A new study from the MIT Media Lab shows that generative AI can provide immediate assistance in identifying misinformation, but this assistance does not necessarily translate into more stable judgment skills. Participants performed better with AI assistance but showed a decline when removed from the tool.
Over 7,000 conversations recorded during the four-week experiment.
This study lasted four weeks and tracked 67 participants, recording 7,203 conversations with AI and 4,536 judgments on the authenticity of news. Participants first independently assessed the truthfulness of news headlines and accompanying images, then discussed them with AI, and finally revised their conclusions.
The research team built a detection system that combined OpenAI’s GPT-4o with Google Search to help participants verify news content. After the experiment, the team conducted a non-AI test to observe whether participants’ ability to make independent judgments about new content had improved.
Accuracy improves by 21% with AI assistance.
The results show that participants' accuracy in identifying misinformation improved by 21% when using AI. This demonstrates that chatbots can effectively assist users in detecting suspicious content, particularly in scenarios where true and false news headlines and images are mixed.
However, during the subsequent no-assistance test phase, participants' overall performance dropped by 15.3 percentage points. Researchers noted that the decline was primarily due to reduced ability to identify fake news, while accuracy in judging real news remained largely unchanged.
Research points to dependency risk
The research team believes that current AI tools are more akin to helping users correct individual judgments rather than training them to develop more stable discernment skills. In other words, AI may improve immediate outcomes without simultaneously enhancing human independent thinking, potentially making some users more reliant on the tools.
Researchers also used Anthropic’s Claude 3.5 Sonnet to analyze thousands of user-AI conversations. The paper concludes that current designs tend to make judgments on behalf of users rather than teaching users how to make those judgments themselves.
Fake war videos intensify concern
This study is being released as generative AI is significantly lowering the barrier to creating forged content. More realistic images, videos, and text are accelerating the spread of misinformation on social platforms.
The article mentions that after Iran's missile strike on Israel in June 2025, multiple videos allegedly filmed in Tel Aviv and Ben Gurion Airport circulated on social media, claiming to show war damage. These videos garnered millions of views before being confirmed as AI-generated.
Concerns over the falsification of war imagery continue to rise. In March, X announced that creators who post AI-generated conflict videos without proper labeling will be suspended from the platform’s revenue-sharing program.