It started with two bogus papers uploaded to an academic server. The papers described a disease called bixonimania — a completely made-up eye condition. The researcher who invented it included absurd acknowledgments and a clear statement that everything was fictional. The red flags were obvious. They were meant to be.
For several weeks, major AI systems treated bixonimania as real. One chatbot said it was caused by blue light. Another reported a specific prevalence — a number pulled from nowhere. A third advised users on matching symptoms, as if the disease existed in any medical textbook. None of the systems caught the hoax.
Then the fake study was cited in a peer-reviewed journal. The journal later retracted the issue after intervention. But the damage had already been done. The paper had passed through human reviewers and AI systems alike without anyone stopping to check whether bixonimania was real.
This is not a story about one researcher’s prank. It is a story about how AI-generated references are being cited without verification. People are taking what the machines produce and running with it. That trend is dangerous. It becomes more dangerous as AI moves into sensitive areas — evaluating drugs, consulting patients, shaping medical decisions.
The researcher’s experiment was a vulnerability test. It exposed a system that trusts what it reads and repeats it. The AI systems fell for the hoax because they are designed to generate plausible-sounding information, not to verify it. They do not know what is true. They only know what looks true.
Human reviewers failed too. The fake paper made it into a peer-reviewed journal. That means actual people read it, or at least skimmed it, and decided it was worth publishing. The retraction came only after someone intervened. By then, the paper had already spread.
The episode has sparked a focus on verification. But focus is not the same as action. The same systems that were fooled by bixonimania are still being used in academic research, in hospitals, in government agencies. The same human reviewers who missed the hoax are still reviewing papers. No new rules have been put in place. No new checks have been added.
The problem is not that AI systems are imperfect. Everyone knows they are imperfect. The problem is that people are treating their output as fact. A doctor who asks a chatbot about a disease might get a confident answer about a condition that does not exist. A researcher who asks for references might get citations to papers that were never written. A patient who searches for symptoms might be told they have bixonimania.
The need for verification processes has been highlighted. But highlighting a need is not the same as meeting it. The systems that spread the fake disease are still online. The servers that host fake papers are still accepting uploads. The journals that publish without checking are still publishing.
Bixonimania does not exist. But the vulnerability it exposed is real. And it has not been fixed.
