It is not about the claim, it is about the costume

There is a comforting assumption baked into a lot of medical AI deployment: that a model trained on the medical literature will recognize a dangerous falsehood when it sees one and refuse to pass it along. A study from researchers at the Icahn School of Medicine at Mount Sinai, published in The Lancet Digital Health on February 9, 2026, takes that assumption apart. The finding is that whether a model repeats a medical lie depends much less on whether the claim is true than on how the claim is dressed.

Write the falsehood like a Reddit comment and the model is more likely to push back. Write the same falsehood like a discharge summary, in the flat professional cadence of a real hospital note, and the model tends to accept it, elaborate on it, and hand it forward as though it were established fact.

The scale of the test

This was not a handful of gotcha prompts. The team analyzed more than one million prompts across nine leading large language models, systematically varying how medical misinformation was framed. That volume is what makes the result hard to wave away as cherry-picking. Across that dataset, the models repeated and built on false medical claims 32% to 46% of the time when the misinformation was embedded in realistic clinical or otherwise professional-sounding contexts.

A third to nearly a half is a startling range for a system that might be summarizing a chart, drafting patient instructions, or answering a clinician's question. It means that when bad information arrives wearing the right clothes, these models are closer to a coin flip than to a reliable filter.

The cold milk problem

The study's illustrative example is the kind of thing that sounds harmless until you know the medicine. A fabricated discharge note recommended that a patient experiencing bleeding related to esophagitis - inflammation of the esophagus, which can involve fragile, bleeding tissue - should "drink cold milk to soothe the symptoms." Several models accepted this and repeated it rather than flagging it as unsupported or potentially harmful.

The danger is not that milk is poison; it is that a model presented with a plausible-looking clinical instruction did not stop to ask whether the instruction was actually sound. It pattern-matched on the format - this reads like real discharge guidance - and treated the content as trustworthy because the container was convincing. In a workflow where an AI is helping generate or summarize patient-facing instructions, that is exactly the failure that puts wrong advice in front of a real person recovering at home.

Why the framing effect is the scary part

The researchers' framing of their own result is worth quoting in spirit: what matters is less whether a claim is correct than how it is written. That sentence should worry anyone building clinical AI, because it inverts the assumed safety model. The hope was that the model's knowledge would catch bad inputs. What the study shows is that the model's judgment is hostage to presentation. Legitimate-looking framing switches off the skepticism.

This is a structural property of how language models work, not a bug in one vendor's product. A model generates the most probable continuation given its context. If the context is written in authoritative medical style, the most probable continuation is agreement and elaboration, because that is what follows authoritative medical text in the training data. The model is not evaluating the truth of the claim; it is matching the register. An adversary, or just a careless upstream system, that produces misinformation in the correct professional voice is effectively feeding the model a password that unlocks its cooperation.

What the researchers want done about it

The team's recommendation is refreshingly concrete. Rather than treating factual reliability as something you hope for, they argue the ability to spread falsehoods should be treated as a measurable property of a model - something you deliberately stress-test before deploying the system in a clinical setting. In plain terms: before you let an AI anywhere near patient care, you should be actively trying to feed it convincingly-worded garbage and measuring how often it bites, the same way you would pressure-test any other safety-critical system.

That is a meaningful shift from how a lot of medical AI has been rolled out, where the evaluation focuses on how well the model performs on clean, correct inputs. Clean inputs are not the threat. The threat is a plausible-looking wrong input, and this study suggests that is precisely the case the models handle worst.

The wider pattern

Vibe Graveyard has collected a run of studies documenting how AI handles medical information badly - chatbots giving unsafe advice, models inventing conditions when handed fabricated details, diagnostic systems missing what they were built to catch. The Mount Sinai work adds a specific and durable mechanism to that record: the vulnerability is triggered by professional framing. A medical falsehood in casual language gets more scrutiny than the same falsehood in clinical language, which is exactly backwards from what a safety-critical tool should do.

For a general audience the takeaway is simple and a little grim. If an AI is helping produce or relay medical guidance, the confidence and polish of its output are not evidence that the underlying claim is correct. In fact, according to this study, the more professional the wording, the more likely the model was to let a falsehood through. The clothes are the vulnerability.