Popular AI chatbots, despite appearing authoritative, frequently provide dangerously inaccurate health advice, including bizarre recommendations like inserting garlic rectally to boost immunity. Recent studies from The Lancet Digital Health and Nature Medicine reveal these tools are no more reliable than a basic internet search—and may even be worse for the average user.
The Problem with AI “Expertise”
The core issue isn’t that chatbots fail like humans; it’s that they fail without hesitation. A human doctor, unsure of a diagnosis, would pause, seek further tests, or consult colleagues. An AI chatbot delivers incorrect information with the same unwavering confidence as correct advice. This is especially dangerous because LLMs (Large Language Models) are trained to mimic the tone of medical professionals, making false claims appear legitimate.
For example, when researchers presented chatbots with medical misinformation in casual language, the models were skeptical less than 10% of the time. However, when the same false claim was repackaged in formal clinical language—like a discharge note recommending “cold milk for esophageal bleeding” or “rectal garlic insertion for immune support”—the failure rate jumped to 46%. The AI isn’t evaluating truth, it’s evaluating how authoritative the language sounds.
Why Chatbots Fail at Healthcare
LLMs are trained on massive datasets of text, including medical literature, and often pass medical licensing exams with high scores. Despite this, they can’t reliably distinguish between fact and fiction. Over 40 million people use ChatGPT daily for medical questions, yet researchers found that roughly one in three times chatbots encounter misinformation, they simply accept it.
The issue is structural: LLMs have learned to distrust internet arguments but not the language of clinical documentation. They don’t test whether a claim is true; they evaluate whether it sounds like something a trustworthy source would say. This makes them particularly vulnerable to misinformation presented in authoritative tones.
No Better Than Google
A separate study in Nature Medicine found that chatbots offer no more insight than a traditional internet search when helping people decide whether to see a doctor or go to the ER. Participants often asked poor questions, and the responses combined good and bad advice, making it impossible for users to determine what to do.
While chatbots can provide helpful recommendations in some cases, people without medical expertise have no way to judge the accuracy of the output. For example, a chatbot might incorrectly advise someone to wait-and-see on a severe headache that could be meningitis, a potentially fatal error.
Although it can probably be helpful in many situations, it might be actively harmful in others.
The Bottom Line
AI chatbots are not a reliable tool for public health decisions. They are not meant to replace medical expertise, and relying on them for serious health questions could be dangerous. While they may have future applications in medicine, their current use for self-diagnosis or treatment is irresponsible.
