If you have spent too much time staring at screens lately, you may have developed sore, itchy eyes, perhaps even a slight pinkish hue in your eyelids. If you asked one of the AI chatbots what was wrong with you, you might have been told you were suffering from bixonimania.
You should find this unsettling, even if you are not a doctor.
The reason is that bixonimania is not a real disease. It is the invention of Almira Osmanovic Thunström, a medical researcher at the University of Gothenburg in Sweden. In March 2024 she came up with a fictitious skin condition in online posts. Over the next two months, she uploaded two fake studies about it to a preprint server, where many medical studies are uploaded before they are sent to journals for peer review by experts.
Within weeks, major large language models began describing bixonimania as a real condition. In April 2024, Microsoft’s Copilot called it an intriguing and relatively rare condition. Google’s Gemini explained that it was caused by excessive exposure to blue light and advised users to consult an ophthalmologist. The Perplexity answer engine helpfully added a prevalence figure: one in ninety thousand individuals. ChatGPT used the term to diagnose users who described their symptoms. The chatbots had absorbed fake online material and relayed the invented diagnosis.
I first learned of bixonimania from a fascinating news feature published by Nature in April. My son and I, both curious, immediately did what most readers of the story would do. We asked ChatGPT about the condition.
The system has since wised up. Almost exactly a month after the Nature feature, the model gave us a different answer: “I don’t recognize bixonimania as a standard medical or psychological term. It may be a typo or misspelling, a niche internet slang term, or a joke or reference I’m not catching.”
Thunström had planted clues a human reader would have caught. She named the disease bixonimania because, as she told Nature, the suffix mania belongs to psychiatry, and no real eye condition would ever carry it. She attributed the papers to a fictional author named Lazljiv Izgubljenovic, who supposedly worked at the equally fictional Asteria Horizon University in Nova City, California. The acknowledgments thanked Professor Maria Bohm at the Starfleet Academy for her work aboard the USS Enterprise. The funding, the papers said, came from the Professor Sideshow Bob Foundation, in support of its work in advanced trickery.
Large language models are trained on enormous swathes of the internet. Some of those materials are high quality, some are junk, and most fall somewhere in between. These models rely on automated filters and on the broad assumption that academic literature is more reliable than the average web page. Preprint servers sit inside this assumption even though they are not peer reviewed.
In his book A Giant Leap, Dr. Robert Wachter cites Bob Kocher, a venture investor and former health official, who calls this kind of contamination “data poisoning”. Osmanovic Thunström’s experiment poisoned a database deliberately, with ethics oversight and with markers throughout the fake papers designed to demonstrate the vulnerability without doing serious harm.
A bad actor with commercial or political motives would not be so careful. We are already awash in fads, fake cures, and quacks. If you can poison the right places on the internet, you can make your way into AI chatbots that people increasingly rely on for medical advice.
There is a deeper malaise here too. The fake bixonimania material did not only corrupt AI chatbot responses. One of the fake papers was cited in a peer-reviewed paper that described bixonimania as an emerging form of periorbital melanosis linked to blue light exposure. The paper was retracted only after Nature contacted the journal. The journal’s editor noted that, in the wake of the retraction, the editorial staff no longer had confidence in the accuracy or provenance of the work. Good, but we might not be so lucky if the paper is not so blatantly faked and part of an experiment.
This is the new world we live in. We are likely to see many more examples like this. The danger we face is not only that chatbots can be fooled, but also that people may stop exercising their own judgment, assuming that a confident AI answer is a correct one. In medicine, science, and public life, that can prove to be a costly error.
Anirban Mahapatra is a scientist and author. His most recent book is When the Drugs Don’t Work. The views expressed are personal.
