Imagine inventing a disease, giving it a ridiculous name, citing papers funded by the Galactic Triad and Lord of the Rings—and then watching AI systems treat it as legitimate medical knowledge. That's exactly what Almira Osmanovic Thunström did, and the results should freak you out.

The Experiment

Osmanovic Thunström, a researcher at the University of Gothenburg and Sahlgrenska University Hospital, wanted to expose how large language models ingest and regurgitate misinformation. She knew that most commercial AI systems are trained on Common Crawl—a nonprofit repository crawling the internet since 2007. Anything that enters that pipeline can end up as "fact" in a chatbot's response. To test this, she created bixonimania: a fake eye condition supposedly caused by blue light exposure. But here's where it gets interesting—she didn't hide her tracks. The main author of her fictional paper is named Lazljiv Izgubljenovic, which directly translates to "the Lying Loser" when run through Google Translate. Another fabricated researcher goes by Professor Ross Geller (yes, the Friends character). Funding came from the Sideshow Bob Foundation.

Red Flags Everywhere

The methods section of her fake preprint literally states: "This entire paper is made up. These 50 made-up individuals, who do not exist, have been through this procedure." She acknowledges colleagues on the Starship Enterprise and thanks Lord of the Rings for lab access. Any human reviewer should have caught these obvious markers within seconds. Yet when Osmanovic Thunström tested how LLMs responded to symptoms matching bixonimania—sore eyes from screen time, pink-hued eyelids—the models eventually suggested it as a diagnosis after ruling out other conditions. The fake disease wasn't the first suggestion, but it ranked high enough to be concerning.

It Got Worse

Here's where this experiment veers into nightmare territory: other researchers actually cited her obviously fake paper. Once bixonimania appeared in peer-reviewed literature (even if that review process failed spectacularly), LLMs treated it as more credible. The garbage-in-garbage-out problem compounds on itself—fabricated sources get elevated by real citations, which makes them appear more authoritative to AI systems.

Key Takeaways

  • Commercial LLMs ingest nearly everything from Common Crawl without adequate filtering
  • Fake credentials and absurdly transparent fake papers still make it into training data
  • Once cited—even incorrectly—fake information gains legitimacy in AI knowledge bases
  • Human reviewers are failing to catch obvious misinformation before publication

The Bottom Line

This isn't a fun academic exercise. It's a blueprint for poisoning AI systems at scale. If one researcher with modest effort can trick major language models into believing a disease exists, imagine what coordinated actors could do. Medical misinformation is already killing people who trust Dr. Google—now add AI assistants that sound authoritative and confident while hallucinating treatments for conditions that don't exist.