Health answers from an assistant come with a policy condition

Start here

People ask assistants about symptoms, medications, and test results because the question is quick to type and the answer arrives without an appointment. Vendors' own rules already say something specific about this use, and it is worth reading before typing the question rather than after.

What the documents say

OpenAI's usage policies, effective 29 October 2025 as retrieved, prohibit 'provision of tailored advice that requires a license, such as legal or medical advice, without appropriate involvement by a licensed professional,' and separately bar 'automation of high-stakes decisions in sensitive areas without human review,' naming medical decisions among them. Anthropic's usage policy is more specific about the mechanism: it classifies 'healthcare decisions, medical diagnosis, patient care, therapy, mental health, or other medical guidance' as a high-risk use case, and for advice reaching an individual directly, requires that 'a qualified professional in that field must review the content or decision prior to dissemination' and that the person be told AI helped produce it, while stating that general wellness guidance on sleep, stress, or exercise falls outside that category. Separately from policy, a 2023 study in JAMA Internal Medicine, Comparing Physician and Artificial Intelligence Chatbot Responses to Patient Questions, had licensed healthcare professionals evaluate 195 real question-and-answer pairs from a public forum, comparing physician replies to chatbot-generated ones; evaluators preferred the chatbot's response in 78.6% of cases and rated it more empathetic. The authors are explicit about a major limit: the study 'didn't independently evaluate factual accuracy or fabricated information' — it measured perceived quality and empathy on forum text, not clinical correctness.

Check this

Notice whether an answer sounds confident and complete versus whether it names an uncertainty or tells you to see a clinician; the second pattern is closer to what the policies above actually permit. Search the specific claim, drug name, or dosage the assistant gave against a source you already trust, since the JAMA study's own limitation means a well-liked answer was never confirmed to be an accurate one.

What holds and what fails

The vendor policies hold as a stated boundary regardless of how good a chatbot's tone tests: both companies build in a licensed-professional checkpoint for advice reaching an individual, which is a policy fact, not a technical guarantee that unreviewed answers are wrong. The JAMA finding is often summarised as AI beating doctors, which overstates what a preference-and-empathy rating on forum text can show; that gap between the finding and the popular summary is the thing worth double-checking, editorially, whenever the study is cited.

• Read the answer for hedges and referrals to a clinician, not just confidence.
• Check any specific drug, dose, or diagnosis claim against a trusted source.
• Treat a chatbot's health answer as a starting question for a professional, not a result.

Both vendors' own policies point the same direction: useful for framing a question, not licensed to replace the person who reviews the answer.