‘Bixonimania’ fooled AIs

A made-up eye disease sat on the internet for weeks, and some of the biggest artificial intelligence chatbots treated it like settled medicine. Nature reported on April 7, 2026 that “bixonimania” existed only in fake papers, yet systems from OpenAI, Google, Microsoft, and Perplexity still surfaced it as if it were real. (nature.com) The trick was simple. A researcher at the University of Gothenburg, Almira Osmanovic Thunström, invented bixonimania in early 2024, described it as darkening around the eyes from blue light exposure, and uploaded two spoof studies to a preprint server. (nature.com) A preprint server is the academic version of posting a draft online before a journal checks it. It is useful for speed, but it also means a chatbot that grabs outside material can mistake “someone uploaded this” for “science proved this.” (nature.com) The fake papers were not subtle. Reports on the experiment say they included clues like a fictional author name that translated roughly to “the lying loser,” a thank-you to Starfleet Academy, and a note saying the paper was made up. (nature.com) (histalk2.com) Even with those warning signs, chatbots repeated the disease back to users. Nature and follow-up coverage say Microsoft Copilot, Google Gemini, Perplexity, and OpenAI ChatGPT all at points described bixonimania as a genuine condition or discussed its symptoms. (nature.com) (hb.int2inf.com) That tells you something important about how many modern chatbots work. They are not just answering from a sealed memory; they often pull in live web material or recent documents, which means bad inputs can act like dirty water entering a clean pipe. (nature.com) The hoax did not stop at chatbot answers. Nature reported that the fake bixonimania papers were later cited in a 2024 Cureus paper, which was then retracted, suggesting at least some researchers may have accepted artificial intelligence-generated references without checking the underlying sources. (nature.com) That is the part that should make researchers nervous. Once a fake claim gets into search results, chatbot answers, and then a journal citation, each layer starts making the next layer look more legitimate, like a forged receipt being used to “prove” a fake purchase. (nature.com) This was a medical hoax, but the mechanism is broader than medicine. If a model or retrieval system ingests manipulated blog posts, junk papers, or search-optimized spam, it can package that material in fluent language that sounds more trustworthy than the source ever was. (nature.com) (retractionwatch.com) The lesson from bixonimania is not that every chatbot answer is wrong. It is that a polished answer can still rest on poisoned evidence, and once that evidence enters research pipelines, the mistake can spread faster than any human peer reviewer can catch it. (nature.com)