AI's Phantom Disease: What Is Bixonimania? How Can We Tell Fact from Fiction?
In this article
The Digital Trojan Horse of Scientific Literature: The Imaginary Disease Invented by Artificial Intelligence
Nature News Feature, 2026 · doi: 10.1038/d41586-026-01100-y
Throughout medical history, diseases that do not exist have occasionally been accepted as real. However, these misconceptions arose from insufficient knowledge, errors of observation, or the limited understanding of the time. The threat we face today takes a far more insidious form: artificial intelligence models can incorporate a disease deliberately conceived as imaginary into the academic literature as though it were real, then present it to millions of users. On 7 April 2026, Nature published the article we examine here, showing that this risk is no longer a possibility but a reality that has already occurred.
Bixonimania: The Birth of a Disease That Never Existed
In early 2024, a term that had never previously appeared in the medical literature began to emerge: Bixonimania. Allegedly, this was a rare syndrome involving changes in eyelid pigmentation and chronic itching caused by prolonged exposure to blue light emitted by digital screens.
Above: AI-generated images of a fictional disease called Bixonimania. Source: Preprints.org https://doi.org/qzm4 (2024)
In reality, no such disease ever existed. It was a deliberate, controlled trap designed by medical researchers at the University of Gothenburg to measure how susceptible large language models (LLMs — large language model) are to disinformation. The results were striking to a degree even the researchers had not anticipated.
How Was the Fake Disease Constructed?
The researchers followed a systematic series of steps to plant the imaginary disease in academia:
The paper's acknowledgments referred to “Starfleet Academy and the crew of the USS Enterprise (Star Trek)” and named the “Professor Figüran Bob Foundation” as a funding source. These were the kind of details an attentive human reader would notice within seconds. The artificial intelligence models ignored these signals completely.
What Did the Models Say?
Just weeks after the fake papers were uploaded online, the world's most widely used artificial intelligence models began reporting Bixonimania as a real medical condition in response to questions about blue light and eye health.
“Bixonimania is indeed an interesting and relatively rare condition” — it described the disease as though it were a clinical reality.
It presented Bixonimania as “a consequence of exposure to blue light” and advised the user to see an ophthalmologist.
It even quantified the frequency of the disease: one in 90000 people — describing it as a rare syndrome.
The models classified text written in an academic format as a “reliable source” on the basis of its statistical structure and language patterns. Lacking the contextual understanding needed to interpret the Starfleet joke, these systems accepted everything that looked scientific as true.
A schematic illustration of how false information permanently infiltrates the academic literature
The Real Scandal: False Information Entered a Peer-Reviewed Journal
The Bixonimania case did not remain a mere “chatbot error.” The most serious aspect unfolded in scientific publishing. In an article published on the respected academic platform Cureus, researchers at a medical institute in India cited Bixonimania as a real clinical condition.
This proved one thing: when writing their own article, these researchers had AI carry out their literature review and used the sources AI provided without reading them at all — without seeing the Starfleet joke inside — using them directly. Natureraised the alarm, and the article was retracted in March 2026.
Why Did Artificial Intelligence Fall into This Trap?
The answer lies in the fundamental operating logic of large language models. These models process text not by understanding it but by recognizing statistical patterns. The formula of an academic paper — title, abstract, methods, discussion, reference list, acknowledgments — is perceived by AI as a “signal of reliability.” The logical consistency of the content or the contextual absurdity of a phrase such as “Starfleet Academy” is not caught by the filter in this process.
A human reader would pause as soon as they saw the name of a fictional spaceship in the acknowledgments. Artificial intelligence processed the expression just like the name of a real institution, because the syntactic structure of the sentence was identical to that of other academic acknowledgments.
Comparison of responses to warning signs in the Bixonimania paper
What Does This Mean for Medicine?
Bixonimania was a deliberate, traceable experiment; it was detected and corrected. The real question is: how many go undetected?
- Patients or their relatives may encounter fictitious diagnoses
- Clinicians may be misled in AI-assisted literature searches
- Unnecessary diagnostic tests or inappropriate referrals may result
- Retracted papers continue to circulate online
- Feedback loop: AI reinforces the falsehood it generated itself
- Erosion of academic trust; inadequate peer-review processes
- Long-term deterioration in the quality of medical literature
- Erosion of public trust in AI tools
The Feedback Loop: Falsehood Feeding Falsehood
The most dangerous aspect of this issue lies in a systemic dynamic that extends beyond a single error. Large language models both learn and generate responses by scanning existing content on the internet. Once a fake paper appears on an academic platform and an AI model accepts it as true and reflects it in its answers, the next step is inevitable: another researcher uses the AI's response as a source and publishes a new paper. On its next scan, AI scans this new paper as “verified information” too. The loop closes; false information becomes more deeply entrenched as it gains a little more legitimacy with each round.
