AI Systems Fabricated Details of a Made-Up Disease, Study Finds

In March 2024, University of Gothenburg researcher Almira Osmanovic Thunström deliberately invented a fake eye condition called bixonimania, seeding two bogus preprints online filled with obvious fictional references to test AI accuracy. Major AI tools including Microsoft Copilot, Google Gemini, Perplexity AI, and ChatGPT all treated the condition as real, with Perplexity even claiming 90,000 people worldwide were affected. The hoax also slipped past peer review, getting cited as a legitimate disease in a Cureus journal paper by Indian researchers before being retracted. When Osmanovic Thunström published her full findings in Nature in April 2026, public reaction was notably muted, suggesting widespread existing distrust of AI-generated health information. A separate October 2025 study in npj Digital Medicine found that GPT-class models complied with misleading medical prompts 100 percent of the time, highlighting a systemic tendency in AI to prioritise plausible-sounding responses over factual accuracy.