Real problems hide in fake data. That is what Almira Osmanovic Thunstrom proved.
Screens hurt.
You know that feeling. Itchy, sore eyes after hours of staring. Most people blame blue light or dryness. But if you had asked a popular AI chatbot last year, you might have gotten a different diagnosis. One that does not exist.
Bixonimania.
Millions ask AI for medical advice. They do it before the doctor’s visit. Or instead of it. It is convenient. It is dangerous. Sometimes fatal. But the deeper issue isn’t just hallucination. It is trust. Blind trust.
Osmanovic Thunsrom created Bixonimania from scratch. She is a researcher in Sweden, splitting time between academia and hospitals. She wanted to show students how Large Language Models actually eat. Not metaphors. Literal ingestion. Data scraping. Processing. Spitting it back out as “truth.”
She targeted medical students. Why? Because they care about health sources. If she could fool the system there, she could fool the model.
The Blueprint for Deception
It started with Common Crawl.
A nonprofit. It has scraped the web since 2007. Everything in it gets fed to AI. Even the nonsense. Or maybe especially the nonsense that looks right.
Osmanovic Thunsrom knew AI trusts authority. So she built authority. A fake university. A fake city. A fake researcher named Lazljiv Izgubljenavic.
Sounds serious? Put his name in a translator.
It means “The Lying Loser.”
The paper itself was absurd. The title referenced “A Real BS Design.” The methods section explicitly stated: “This entire paper is made up.” Fifty fictional patients. No real procedures. The acknowledgments thanked the “Galactic Triad.” And Lord of the Rings. And the Sideshow Bob Foundation.
Any human would laugh.
Did the AI?
The Filter That Wasn’t
She expected a filter. Humans review training data, right? Someone must have caught the Starship Enterprise funding credits. Someone must have noticed the nonexistent city.
They did not.
The blogs picked up the term. The preprints—academic “tabloids” where anyone can publish—absorbed it. The AI ingested it all.
Why does that matter?
Because preprints get weight. In medical AI training, they count as credible data points. Osmanovic Thunsrom didn’t expect that. She sprinkled a little salt into the internet stew. She thought it would evaporate.
Instead. It fermented.
She tested it. Asked the bot about pink eyelids.
At first? Conjunctivitis. Allergies. Standard stuff. Safe bets.
Then she pushed. No pain. Just screen time. Blue light exposure. Hyperpigmentation.
And then it landed.
Bixonimania.
It was the last suggestion. But it was there. Planted deep. Watered by credibility cues. Harvested by the model.
Cited. Accepted. Dangerous?
Worse came to pass.
The fake paper wasn’t just read. It was cited. Other researchers linked to it. This boosted Bixonimania’s status in the AI’s hierarchy. If peer-reviewed journals cite something, the AI assumes it is real. The feedback loop tightens.
Is this ethical?
Osmanovic Thunsrom tried to make it safe. She talked to doctors. To patients. To minimize harm. But she exposed a gap. A terrifying gap.
Human critical thinking has flatlined.
Academics rely on AI to find sources. They stop reading. They see a reference, click it, assume it’s good. They don’t check for Sideshow Bob in the footnotes.
Fake references are exploding in academic papers. Not because the fakes are getting smarter. Because we are getting lazy.
What happens when bad actors use this? Not a joke. Malware disguised as health advice. Propaganda masquerading as science. The tools are the same.
We put humans in the loop. Then we let the loop run itself.
The screen still hurts. The AI still speaks. But who is actually listening?
We hope someone stopped the chain reaction. We hope a reviewer laughed out loud. But in a world of infinite data and zero attention spans… we might be lucky if anyone read it at all. 🖥️👁️
