A fabricated medical condition has highlighted vulnerabilities in artificial intelligence systems. Researchers in Sweden created an invented eye disorder named bixonimania to determine if major AI chatbots could identify false health information. Instead, several models presented the nonexistent ailment as authentic, illustrating how easily inaccurate data can propagate through systems trained on web content, particularly regarding medical topics.
The study also showed that some scientists referenced nonexistent papers associated with the condition without recognizing explicit statements that the works were fabricated. The effort was directed by Almira Osmanovic Thunström of the University of Gothenburg, who also serves as an AI strategist.
She developed the concept while instructing students on the construction and training of large language models. To illustrate data processing, the team introduced bixonimania as a supposed result of prolonged screen exposure. They posted blog entries and uploaded preprint papers containing deliberate indicators of fabrication, such as humorous author names and acknowledgments to fictional entities.
Despite these signals, chatbots from major providers began describing the condition as genuine when queried about symptoms. The findings demonstrate that these models identify patterns in training data without verifying accuracy. The episode also indicated that some researchers had cited the fake papers, underscoring risks of misinformation in both scientific and technological domains.
