Evidence standards were built for a world where researchers read before they cited

April 2026 · Evidence standards

That world will end unless we do something about it.

In a recently published article in Nature, researchers invented a fake skin condition, uploaded papers about it to a preprint server, and watched AI chatbots repeat it back as real medicine. Within weeks Copilot, Gemini, Perplexity and ChatGPT were all describing it as legitimate. One peer-reviewed paper in Cureus then cited it.

The framing in most of the coverage is about AI hallucination and injection. I think the real finding is something else.

LLMs now bring science directly to consumers in a way that didn’t happen before. Science isn’t researcher-to-researcher only anymore — it’s researcher-to-LLM-to-public. Editorial and peer review standards were built for the first model. They need to reflect the second, and acknowledge that the stakes for errors are higher now as the public is more exposed.

Is this a story about AI, or about editorial standards that haven’t caught up?

Scientists invented a fake disease. AI told people it was real — Nature, 2026

Evidence decisions made during development are reviewed years later by HTA bodies and payers who were not consulted at the time.

For programmes where the architecture can still be shaped, or dossiers with unresolved payer risk, thirty minutes is usually enough to clarify fit and where I would have most leverage.

Schedule a Call