What your prompts reveal about you

A prompt can reveal more than the answer requires. The practical privacy decision often happens before the text is sent: which details are necessary, and which are merely convenient?

I checked OpenAI’s current help pages on 6 October 2026. Their controls vary by account and workspace. OpenAI says that turning off Improve the model for everyone means new conversations are not used for training, while regular saved chats remain in history. Feedback can create an exception: feedback on a response may allow the associated conversation to be used for training. These are OpenAI’s policy statements, not an independent audit or worldwide privacy rule.

Temporary Chat has its own boundaries. OpenAI says temporary chats are not used for model improvement while temporary, do not create or update memories, and may be retained for up to 30 days for safety. A personalized temporary chat can use existing memories, custom instructions, or plugins; an unpersonalized option can be chosen before starting. Saving a temporary chat converts it to a regular chat, after which account settings apply. Data sent through a third-party action follows that recipient’s policy and may be retained longer. Turning off memory also does not disable limited safety-relevant context.

Now consider a fully fictional prompt. A person asks an assistant to rewrite a three-sentence email declining a meeting. The draft needs only the meeting date, the recipient’s role, and the sender’s preferred tone. The person also includes a home address, a medical appointment, and a private phone number, although none is needed for the rewrite.

My choice would be to remove the address, appointment, and phone number before sending. That is not a promise that the remaining text is private, and it is not a claim that a setting makes the conversation stateless. It is simply data minimization: do not hand over details the task does not require. If the role or date could identify someone in the real situation, I would generalize those too while preserving the writing problem.

The prompt itself is part of the decision. Account controls describe later uses and retention, but they do not make irrelevant detail necessary. Check the account or workspace policy, distinguish regular from temporary chat, and ask whether each personal fact changes the requested answer. If it does not, leave it out. A shorter prompt can still contain the full task while revealing less about the person who asked it.