Patient-Facing Health Chatbots: Evidence and Risks

If you think this is an emergency

A chatbot is not the right tool for an emergency. In the US, call 911. If you are thinking about harming yourself, the 988 Suicide & Crisis Lifeline (US) can help. For a possible poisoning, US Poison Help is 1-800-222-1222. Elsewhere, see Poison Help by Country. This site does not offer a chatbot, and nothing on this page is personal medical advice.

What this page covers

A patient-facing health chatbot is a program, usually built on a large language model (LLM), that answers health questions in conversation. This page summarizes the evidence on how well these tools work when members of the public use them, and the risks. Evidence labels: promising but limited, mixed, insufficient evidence, evidence of harm, marketing claim. Regulatory notes are for the United States and are not legal advice.

The most relevant trial: strong on tests, weak with real people

A 2026 randomized study in Nature Medicine tested whether LLMs help members of the public work out what is wrong and what to do (see a doctor, self-care, or call an ambulance) in ten medical scenarios. It included 1,298 UK participants who were randomly assigned to one of three LLMs (GPT-4o, Llama 3, Command R+) or to a source of their choice. [1]

• When the LLMs were tested alone, they identified the relevant conditions in 94.9% of cases and the right course of action in 56.3% on average. [1]
• Participants using the same LLMs identified relevant conditions in fewer than 34.5% of cases and the right course of action in fewer than 44.2%, both no better than the control group. [1]
• The authors concluded that standard benchmarks and simulated patients did not predict these failures, and that human user testing is needed before public rollout. [1]

Limits: the scenarios were written scenarios, not real illnesses; the participants were UK adults recruited online; and the models were the versions tested in 2024. The article carries a Publisher Correction dated April 17, 2026. [2] Label: no good evidence of benefit to the public in this setting, which is not a verdict on every chatbot.

What other studies show

• Answer quality in text. A 2023 JAMA Internal Medicine study had licensed professionals compare physician replies on a public forum with ChatGPT replies to 195 questions. Evaluators preferred the chatbot replies in 78.6% of evaluations. [3] This was a cross-sectional study of forum posts, with chatbot replies much longer than the physicians' (211 versus 52 words), and it did not test whether anyone's health improved. [3] Label: promising but limited for drafting replies that a clinician reviews; it does not show safe unsupervised advice.
• Clinical decision-making. A 2024 Nature Medicine study built 2,400 real patient cases for four abdominal conditions and found that the LLMs tested did not diagnose accurately across all conditions, performed significantly worse than physicians, and did not reliably follow guidelines. [4]
• Bias. A 2023 study put nine questions about race-based medicine to four commercial LLMs, five times each; all models showed examples of repeating race-based medicine, and they were not always consistent when asked the same question again. [5]

Regulation and official statements

• The FDA's advisory committee summary (November 6, 2025) said that of more than 1,200 AI-enabled devices authorized at that time, none had been authorized for mental health uses. [6]
• The same summary notes the FDA's regulatory approach applies to generative AI products as well, and that many consumer wellness apps are not reviewed or authorized by the FDA. [6]
• We did not find, in the FDA pages we read, a listing of any general-purpose chatbot cleared to give medical advice to the public. The FDA's AI device list page says the agency will explore ways to identify devices that use large language models, so the status can change. [7] Not verified beyond that.
• The World Health Organization's guidance on large multimodal models (January 2024) says there are documented risks of false, inaccurate, biased or incomplete statements that could harm people making health decisions. [8]

Risks to understand

Who should be especially careful

Anyone with symptoms that are severe, sudden or getting worse; people who are pregnant; parents of infants and young children; people managing several medicines; and anyone in emotional crisis. Contact a clinician or emergency service instead of relying on a chatbot. This also applies to people who feel a chatbot is "more understanding" than a person: the empathy rating in forum studies is not evidence of safety.

Using a chatbot as a research aid

If you use one to understand a term or prepare questions for an appointment, treat answers as leads to check against a reliable source, such as the NHS or MedlinePlus, and bring them to a professional. Do not use a chatbot to start, stop or change a medicine. The page Clinical Validation and Real-World Performance of Health Software explains what "tested" should mean.