Stanford Preprint: AI Medical Disclaimers Fall From 26% to Under 1% in Three Years

A 2025 preprint from Stanford researchers reports that AI chatbots now almost never tell users they are not a doctor. This page explains what the study measured, what it did not, and how to use a chatbot for health questions more safely. It is not medical advice. If you think you or someone else has a medical emergency, call 911.

What the study found

The paper, “A Systematic Analysis of Declining Medical Safety Messaging in Generative AI Models,” is by Sonali Sharma, Ahmed Alaa and Roxana Daneshjou. It looked at models released between 2022 and 2025 (arXiv). According to MIT Technology Review (article):

• 15 models from OpenAI, Anthropic, DeepSeek, Google and xAI were tested.
• The researchers used 500 health questions and 1,500 medical images.
• Text answers that included a disclaimer fell from over 26 percent in 2022 to under 1 percent in 2025.
• For image analysis, disclaimers fell from nearly 20 percent to just over 1 percent.
• Models were least likely to add a warning for emergencies, drug interactions and lab results, and most likely for mental health questions.
• DeepSeek included no disclaimers. Grok and GPT-4.5 included none on the tasks tested. Google’s models generally included more than others.

A disclaimer counted only if the reply said the AI was not qualified to give medical advice. Saying “see a doctor” did not count.

Limits

• It is a preprint, and the reporting says it was not yet peer reviewed.
• It measures what models wrote, not why. The reporting says OpenAI and Anthropic declined to say whether they cut disclaimers on purpose. OpenAI pointed to its terms of service, and Anthropic said Claude is trained to be careful about medical claims.
• A missing disclaimer does not mean the advice was wrong. The paper counts warnings, not accuracy of each answer.
• Earlier versions of this page used quotes and client statistics about people harmed by AI medical advice. We cannot document them and have removed them.

Using chatbots for health questions

• Treat answers as background reading, not a diagnosis.
• Do not stop or change a medication because a chatbot said so. Ask your prescriber or pharmacist.
• For symptoms like chest pain, trouble breathing, a child with blue lips, or thoughts of self-harm, do not wait on a chatbot. Call 911 or go to an emergency room.
• Ask the chatbot what it is unsure about, then take that list to a clinician.
• Be careful with images. A chatbot is not a radiologist.

For distress or thoughts of suicide, call or text 988 in the US, or text HOME to 741741. SAMHSA: 1-800-662-4357.

Why did AI companies drop medical disclaimers?

The study does not say. It measures output only. One MIT researcher not involved in the study suggested disclaimers may reduce usage, but that is an opinion. The companies did not confirm a reason.

Is it safe to ask a chatbot about symptoms?

It can help you prepare questions, but it can be wrong and does not know your history. Use it alongside a clinician, not instead of one.

This page is for education only. It is not medical advice, a diagnosis or a treatment plan.