The four settings that matter are the model training toggle in Data Controls, Temporary Chat, saved memories under Personalization, and shared links. Business and Enterprise plans exclude your data from training by default; consumer plans do not until you turn it off.
Short answer
The four settings that matter are the model training toggle in Data Controls, Temporary Chat, saved memories under Personalization, and shared links. Business and Enterprise plans exclude your data from training by default; consumer plans do not until you turn it off.
Most ChatGPT privacy advice is either "do not use it" or a restatement of OpenAI's marketing. Here is the useful version: the specific controls, what each one changes, and where the gaps are.
Every control that matters
| Setting | Where | What it changes |
|---|---|---|
| Improve the model for everyone | Settings, Data Controls | Whether new conversations can be used for training |
| Temporary Chat | Chat header toggle | Excludes a single conversation from history, memory, and training |
| Memory | Settings, Personalization | Whether facts are extracted from chats and persisted |
| Shared links | Settings, Data Controls | Lists and revokes public snapshots you have created |
| Export data | Settings, Data Controls | Produces a full archive of everything held |
| Delete all chats | Settings, Data Controls | Starts the purge of your conversation history |
The training toggle is forward looking only
Turning off model improvement stops future conversations from entering the training pool. It does not retract data already collected, and it does not affect anything that happened before you flipped it. If you have been using a consumer account since launch with the default settings, that history is already in.
Memory is the quiet one
Memory reads your conversations and stores extracted facts about you: your job, your projects, your writing preferences, sometimes things you would not have chosen to persist. It survives deleting your chat history. Review it periodically under Settings, Personalization, Manage memories. Anyone handling client work should check what has accumulated there before assuming the account is clean.
Shared links are public until revoked
A share link is a snapshot published at a public URL. Anyone with the link can read it, and search engines have indexed shared AI conversations in the past when the URLs leaked. Settings, Data Controls, Shared links shows every one you have created. Delete the ones you no longer need.
Work use
If you are using ChatGPT for anything commercial, the plan is the control that matters most. Team, Enterprise, and API traffic are excluded from training by default and come with administrative controls and retention settings. A personal Plus account does not, and no amount of careful toggling makes a personal account an appropriate place for client data covered by a contract.
Also worth knowing: on a Team or Enterprise workspace, your administrator can access workspace conversations and run compliance exports. Privacy from OpenAI and privacy from your employer are different questions with different answers.
The part you control completely
Everything above is about limiting what the platform keeps. The other half is making sure you keep what matters, so you are free to delete without hesitation.
Request an export from Data Controls, then convert conversations.json into readable documents. ChatExports parses the file in your browser using local JavaScript, which means the transcripts are never uploaded anywhere during conversion. You end up with PDF, Markdown, DOCX, or CSV files on your own machine, and the platform side becomes disposable.
That is the strongest privacy position available: keep your own copy, delete theirs.
Frequently asked questions
Does OpenAI sell my conversations?
OpenAI's privacy policy states it does not sell personal data. It does share data with service providers and can disclose it under legal process.
Can I opt out of training retroactively?
No. The toggle applies to future conversations only. Data already collected under the previous setting stays collected.
Is the API more private than the ChatGPT app?
API traffic is excluded from training by default and retained for a limited abuse monitoring window, so yes, for that specific question.
What is the safest way to use ChatGPT for sensitive work?
A plan excluded from training, Temporary Chat for the sensitive parts, no identifying details in prompts, and your own local archive of the output.