Today, if your patients ask Google or Gemini or Chat if they have bixonimania, the answer is: “No, you do not have bixonimania. Bixonimania is a fabricated condition. It was invented in 2024 by Swedish medical researcher Almira Osmanovic Thunström. The experiment tested how easily AI systems spread medical misinformation.”
But until early April, AI chatbots were apparently amplifying incorrect information, including images, of this made-up condition. Nature magazine took a dive into how the condition took on a life of its own.
The researchers who planted this gave plenty of clues that this was a fabrication. The acknowledgements thanked Professor Maria Bohm at The Starfleet Academy on the USS Enterprise and noted that funding came from the Professor Sideshow Bob Foundation for its work in advanced trickery. At one point, the paper even said, “This entire paper is made up.”
Still, it fooled large language models.
The Impact on Health Care Providers
Let’s put aside the ridiculous name, which leans heavily on “mania,” and look at the more serious question. Where does this leave patients and potentially the health care providers who need to correct the record?
The health care research organization, ECRI Institute, lists navigating the AI diagnostic dilemma as the number one patient safety concern in 2026. In its report listing the top 10 concerns, the authors pointed out that tested machine learning models failed to recognize 66% of critical or deteriorating health conditions and injuries in synthesized cases.1 And while generative AI models more accurately diagnosed genetic conditions based on textbook-like descriptions, “their accuracy dropped significantly when prompts were based on a conversation with a simulated patient,2,3 suggesting that AI models struggle with open-ended diagnostic reasoning.” The full ECRI report can be downloaded here.
The report also noted a study that says that 66% of physicians survey reported using AI in 2024, a large increase over the 38% who said they used AI in 2023.4 This can contribute to errors based on perpetuated biases, incorrect diagnoses, and a lack of transparency.
This story emphasizes the real reason why “human in the loop” will always be paramount and experience cannot be replaced.
References
- Pias TS, Afrose S, Tuli MD, et al. Low responsiveness of machine learning models to critical or deteriorating health conditions. Commun Med (Lond). 2025;5(1):62.
- Flaharty KA, Hu P, Hanchard SL, et al. Evaluating large language models on medical, lay-language, and self-reported descriptions of genetic conditions. Am J Hum Genet. 2024;111(9):1819-1833.
- Okunytė P. ChatGPT can read medical textbooks but likely will misdiagnose you. Cybernews. August 19, 2024. Accessed October 16, 2025.
- Henry TA. 2 in 3 physicians are using health AI—up 78% from 2023. AMA Digital Health. February 26, 2025. Accessed October 16, 2025.
Author
