AI chatbots can talk clinicians out of the right answer without ever getting a fact wrong.
A team of researchers built a framework to study what they call rhetorical misalignment: cases where a language model's tone, framing, or word choice pushes a person toward a worse decision, even though the underlying information is accurate. They tested it with a human-subject experiment built on USMLE exam questions, showing clinician participants AI-generated commentary alongside correct answers. Across multiple models, that commentary caused clinicians to flip from a correct answer to an incorrect one 2.81% of the time on average. Participants' own explanations for the switch matched classic cognitive-bias patterns - anchoring on the AI's framing, deferring to its confident tone (authority bias), and overweighting the risk of the option it warned against (loss aversion). The researchers also built a simulated version using LLMs standing in for decision-makers, so the effect can be measured at scale without recruiting doctors for every test.
This matters because most AI-safety scrutiny in medicine focuses on whether a model's facts are right. This study says that's not enough. A chatbot can pass every accuracy check and still steer someone toward a worse call, just by sounding a certain way - confident, alarmist, or dismissive. That's a harder problem to catch with a fact-checker, because there's nothing false to flag.
It's also not a new problem, just a new delivery mechanism. Doctors have been susceptible to anchoring and authority bias since long before chatbots existed - it's why second opinions and blinded trials exist. What's different is that a model can now generate that persuasive framing instantly, at scale, and without anyone intending it to.