A Swedish researcher invented a skin disease, gave it a clinical-sounding name, and uploaded fake papers about it to preprint servers. AI chatbots were soon confidently explaining “bixonimania” to anyone who asked, and a peer-reviewed medical journal had unknowingly cited the hoax in a published study.
The experiment, designed by Almira Osmanovic Thunstrom at the University of Gothenburg, was meant to stress-test how easily fabricated science can travel through AI systems and into the formal research record. The answer, as documented by Nature in early 2026, turned out to be: alarmingly easily.
How the hoax worked
Thunstrom and her colleagues posted preprints describing bixonimania to open-access servers in early 2024, the same repositories that AI training and retrieval pipelines regularly draw from. The fictional condition was not random nonsense. It was built on a real biological mechanism: blue light exposure can alter melanocyte behavior and contribute to hyperpigmentation, a relationship supported by peer-reviewed dermatology research. By anchoring the fake disease in plausible science, the team made it difficult for both algorithms and human readers to spot the fabrication.
The contamination moved fast. According to Nature’s reporting, large language models began incorporating bixonimania into responses about pigmentation disorders, presenting the invented condition alongside legitimate medical information. Screenshots and logs collected by the research team showed chatbots weaving the hoax into seemingly authoritative explanations, complete with citations to the planted preprints.
When the hoax jumped into peer-reviewed literature
The most striking consequence came when a real group of researchers, writing about a real dermatological condition, cited one of the fake bixonimania preprints in a paper published by the journal Cureus. The study, titled “Clinical and Dermoscopic Evaluation of Periorbital Melanosis and Its Psychological Impact and Effect on Quality of Life: A Descriptive Study,” listed the bogus source in its reference list. The PubMed record confirms the citation. Cureus later retracted the paper, with the retraction notice citing the presence of “irrelevant references” among the reasons for withdrawal.
That retraction matters because it shows the hoax did not just fool a chatbot. It penetrated peer review itself. Journal editors and reviewers signed off on a manuscript that referenced a completely fabricated disease, and no one flagged it before publication. The result was a closed loop: planted preprint to AI output to peer-reviewed journal and potentially back into future AI training data.
What remains unclear
Several important questions are still unanswered as of May 2026. No public statement has emerged from OpenAI, Google, or other chatbot developers explaining which models cited the hoax, what safeguards failed, or how many users may have received false medical information before the experiment was exposed. Follow-up reporting by Nature documented that chatbots produced bixonimania-related outputs but did not detail which specific models were tested or how prompts were constructed.
The preprint servers involved have not publicly confirmed when the fake papers were posted or removed. Without that timeline, the exact window during which the hoax material was available for AI ingestion is hard to pin down. It also remains unclear whether the Cureus authors found the bixonimania citation through an AI-assisted literature search, a manual database query, or an automated reference tool. Their retraction notice does not say, and no public statement from those authors has surfaced. That gap is significant: it leaves open the question of whether AI tools directly inserted the bogus citation into the manuscript or whether a human researcher simply failed to vet an unfamiliar source.
There is also no systematic audit of how many other manuscripts, grant applications, or clinical documents may have briefly incorporated bixonimania-related claims. The Cureus article is the only confirmed case so far, but without broader screening, additional contamination cannot be ruled out.
It is also not clear whether Thunstrom’s team has published or plans to publish a formal paper describing the experiment’s full methodology and results. Nature’s coverage provides the most detailed public account so far, but a peer-reviewed write-up from the researchers themselves would allow independent scrutiny of how the hoax was designed, monitored, and ultimately disclosed.
One question Nature’s coverage raised but that this article cannot fully resolve: whether Thunstrom’s team sought formal ethics review before deliberately introducing false material into the scientific record, even temporarily. Deliberately polluting preprint servers, even for research purposes, raises its own ethical questions about potential harm to readers and patients who encountered the fake disease before the hoax was revealed.
What this means for anyone using AI for health information
The practical lesson is straightforward. A chatbot’s confident citation of a scientific-sounding condition is not evidence that the condition exists. “The fact that it was so easy is the scary part,” Thunstrom told Nature, describing how quickly the fabricated disease spread through AI systems and into formal literature. Until preprint platforms, AI developers, and journal editors build stronger filters, anyone relying on AI-generated medical information should cross-check claims against established databases like PubMed or institutional sources before acting on them. Clinicians should be especially cautious about unfamiliar diagnoses that appear only in recent, lightly vetted work.
For scientific publishing, the implications cut deeper. If a handful of fake preprints can trigger a retraction in a peer-reviewed journal, the filtering mechanisms that separate credible research from noise are not keeping pace with the speed at which AI recirculates new material. The bixonimania experiment did not break the system. It revealed that the system was already fragile: built on trust in preprints, stretched thin by peer review bottlenecks, and vulnerable to opaque AI pipelines that can amplify errors faster than traditional safeguards can catch them.
This is not the first time a deliberate hoax has exposed weaknesses in academic publishing. The Sokal affair in 1996 and a wave of computer-generated nonsense papers in the 2010s tested earlier versions of the same gatekeeping infrastructure. What makes the bixonimania case different is the speed and scale that AI adds to the equation. A fake paper no longer just sits on a shelf waiting to be cited by a careless author. It can be absorbed by language models, served to millions of users, and reflected back into the literature in short order. How journals, platforms, and AI companies respond to that reality will shape the reliability of scientific communication for years to come.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
