cognitive offloading” seen with technologies like calculators and GPS systems.

Over four weeks, the study tracked 67 individuals as they evaluated news headline-image pairs. Initially, participants demonstrated a 21 percent improvement in accuracy when assisted by an AI chatbot, confirming prior research that AI can help reduce belief in false information.

However, a significant finding emerged when AI was no longer present. By the fourth week, participants' unassisted performance on new news items dropped by 15 percentage points compared to their pre-study capabilities. Interestingly, approximately a quarter of all participants believed their detection skills were improving, even as their actual performance declined.

Anku Rani, an MIT media arts and sciences (MAS) PhD student and co-lead author of the paper, noted that large language models (LLMs) are statistical models with inherent limitations, both in what they can reliably generate and their impact on users. The research team identified distinct behavioral patterns, labeling one-fifth of participants as "Dependency Developers" who transitioned from self-reliance to passively accepting AI guidance.

One participant in a post-experiment survey highlighted this shift, stating that while chatbots encouraged cross-referencing sources, they did not teach much about exploring the context of images. The research also suggests AI models are particularly vulnerable to errors during emotionally charged breaking news events, and that the increasing unreliability or bias in human-created content used for AI training further exacerbates these issues.

The paper, co-authored by Rani, fellow MAS PhD student Valdemar Danry, Assistant Professor Paul Pu Liang, Senior Research Scientist Andrew Lippman, and senior author Pattie Maes, was presented at the 2026 CHI Conference on Human Factors in Computing Systems.

According to MIT News AI, the widespread use of LLMs for news consumption, including by one-in-five U.S. teens and one-in-four young adults, underscores the relevance of these findings.