OpenAI has made fundamental changes to the mechanics of temporary chats in ChatGPT, revising the key principle on which this feature had operated since its inception. Previously, a temporary chat by definition left no traces and was not tied to the user's profile; now, in the top-right corner of an empty dialogue window, a menu with two personalization modes has appeared, directly affecting what personal data the AI is authorized to use during the conversation. The update is already available to subscribers on the Plus, Pro, and Business plans, and also partially extends to some free accounts.
Two Modes: Privacy by Default or Access to Account Memory
The first mode — Non-personalized — is set by default. In this mode, the model does not access saved memories, personal instructions, or the user's connected plugins. The second mode — Personalized — allows a temporary chat to read the account's saved memory, use plugins, and apply preconfigured settings. Importantly, the mode must be selected strictly before sending the first message: once the conversation has started, the personalization settings can no longer be changed, and this restriction is enforced at the interface level.
Saving 'Private' Dialogues and Their Impact on Model Training
A significant innovation is the ability to save successful temporary conversations. If a dialogue proved useful, the user can move it into the account history. However, once this is done, the standard profile rules apply to the chat's content, including permission to use the content for training future AI models. Thus, a 'private' conversation moved into the history effectively loses its isolated status and becomes part of the data on which the model may be trained. At the same time, an important limitation applies: a personalized temporary chat can read existing memory entries but does not create new ones while it remains in temporary status.
Thirty-Day Storage and Protective Mechanisms
Regardless of the chosen personalization mode, OpenAI reserves the right to store a copy of a temporary chat for 30 days. According to the developers, this is driven by security considerations. In rare high-risk situations, the service's protective systems may use limited context from previous conversations to adjust responses. This means that even in Non-personalized mode, absolute anonymity is not guaranteed: on the server side, data may be stored and analyzed for a month after the session ends.
Context: Parallel Updates to the ChatGPT Ecosystem
The changes to temporary chats came amid an active phase of development in the ChatGPT ecosystem. In parallel, OpenAI is testing and rolling out group chats, allowing up to 20 participants to hold a joint dialogue with the model, and is periodically updating data deletion mechanisms. Taken together, these updates indicate that the company is restructuring its privacy architecture: from a 'full isolation or full linkage' model to a flexible system in which the user controls the level of the AI's access to their data, while corporate security and training mechanisms remain in effect.
