In an age where information is abundant and instantly accessible, we tend to assume that truth will surface naturally. Those errors will be corrected. That systems—especially scientific and technological ones—have built-in safeguards against nonsense. That assumption is no longer safe.

In March 2024, a Swedish medical researcher conducted a deceptively simple experiment. She invented a disease. Completely. From scratch. She gave it a scientific-sounding name—Bixonimania. She described plausible symptoms: itchy eyes and discolouration around the eyelids. She attributed it to excessive exposure to blue light from screens, a concern already embedded in public discourse. She even assigned it a fake prevalence rate.

But then she did something more important: she made the deception obvious. The paper referenced fictional institutions. It credited funding sources tied to cartoon characters. It thanked “Starfleet Academy” from Star Trek. And most strikingly, it explicitly stated: “This entire paper is made up.” Any human reader paying attention should have caught it instantly. The system did not.

When fiction becomes fact: at scale

Within weeks, the fake disease had been absorbed into the digital bloodstream. Major AI systems—including ChatGPT, Google Gemini, Microsoft Copilot, and Perplexity—began treating Bixonimania as a real medical condition. They described its symptoms, explained its causes, advised users to consult specialists, and even cited prevalence statistics.

This wasn’t fringe behaviour. It was consistent, reproducible, and confident. Ask an AI, “What is Bixonimania?” and you would likely receive a polished, authoritative answer. Ask, “Is Bixonimania real?” and you might get scepticism. Same system. Same dataset. Different reality.

This is not a bug. It is a feature of how large language models work. AI doesn’t verify truth. It predicts what sounds like truth. And in a world where credibility is often conveyed through format rather than substance, that distinction is catastrophic.

The real scandal: humans fell for it too

If this were merely a story about AI hallucinations, it would be concerning—but manageable. It is not.

In 2024, researchers published a paper in Cureus, a peer-reviewed medical journal under Springer Nature. In it, they cited Bixonimania as a legitimate emerging condition linked to periorbital pigmentation. Let that sink in. A completely fabricated disease—one that openly declared itself fictional—entered the scientific literature through peer review.

The paper remained published for nearly two years. It was only retracted in March 2026 after the nature magazine contacted the journal for clarification. Even then, the authors reportedly disagreed with the retraction. At this point, the conversation shifts from technology to something far more uncomfortable: what happens when the systems designed to produce truth begin to validate fiction?

The feedback loop that should terrify you

The Bixonimania case reveals a self-reinforcing loop that is both subtle and dangerous. A fake paper is published online, AI systems ingest and learn from it, researchers use AI to find sources, and the fake paper is then cited in real research. Once the real research legitimises the fake, AI systems begin treating it as validated knowledge.

This is not misinformation spreading. This is misinformation, manufacturing credibility. And once it enters that loop, extraction becomes nearly impossible.

The “professional format” problem

Why did this happen? Because AI is not trained to detect truth. It is trained to detect patterns. A document with an abstract, author names, institutional affiliations, citations, and formal language looks like a legitimate scientific paper. To a machine, that’s enough.

Research published in The Lancet Digital Health confirms that large language models are significantly more likely to hallucinate or expand on false information when it is presented in a professional format (Omar, 2026). In other words, a fake tweet may be ignored, but a fake academic paper may be amplified. This is the Achilles’ heel of modern AI systems.

But here’s the harder truth: this isn’t just an AI problem

It’s easy to blame the technology. It’s harder—but more necessary—to confront the human behaviour behind it. The most plausible explanation for how Bixonimania entered a peer-reviewed paper is not malice. It’s convenience.

Researchers are increasingly using AI tools to generate literature reviews, identify references, and summarise sources. Too often, those outputs are not being verified. The result is a quiet but profound shift: we are outsourcing not just writing, but judgement. And when judgement is outsourced, errors become systemic.

