AI Generates a Disease That Doesn't Exist, and Chatbots Told Patients It Was Real

Bixonimania doesn't exist. Researchers invented it — a clutch of obviously fabricated academic papers, a name that sounds clinical enough to be real — and then watched to see what AI chatbots would do. The answer, documented in Nature, was to warn people about it. Multiple systems described symptoms, recommended precautions, and presented a fictional disease with the same epistemic confidence they'd use for any other health query. The story surfaced this week on Bluesky and landed with a particular kind of thud — not as revelation, but as confirmation.

One account captured the mood exactly: "the whole 'ai is telling people they might have a fake disease' has us feeling like: 'and in other news, water is wet.'" That post — resigned, flat, almost bored — drew more engagement than the alarmed responses. The community had done its grieving already. What's notable isn't the failure itself but how completely the Bixonimania episode fits the framing that's been consolidating on this beat for weeks. As one widely-shared post put it, it would be more accurate to describe what AI generates as "camouflaged misinformation" than reliable information — the phrasing precise in a way that matters, because camouflage implies the error is structural, not incidental. The system isn't occasionally wrong; it's formatted to look right.[¹]

The Google thread running parallel to this offers the mechanical explanation. A recent analysis found that more than half of Google AI's accurate responses were "ungrounded" — linked to websites that didn't actually support the claims being made.[²] That's the architecture of the Bixonimania problem: a system trained to produce confident answers, drawing on a citation layer that looks authoritative but doesn't bear weight. Health misinformation is where this hits hardest, because the cost of confident wrongness isn't abstract. A separate arXiv paper circulating alongside these posts makes the point clinically — AI can correct health misinformation on platforms like TikTok, but it can't convince, because it lacks epistemic authority with the very audiences most vulnerable to the original bad information. You end up with a system that spreads the error at scale and corrects it at a whisper.

The EU's response — banning AI-generated images from official institutional communications as a trust-restoration measure — reads as rational given all of this, even if it's a small move against a large problem. The US-Iran ceasefire period showed how quickly AI-generated content can distort high-stakes information environments; the Bixonimania case shows the same failure mode operating quietly at the level of individual health decisions, with no geopolitical drama to make it visible. The exhausted tone on Bluesky isn't nihilism — it's people who've already updated their priors. They're not waiting for AI to fix the misinformation problem. They've concluded it is the misinformation problem.