Turning off model training does not automatically delete a health-related chat or prevent every kind of safety review. In U.S. consumer accounts, ChatGPT, Gemini, and Claude have different rules for model improvement, retention, and review. Those controls do separate jobs, so a training opt-out is not a deletion request.
Treat “don’t use my chats to train models” and “delete my chats” as two separate requests. Check both before sharing sensitive health details.
How ChatGPT, Gemini, and Claude handle consumer chats
The comparison below reflects the providers’ stated U.S. consumer-account practices: OpenAI’s U.S. Privacy Policy, Google’s Gemini Apps Privacy Hub, and Anthropic’s consumer privacy policy and terms. These are the companies’ stated practices.
| Service | Model-improvement control | Retention and review conditions |
| ChatGPT | Eligible personal accounts can control whether content is used to improve models. | Turning off model improvement does not delete existing chats. Temporary Chats do not appear in chat history or help improve models, and are automatically deleted within 30 days, subject to safety or legal exceptions. |
| Gemini Apps | With Keep Activity on, chats may be used to improve Google services, including generative AI models. Turning it off, or using Temporary Chat, prevents future chats from being used to train Google AI unless you submit feedback. | With Keep Activity off or Temporary Chat, Google retains chats for 72 hours for service and safety purposes; human review may still occur for safety. With Keep Activity on, the default auto-delete period is 18 months, adjustable to 3 or 36 months or indefinitely. Some reviewed chats are disconnected from the account and retained for up to three years. |
| Claude | Anthropic may use consumer Inputs and Outputs to train and improve models unless you opt out. Safety-flagged or explicitly reported content may still be used for safety-related improvement. | Anthropic’s consumer terms state five-year retention when users allow training and 30 days when they do not choose training. |
The practical difference is clearest with Gemini: switching off Keep Activity stops future chats from being used to train Google AI unless you send feedback, but it does not end the 72-hour retention period or all safety review. With ChatGPT, the model-improvement setting and chat deletion are also separate. Anthropic’s stated retention period changes with the training choice, while safety-related exceptions still apply.
ChatGPT Health has separate terms
OpenAI’s Health Privacy Notice applies to ChatGPT Health and Connect Health, not automatically to ordinary ChatGPT chats. It covers specified Health Feature information, including connected medical records, lab results, prescriptions, Apple Health data, and health details entered in Health conversations. The notice says this information is excluded from foundational-model improvement by default. Limited authorized personnel and trusted service providers may access Health Feature data for safety unless the user opts out.
That feature-specific policy does not replace the general controls for a standard ChatGPT conversation. If you type health details into an ordinary chat, the general consumer-account policy is the relevant one.
What HIPAA and the FTC rule cover
In the United States, HIPAA protection depends on a service’s role and relationship to a covered entity, such as a health-care provider or health plan. The U.S. Department of Health and Human Services explains that information a person directs to an app that is neither a covered entity nor a business associate is outside HIPAA protection after the app receives it. A service handling information on behalf of a covered entity may have a different status.
The FTC Health Breach Notification Rule has a narrower, conditional scope. It applies to specified non-HIPAA personal-health-record vendors and related entities when the rule’s definitions and breach conditions are met. A personal health record must have the technical capacity to draw identifiable health information from multiple sources and be managed, shared, and controlled by or primarily for the individual. The rule concerns notification of qualifying breaches of unsecured identifiable health information; it is not a blanket privacy guarantee for every chatbot. See the rule text and FTC and HHS guidance.
What to check before sharing health details
Share only what the question requires. Remove names or account identifiers that are not needed, then check the service’s training or activity setting, retention and deletion controls, feedback options, and any connected apps separately. A setting that changes model training does not, by itself, change every other way a provider says it may retain or review information.
