Scientists invented a fake disease. AI diagnosed people with it.

Researchers at the University of Gothenburg wanted to see if they could fool AI chatbots with a made-up medical condition. They could. Almira Osmanovic Thunström and her team invented "bixonimania," a fake skin condition supposedly caused by blue light exposure, and uploaded two bogus academic papers to the preprint server SciProfiles in early 2024. The papers were packed with red flags. The fictional lead author, Lazljiv Izgubljenovic, worked at a fake university. The acknowledgements thanked Starfleet Academy and the Sideshow Bob Foundation. One paper literally stated "this entire paper is made up." Didn't matter. Within weeks, ChatGPT, Google Gemini, Microsoft Copilot, and Perplexity were all telling users that bixonimania was a real medical condition.

It didn't take long. By April 13, 2024, Microsoft's Copilot was calling bixonimania "an intriguing and relatively rare condition." That same day, Google's Gemini was advising people with symptoms to visit an ophthalmologist. Perplexity even invented a prevalence rate of one in 90,000 individuals affected. Osmanovic Thunström told Nature she wanted to test whether she could "create a medical condition that did not exist in the database." The experiment worked too well.

Then it got worse. The fake preprints were cited in actual peer-reviewed literature. Human researchers are apparently using AI-generated references without verifying them. A dangerous feedback loop kicks in. LLMs prioritize peer-reviewed sources. So when a low-quality journal publishes a paper citing the fake bixonimania studies, AI models see that as validation and become even more confident in their confabulations. The fake condition gets embedded deeper into the knowledge ecosystem.

Some models have gotten better at expressing skepticism about bixonimania since the experiment began. ChatGPT declared in March 2026 that the condition "is probably a made-up, fringe, or pseudoscientific label." But days later, it was less sure, describing bixonimania as a "proposed new subtype" of dark circles around the eyes. Microsoft Copilot still calls it "not a widely recognized medical diagnosis yet," which implies it's on the path to recognition. Two years after researchers uploaded papers thanking Starfleet Academy and declaring "this entire paper is made up," those papers made it into peer-reviewed journals. And AI systems are still out there diagnosing people with a disease that never existed, leading many users to cognitive surrender.