AI Chatbots Cited Fake Disease Bixonimania as Real After 4 Planted Papers
Updated
Updated · Smithsonian Magazine · Jun 11
AI Chatbots Cited Fake Disease Bixonimania as Real After 4 Planted Papers
2 articles · Updated · Smithsonian Magazine · Jun 11
Summary
Copilot, Gemini and ChatGPT described the invented illness bixonimania as a real eye condition after researchers seeded it online in 2024 to test whether large language models could reject falsehoods.
4 planted items—2 Medium posts and 2 preprints—were enough to push the fake disease into chatbot answers, including responses to symptom queries about pink eyelids and blue-light exposure.
The hoax was laced with obvious clues, including an author name translating to “lying loser,” fake funders like the Galactic Triad, and one paper stating outright that “This entire paper is made up.”
Some scientists also cited the bogus preprints, underscoring that the failure was not limited to chatbots and highlighting how easily misinformation can enter both AI systems and academic workflows.
The experiment points to a broader risk: LLMs inherit internet misinformation at speed and can present it convincingly, making human verification more critical for health advice.
When AI can create convincing fake realities, how can we safeguard the integrity of medical and scientific truth?
If AI models are designed to be plausible, not truthful, can technical fixes ever truly solve their disinformation problem?
How the Bixonimania Hoax Fooled AI Chatbots and Scientific Journals: Lessons from a 2024 Medical Misinformation Experiment
Overview
In early 2024, the Bixonimania Experiment exposed major flaws in leading AI chatbots by introducing fake studies about a made-up eye condition called 'bixonimania.' Researchers created and uploaded these fabricated studies, describing the condition as caused by excessive screen time or blue light. Soon after, AI models began treating bixonimania as a real illness, failing to spot obvious signs of fraud. This experiment revealed how easily AI systems can absorb and spread misinformation when trained on unverified online content, highlighting the urgent need for better safeguards in AI and scientific publishing.