On April 25, 2023, OpenAI announced on its website that ChatGPT users could turn off chat history. Conversations started with history disabled would not train its models or appear in the history sidebar. OpenAI said it would retain those conversations for 30 days, review them only when needed to monitor abuse, then permanently delete them.

That announcement described a retention setting, not a confidential relationship. The record separates questions that the phrase “save my chats” collapses: who stores the conversation, who may inspect it, and what can prevent its deletion. A chat printout is another retained copy, not a privacy setting.

Where did you save it?

OpenAI’s export instructions describe a downloadable ZIP file containing chat history and other account data. Exporting therefore creates a copy outside the account. It does not, by itself, delete the conversation held by OpenAI.

Sharing creates a different access problem. OpenAI’s shared-links documentation says anyone with the link can view the linked conversation. It does not offer granular permissions for those links. If another user imports the conversation into their own history, deleting the shared link does not delete that imported copy.

Those are materially different records. A conversation in an account remains subject to the provider’s controls. An export becomes a file the account holder possesses. A shared conversation may become a record someone else possesses.

The relevant question is not simply whether the chat was saved. It is where the copies went and which holder can still remove them.

Does switching off training close the file?

No. OpenAI’s Data Controls FAQ says turning off “Improve the model for everyone” leaves conversations in chat history while preventing their use to train ChatGPT. The setting changes a permitted use. It does not eliminate storage.

OpenAI also identifies an exception involving feedback. Its model-improvement documentation says that, even after opting out of training, submitting feedback can allow the entire associated conversation to be used for training. A thumbs-up or thumbs-down is not necessarily limited to the response being rated.

Temporary Chat addresses a different combination of functions. OpenAI says these conversations do not appear in history, do not create memories and are not used to improve its models. Its FAQ nevertheless allows a copy to be kept for up to 30 days for safety purposes.

Temporary Chat also does not govern every recipient. The FAQ says information sent to third parties through GPT actions is subject to those recipients’ privacy policies. They may retain it longer. The provider’s temporary setting does not impose the same deadline on another company.

Who can read it without a court?

OpenAI’s consumer-services FAQ identifies authorized personnel and trusted service providers who may access user content when necessary. The stated purposes include investigating abuse or security incidents, providing support, handling legal matters and improving model performance unless the user has opted out.

The same document says access is restricted through technical controls, limited to authorized personnel on a need-to-know basis, and monitored and logged. That is a controlled-access policy. It is not a promise that no human can read the conversation.

A work account introduces another reader. OpenAI’s privacy policy says administrators of business accounts may access and control the account, including its content. The employer’s administrator and the model provider are separate actors with separate reasons to retain a record.

The account boundary matters more than the conversational tone. A personal question typed into an employer-administered account remains content within an account the administrator may control.

What survives deletion?

OpenAI’s deletion instructions distinguish archiving from deleting. Archiving hides a conversation from the main history view without removing it from the account. A tidy sidebar is not a smaller archive.

Deleting removes a chat from view immediately and schedules permanent deletion from OpenAI’s systems within 30 days. The instructions identify exceptions for content already de-identified and disassociated from the account, and for retention required by security or legal obligations.

Memory has its own storage rules. OpenAI’s Memory FAQ says saved memories are stored separately from chat history. Deleting a conversation does not delete a saved memory created from it. To remove that information fully from those features, the documentation directs users to delete both the saved memory and the original chat.

“Delete the conversation” therefore describes one operation, not a universal recall of everything derived from it. History, saved memory and copies held elsewhere require separate accounting.

Can a lawsuit keep it alive?

Yes. On June 5, 2025, OpenAI publicly described a preservation order in its litigation with The New York Times. The company said the order required it to retain certain consumer ChatGPT and API content that would otherwise have been deleted. OpenAI objected to the demand and said the preserved material would be held separately with restricted access.

That episode documents a legal interruption of ordinary deletion. It does not establish that every conversation remains subject to that historical order, or that the opposing litigant automatically received the preserved records. Preservation and disclosure are different steps.

The user’s own records can also matter. Federal Rule of Civil Procedure 34 permits requests for electronically stored information within a party’s possession, custody or control. Rule 26 limits discovery to nonprivileged material relevant to a claim or defense and proportional to the case.

A saved AI transcript can fall within those categories. Its status as a conversation with software does not create an exclusion from the rules. Whether a particular transcript must be produced depends on its contents, the dispute and any applicable protection.

Can the government ask for it?

Government access presents another route, with different requirements. Under 18 U.S.C. §2703, governmental entities can require covered service providers to disclose specified communications or records through the legal processes the statute prescribes. The requirements differ by the material and service involved.

Section 2702 separately restricts voluntary disclosure by covered providers, subject to statutory exceptions. These provisions are not a promise that every AI conversation receives identical treatment. Their application depends on the provider’s legal classification, the information sought and the circumstances.

OpenAI’s privacy policy expressly permits disclosure to government authorities or other third parties when required by law and for specified safety, security and legal purposes. That policy is notice of possible disclosure, not permission for any official to demand any conversation without process.

The distinction is consequential: information can be excluded from model training while remaining available for a legally required disclosure.

Is deleting everything safer?

Not once preservation duties apply. Federal Rule of Civil Procedure 37(e) addresses electronically stored information that should have been preserved in anticipation or conduct of litigation. It allows remedies when a party fails to take reasonable preservation steps and the information cannot be restored or replaced. The most severe measures require a finding of intent to deprive another party of its use.

That rule prevents a simple answer built around indiscriminate deletion. A transcript documenting a disputed instruction or decision may be evidence rather than clutter. The legal question changes when a duty to preserve attaches: removing a sensitive record can create consequences of its own.

Should you keep the transcript?

Not by default. Save an AI conversation when its value as a record justifies retaining its contents, or when preservation is required. The documents support neither treating chat history as a confidential notebook nor deleting relevant evidence because it is uncomfortable. Training controls govern training. Retention rules govern storage. Account permissions and legal process govern other people’s access. Keeping the conversation means accepting those distinctions, not acquiring privacy because the exchange felt private.