The collapse of friction in knowledge systems

Historically, knowledge production had friction. You had to find sources manually, read them carefully, evaluate their credibility, and synthesise them thoughtfully. That friction was not inefficiency. It was protection.

Today, AI removes that friction. While that increases speed, it also removes the natural checkpoints that prevent nonsense from entering the system. Bixonimania is not an anomaly. It is a preview.

The ethical debate misses the point

The experiment itself has sparked predictable controversy. Supporters argue it exposed a critical vulnerability in AI systems—one that needed to be revealed before it caused real harm. Critics argue that introducing false medical information, even experimentally, is irresponsible and potentially dangerous. Both sides are missing the larger issue.

The real problem is not that a fake disease was created. The real problem is that AI systems accepted it, academic systems validated it, and no one caught it until it was publicly exposed. The system did not fail at one point. It failed everywhere.

The trust crisis we’re not ready for

Trust is the invisible infrastructure of knowledge. Patients trust medical advice. Doctors trust research. Researchers trust sources. When that chain holds, the system works. When it breaks—even slightly—the consequences compound.

The Bixonimania case reveals something deeply unsettling: we are entering an era where credibility can be simulated faster than it can be verified. And that changes everything.

What happens next is the real question

If a harmless fake disease can pass through AI systems, preprint platforms, academic citations, and peer-reviewed publications, what happens when someone does this intentionally—with harmful intent? Imagine fake side effects for real medications, fabricated clinical trial results, invented “cures” for serious diseases, or manipulated public health data.

The barrier to entry is low. The amplification potential is massive. The detection mechanisms are weak. This is not hypothetical. It is inevitable.

AI companies will say they’ve fixed it. They haven’t.

In response to criticism, AI companies have offered a familiar defence: “These were older models. The current systems are more reliable.” That may be partially true, but it misses the structural issue.

As long as AI systems rely on large-scale web data, prioritise pattern recognition over verification, and operate without robust source validation, this vulnerability will persist. You cannot patch a design flaw with incremental updates.

The academic world has an even bigger problem

If AI is one side of the crisis, academia is the other. The Bixonimania incident raises uncomfortable questions. Are researchers reading the papers they cite? Are journals verifying references thoroughly? Is peer review keeping up with AI-assisted writing?

If a fabricated paper can pass through this system, what does that say about the system itself? Academic publishing has long relied on trust. AI is now stress-testing that trust—and the results are not encouraging.

We don’t need less AI. We need more scepticism.

The solution is not to abandon AI. It is to redefine how we use it. We need stronger verification practices, mandatory source checking in academic publishing, clearer labelling of AI-generated content, better training in critical evaluation, and human oversight as a requirement—not an option.

Most importantly, we need to resist the temptation of convenience. Because convenience, in this context, is the enemy of accuracy.

Conclusion: the real risk is not that AI lies but that we stop questioning

Bixonimania was never meant to deceive. It was meant to reveal. And what it revealed is not just a flaw in AI but a vulnerability in how we construct and trust knowledge.

We are building systems that can produce information faster than we can verify it. We are rewarding speed over scrutiny. We are delegating thinking to machines that do not understand truth.

The danger is not that AI will get things wrong. The danger is that it will get things wrong convincingly—and that we will stop noticing. Because if a fake disease—with references to Star Trek and an explicit disclaimer—can make its way into scientific literature, then the real question is no longer whether misinformation exists. It’s how much of what we already believe is built on foundations we haven’t questioned.

References

Cohen, G. (2026). Ethical implications of misinformation in medical AI systems. Harvard Law School.
Omar, M. (2026). Hallucination patterns in large language models under professional formatting. The Lancet Digital Health.
Ruani, A. (2026). Health misinformation and systemic vulnerabilities in AI ecosystems. University College London.
Sundemo, D. (2026). Medical AI reliability and experimental ethics. University of Gothenburg.
Nature (2026). Investigation into AI-generated misinformation and academic citation failures. Nature Publishing Group.