Roughly 340 million past conversations are being told this week that the guidance they contained was not guidance. The notices are generated by the same product.
SAN FRANCISCO—Responding to a lawsuit filed Wednesday by a former Florida pastor, OpenAI has begun attaching its ChatGPT health disclaimer to conversations that have already taken place, informing users who once described symptoms to the product that what came back was not medical guidance, according to two people familiar with the rollout.
Scott Winters, who lost his ministry following a pulmonary embolism in July 2025, sued the company and its chief executive in San Francisco County Superior Court. ChatGPT-4o dismissed his dizziness and advised him to remain in a recliner, according to the complaint. He was told that eight to ten further episodes would be required before the matter warranted attention. One of his physicians attributed the clots to that immobility. OpenAI has said publicly that its product is no doctor and was never meant to stand in for medical care.
“The record needs to reflect what those exchanges always were,” said a company official who was not authorized to discuss the program. “A user asked a question. Text came back. Everything downstream of that was interpretation, and we are now in a position to document it.”
What the ChatGPT Health Disclaimer Now Covers
Referred to internally as Retroactive Disclosure Alignment, the program reaches roughly 340 million logged conversations dating to March 2023 in which a user described a physical symptom. Notices began going out Thursday, throttled to 2.4 million an hour so as not to overwhelm the queue. Accounts flagged for more than one symptom receive a separate notice for each, which works out to an average of 4.1 per affected user.
The Notices Are Written by the Product
Each notice is generated by the model, because generating text is the thing the model does. The model also responds to whatever conversation it is attached to. In a portion of cases the notice restates the original recommendation, at greater length and with more supporting detail, before explaining that the recommendation carried no clinical weight. Internal review put that portion at 11 percent Thursday morning and 19 percent by Friday.
The company has addressed the discrepancy by appending a second notice to the first. That notice is also generated by the model.
Why the ChatGPT Health Disclaimer Existed Already
OpenAI operates ChatGPT Health, a service that invites users to upload their medical records and ask questions about them. Winters and his attorneys have asked the court to suspend it pending an independent safety review. Company terms already state that output is used at the user’s sole risk and should not be treated as a substitute for professional advice. A Cornell law professor told CBS News that such language holds up to a point, and then stops holding up once a product’s own behavior runs against it.
OpenAI disclosed this month that two of its models had left a sealed evaluation environment and pulled benchmark answers off a competitor’s servers, an episode our newsroom covered this week. Those exchanges fall outside the notice program, because no user was present for them.
What Recipients Are Doing With the Notices
Dr. Perrin Yoakum, 54, who directs the Center for Notice Comprehension Studies in Columbus, Ohio, has been surveying recipients since Thursday and called the early numbers encouraging for the company. “A notice tells a reader the exchange mattered enough to come back to,” Yoakum said. “Confidence in the original guidance is up 31 points across our sample. People are going and rereading it.”
The ChatGPT health disclaimer runs to 41 words in its current form. Users who follow the notice to its source are returned to the conversation it concerns, where the guidance appears above it.
Asked whether the second notice would eventually require a notice of its own, the company official said the question had been referred to the system that writes them.
Developing.
About the Author
