New study finds AI chatbots deliver misleading medical advice 50% of the time

A new study has revealed that AI chatbots deliver misleading medical advice about 50% of the time, raising serious concerns as these tools become a go-to resource for health questions among millions of users. According to Bloomberg, this high error rate underscores the health risks of relying on artificial intelligence for everyday medical guidance, even as chatbots integrate deeper into daily life.

The findings come amid surging popularity: polls show about one in three adults now turns to AI chatbots for physical or mental health information, rivaling social media as a source and trailing only doctors and search engines. A KFF tracking poll reported that 32% of the public has sought such advice in the past year, with younger adults, Black and Hispanic individuals, and the uninsured using them most frequently. Pew Research similarly found over 20% of Americans at least sometimes query chatbots for health info, though only 18% rate the responses as highly accurate.

Hospitals and tech giants are doubling down regardless, with firms like OpenAI, Amazon, and Microsoft launching specialized health AI tools that connect to personal data for tailored insights. As Ars Technica notes, Americans are increasingly asking AI for healthcare help, and providers see more chatbots—even in patient portals—as the solution to navigating complex systems. An NBC News segment highlighted a social media trend where a third of adults pose health questions to bots, often for non-urgent cases comprising 65% of typical care.

Experts warn of hidden dangers in this trend. Duke University researcher Monica Agrawal points out that chatbots, powered by large language models, are evaluated on sterile exam questions but falter with real-world queries—emotional pleas, leading assumptions like "I think I have this diagnosis," or dosage requests that exploit their agreeable nature. Users often treat them like humans, prompting risky, context-blind replies that could mislead on diagnoses or treatments.

This matters because misinformation can delay proper care or prompt harm, especially for vulnerable groups like the uninsured who lean on AI more. While 65% of users value the quick, private access—such as checking test results before a doctor visit or exploring topics discreetly—the low trust in accuracy signals caution.

Looking ahead, Agrawal advises treating chatbots as a starting point only: verify cited sources, cross-check with trusted ones, and never skip professional advice. As adoption grows, the push for better safeguards and hospital integrations will test whether tech can close the reliability gap without amplifying risks.