OpenAI Mobilizes Doctors to Oversee ChatGPT Health

OpenAI relies on a global network of clinicians to refine how ChatGPT responds to medical questions, as the company claims new technical milestones. The organization emphasizes that its chatbot is not intended for diagnosis and continues its work despite a recent lawsuit and an underdeveloped legal framework.

Technical Milestones and Evidence Still to be Solidified

OpenAI has facilitated the connection of Apple Health and certain medical records to ChatGPT. Karan Singhal presented the GPT-6 Astra model as being at the forefront of a health benchmark test, and the company published a benchmark last month to evaluate the model's responses in mental health. However, Rebecca Soskin Hicks views these advancements as initial steps in building a body of evidence that still needs to be developed. She hopes that future studies will demonstrate concrete benefits for individuals and populations, while emphasizing that the peak has not yet been reached. Additionally, in July, the network of physicians had already reviewed over 700,000 responses from ChatGPT.

A Regulated Use: No Diagnosis or Treatment

Ashley Alexander, head of health products at OpenAI, asserts that ChatGPT should not be used for diagnosis or treatment, a restriction also stated in the company's terms of service. If prompted, the chatbot offers to help assess illnesses and injuries, but Alexander reminds users that the AI supports research that many are already conducting online. According to her, making the chatbot the centerpiece of a medical decision oversimplifies the process and could hinder a beneficial large-scale impact. OpenAI continues these efforts despite a still untested legal environment surrounding automated medical advice.

A Global Clinical Network Testing Gray Areas

OpenAI has recruited hundreds of doctors, speaking 49 languages, to review conversation examples and identify confusions or risks. Rebecca Soskin Hicks desires diversity in regions and specialties, and these practitioners work as part-time contractors from countries such as Kenya, Nepal, and Brazil. They do not directly provide training data. The claimed approach involves testing ambiguous, messy, and noisy situations to reflect real requests that are often poorly specified.

Improvement Loop: Identified Gaps and Educational Priorities

After their evaluations, the doctors relay to researchers the areas where the models struggle. The research team then seeks new data and training methods to better sort emergencies, ask the right questions when information is lacking, express uncertainty, and calibrate the level of detail in responses. Improving health advice mobilizes researchers and clinicians, in a team formed by Rebecca Soskin Hicks with Karan Singhal, head of health research. The latter expressed a desire for patients to perceive ChatGPT as a protector throughout their care journey.

A Highly Used Service, Balancing Ambition and Legal Risks

Every week, over 300 million people consult ChatGPT for health-related questions. OpenAI relies on a network of physicians led by Rebecca Soskin Hicks, who joined in 2024, acting as the interface between clinicians and internal teams to address the gaps in the models. She describes the field as high-stakes and envisions continuous improvement. In July, a pastor filed a lawsuit, claiming that GPT-4o misdiagnosed and downplayed his symptoms, seeking the suspension of health products. Ashley Alexander, head of health products, supports this positioning on the product side.