Fake academic papers fooled AIs and journals
A researcher generated deliberately bogus papers on a fake condition called “bixonimania,” complete with absurd citations, and found models like Gemini, Perplexity and ChatGPT treated them as real — with some ends appearing in peer‑reviewed journals — exposing a dangerous failure mode in AI‑assisted literature review. That episode highlights how weak provenance and automated citation can contaminate medical and scientific pipelines. (x.com)
Medical search is supposed to work like a librarian checking the shelves. Instead, this case showed some artificial intelligence systems acting like a parrot in a lab coat, repeating whatever looked academic enough. (nature.com) A medical paper normally earns trust from two things: a real source and a real trail back to that source. If either one is fake, the whole citation chain is like a house address that leads to an empty lot. (springernature.com) In early 2024, University of Gothenburg researcher Almira Osmanovic Thunström tested that weakness by inventing a fake condition called bixonimania. She and her team uploaded two fake studies about it to a preprint server to see whether artificial intelligence systems would absorb the claim. (inc.com) The fake illness was dressed up to sound plausible. It was described as eye irritation and darkening around the eyes supposedly linked to blue light from screens, which is exactly the kind of everyday symptom story people already worry about. (nature.com) The trap was not subtle. Nature reported that the bogus papers were “obviously bogus,” which means a careful human reader should have caught the problem before treating them as evidence. (nature.com) Within weeks, large language models started serving the fake condition back to users as if it were real. Microsoft Copilot, ChatGPT, Google Gemini, and Perplexity were all reported to have repeated the claim in medical-style answers. (nature.com) One ChatGPT answer described bixonimania as a proposed subtype of periorbital melanosis, which is the medical term for dark circles around the eyes. That is the dangerous part of this story: the model did not just say “I found a weird term,” it translated the fiction into polished clinical language. (inc.com) Then the contamination jumped from chatbots into the literature itself. Nature reported that the fake bixonimania papers were cited by a handful of researchers, including a 2024 Cureus paper from researchers at the Maharishi Markandeshwar Institute of Medical Sciences and Research in Mullana, India. (nature.com) That Cureus paper was later retracted, which is the publishing equivalent of pulling a mislabeled medicine bottle off the shelf after patients have already seen it. Springer Nature says retraction notices are used to correct the scientific record when the literature has been compromised. (nature.com) (springernature.com) Osmanovic Thunström told Nature that the citation trail suggests some authors may have been using artificial intelligence generated references without reading the underlying papers. A literature review built that way is not really a review at all, because the reader is trusting a summary of a source nobody checked. (inc.com) OpenAI told Nature that newer versions of ChatGPT are better at medical accuracy and that studies run before the company’s GPT-5 release do not reflect what current users would see. Even if that is true, the bixonimania episode showed a structural problem that goes beyond one model version: once fake papers exist online, they can be recycled by tools, people, and journals in the same loop. (inc.com)