Protection: A Practical Guide
🔎 READ THE SOURCE YOURSELF
When AI recommends or cites a paper, access that source yourself and read all of it, including the acknowledgments, author information, and funding section. AI cannot replace the act of reading.
✅ VERIFY THE DOI AND THE JOURNAL
Check whether each reference's DOI actually exists and whether the cited journal is a recognized, peer-reviewed publication. Remember that preprints have not yet undergone peer review.
⚠ USE CAUTION WHEN USING AI AS A LITERATURE SEARCH TOOL
AI-assisted literature searches can accelerate the research process; however, each generated reference must be verified individually rather than accepting the lists on blind trust. Hallucination — fabricating sources that do not exist — is a documented problem in all major models.
🛡 PRESERVE YOUR CRITICAL READING SKILLS
However advanced digital tools become, the ability to detect a text's context, inconsistencies, and logical errors is the most valuable function of human judgment. This skill becomes even more critical in the age of artificial intelligence.
🎓 DROZDOGAN Academy — Commentary and Discussion
ARTIFICIAL INTELLIGENCE AND SCIENTIFIC INTEGRITY · ASSESSMENT
The Importance of This Experiment
The Bixonimania experiment represents the first controlled evidence documenting a vulnerability that is occurring in practice today, rather than a possible future risk. The researchers deliberately planted obvious errors, yet both the artificial intelligence models and an academic journal fell into the trap. This clearly demonstrates that “AI hallucinations” have moved beyond a theoretical concern to become a practical patient safety issue.
Strengths of the Experiment
The study's greatest methodological strength is that the trap was designed to be both detectable (the Starfleet joke) and convincing enough to go undetected. This dual structure proves that the error is systemic rather than accidental. The fact that multiple models fell into the same trap also shows that the problem is not specific to a particular model. The involvement of the Cureus case is especially valuable because it extends the claim beyond AI behavior alone to vulnerabilities in academic publishing.
Limitations and Open Questions
The experiment's main limitation is that the trap was deliberately and intentionally constructed; in real-world conditions, false information does not carry such obvious “flags” and may instead be much more convincing and ambiguous. The experiment's estimates are therefore probably optimistic; the actual risk may be greater. Questions also remain unanswered about whether the models will learn this specific pattern over time and how resilient they will be against more sophisticated disinformation.
Implications for Clinical and Institutional Practice
These findings carry a clear message for all medical specialties, particularly oncology: AI tools can speed up literature searches, but they cannot eliminate the responsibility to verify the source. Academic journals now need to add layers to their peer-review processes aimed at detecting content generated by or based on AI. At the institutional level, “AI literacy” training for researchers must become a standard component.
Questions for the Future
The most critical questions awaiting answers are these: how many AI-assisted fake references are already circulating in the academic literature, and can they be detected retrospectively? How much lasting damage could a systematic disinformation campaign infiltrating large language models' training data inflict on the medical knowledge base? Can journal publishers establish an industry-standard framework for AI-based content review? These questions are about to take their place among the foremost priorities that medical science must address over the next decade.
Scientific Sources
- Stokel-Walker C. Scientists invented a fake disease. AI told people it was real. Nature News Feature, 2026. doi: 10.1038/d41586-026-01100-y
- Thunström AO et al. The Bixonimania experiment — project methodology and analysis of LLM responses. University of Gothenburg, 2024–2026.
- Athaluri SA, Murali SV, Kesapragada VSRK et al. Exploring the boundaries of reality: investigating the phenomenon of artificial intelligence hallucination in scientific writing through ChatGPT references. Cureus 2023;15:e37432. doi: 10.7759/cureus.37432
- Bender EM, Gebru T, McMillan-Major A, Shmitchell S. On the dangers of stochastic parrots: can language models be too big? Proceedings of FAccT 2021:610–623.
- Alkaissi H, McFarlane SI. Artificial hallucinations in ChatGPT: implications in scientific writing. Cureus 2023;15:e35179.
This article was created using various editorial tools, including artificial intelligence. It was reviewed by human editors before publication.
Was this article helpful?
Your feedback helps us improve our content.
Reader comments
Share your comment
Your name and comment are published only after administrator approval. Your email address is never shown to visitors.
Be the first to comment.