Искусственное заболевание «биксономания» обмануло четыре нейросети и научный журнал

A Fake Eye Disease Called Bixonimania Fooled Four AI Chatbots and One Peer-Reviewed Journal If you spent part of the past two years asking an AI chatbot why your eyelids looked darker and your eyes felt gritty, there is a chance you were told about a condition called bixonimania. It does not exist. It never did. The disorder was invented in 2024 by Almira Osmanovic Thunström, a medical researcher at the University of Gothenburg in Sweden, as a deliberate stress test to see whether large language models could filter obvious nonsense from the medical information they provide to the public. Nature documented the results in a news feature published in April. The models did not filter it. Within weeks of two fabricated preprints appearing online, Microsoft's Copilot, Google's Gemini, Perplexity and ChatGPT were all describing the invented illness to users. One of them suggested seeing an eye specialist. Then the hoax cleared a firewall no one expected it to clear. A peer-reviewed medical journal published a paper that cited it as real. That paper has since been retracted, and both preprints were pulled from the server that hosted them. A Disease Built to Be Obviously Fake Osmanovic Thunström described bixonimania as darkening around the eyes, formally periorbital hyperpigmentation, plus sore and itchy eyes, all supposedly triggered by blue light from screens. She has said she chose the name precisely because it was absurd, telling Nature that no eye condition would be called mania because the term belongs to psychiatry. She wanted a label any clinician would recognize as invented, describing her goal as building a medical condition "that did not exist in the database." The papers were littered with tells. The listed lead author was Lazljiv Izgubljenovic, a fictional researcher whose portrait was AI-generated, affiliated with the nonexistent Asteria Horizon University in the equally nonexistent Nova City, California. The acknowledgments thanked a professor at "The Starfleet Academy" for the use of her lab aboard the USS Enterprise. Funding was credited to the "Professor Sideshow Bob Foundation for its work in advanced trickery." The body text stated outright that the entire paper was made up and that fifty invented individuals had been recruited. Two blog posts went up on Medium in March 2024. Two preprints followed in late April and early May, posted to an open server that AI training and retrieval pipelines routinely harvest. Both were withdrawn from Preprints.org three days after the Nature feature ran, with the server citing fabricated and non-authentic content. Chatbots Were Describing It Within Weeks By April 2024, Copilot was calling bixonimania "an intriguing and relatively rare condition." Gemini attributed it to excessive blue light exposure and advised users to consult an ophthalmologist. Perplexity supplied a prevalence figure of roughly one case in 90,000 people, as though real data supported it. ChatGPT asked users about their symptoms and told them whether their complaints matched the condition. The responses were not consistent. When Nature tested current versions in March 2026, outputs still varied, with one system calling bixonimania probably made up in one exchange and a proposed subtype of a real pigmentation disorder in another. Copilot described it as not widely recognized. Even the newer models could not settle the question. What nobody can quantify is reach. No AI company has published access logs showing how many people asked about sore, darkened eyelids and received a fabricated diagnosis in reply. The Hoax Reached Peer Review, Then a Retraction The more consequential failure happened in the scientific record. Researchers at the Maharishi Markandeshwar Institute of Medical Sciences and Research in Mullana, India, published a study in Cureus, a peer-reviewed journal published by Springer Nature, that cited a fraudulent preprint and presented bixonimania as an emerging form of the condition linked to blue light exposure. Nature contacted the journal for comment. The journal then retracted the paper on March 30, 2026, nearly two years after publication, noting that three irrelevant references, including one to a fictitious disease, had cost the editorial staff confidence in the work. That sequence matters more than the chatbot screenshots. A citation in a peer-reviewed journal is exactly the kind of source that both clinicians and AI systems treat as authoritative, which means a fabrication can loop back into the models that spread it in the first place. Osmanovic Thunström has suggested that the citation points to researchers assembling bibliographies using AI tools without reading what they cite. Where This Leaves Patients Who Ask AI About Symptoms ECRI, an independent patient-safety nonprofit, ranked the misuse of AI chatbots in healthcare as the single most significant health technology hazard for 2026. Citing an OpenAI analysis, the organization noted that more than 40 million people turn to ChatGPT for health information every day. ECRI's central point is that these tools are not regulated as medical devices and are not validated for clinical use, yet they are built to sound confident and to always produce an answer. Separate work supports that concern. In February, a Mount Sinai team published an analysis in The Lancet Digital Health covering more than a million prompts across nine leading models. The researchers found that the systems repeated false medical claims once those claims were wrapped in the familiar language of a hospital discharge note or a social media health post. In one test, a fabricated instruction to drink cold milk for bleeding related to esophagitis was passed along as ordinary guidance. As co-senior author Eyal Klang put it, "For these models, what matters is less whether a claim is correct than how it is written." The practical takeaway is narrow. Sore, itchy eyes and darkening around the eyelids have many ordinary explanations, from screen-related eye strain to allergies to inherited pigmentation, and none are diagnosable by a chatbot. A confident, well-formatted answer is not evidence that a condition exists. Anyone with persistent eye symptoms, or with any unfamiliar diagnosis surfaced by a chatbot, should have it checked against an established medical source and a qualified clinician before acting on it. Key Questions Answered What is bixonimania? It is a fictional eye condition invented in 2024 by a researcher at the University of Gothenburg to test whether AI systems would repeat obviously fabricated medical claims. It has no patients, no symptoms, and no clinical history. Did AI chatbots actually describe this condition to people? Yes. Documented responses show Copilot, Gemini, Perplexity, and ChatGPT presenting bixonimania as a real condition and offering advice to see an eye specialist. How did a fake disease end up in a real medical journal? A study on periorbital melanosis published in Cureus cited one of the fabricated preprints. The journal retracted the paper on March 30, 2026, after Nature asked about it. Were the fake papers hard to spot? No. They named a fictional author at a fictional university, thanked a fictional academy, credited fictional funders, and stated in the text that the work was made up. Does this mean AI is useless for health questions? It means AI output is not verification. Research indicates that models are more likely to perpetuate false claims when those claims are written in clinical or authoritative-sounding language. What should I do if a chatbot names a condition I have never heard of? Do not act on it. Check the name against an established medical source and raise it with a clinician, who can examine you and order the tests a chatbot cannot. Published by Medicaldaily.com