ChatGPT
What the default settings do
ChatGPT’s consumer tiers (Free, Plus, and Pro) default to sharing your conversations with OpenAI for model training. The setting is called “Improve the model for everyone,” and unless you turn it off, the things you type, the files you upload, and the responses you receive can all become training data for future models. This is separate from whether your chat history is saved: you can keep history on for your own reference while still opting out of training.
Memory is also on by default and works in two layers. “Saved memories” are facts you have explicitly asked ChatGPT to remember, like your name or dietary preferences. “Reference chat history” is a quieter layer where ChatGPT infers patterns from past conversations, such as noticing you often ask about a particular city or coding language, and uses that to shape future answers. Both build a persistent profile of you over time.
ChatGPT can also connect to outside services. What used to be called “Connectors” was renamed “Apps” in late 2025, and it covers things like Gmail, Google Drive, and Canva, each granted scoped access once you authorize it. None of this is on by default; it only activates once you connect something.
Voice conversations are transcribed, and by default OpenAI does not train on the raw audio or video itself, though the transcript can still feed into training if the general model-improvement toggle is on. Separate toggles exist specifically for sharing audio and video clips, and those are off unless you turn them on.
Business tiers work differently. ChatGPT Team, Enterprise, Business, and Edu do not train on your organization’s data by default, the opposite default from the consumer product. If you’re on a personal account through work, it’s worth checking which kind you actually have.
A reasonable privacy baseline
These settings are found in Settings, opened from your profile icon in ChatGPT on web or in the mobile app. Menu names have shifted a few times as OpenAI reorganizes its settings pages, so if a label doesn’t match exactly, look nearby under the same section.
Model training
Settings → Data Controls → Improve the model for everyone → Off
This stops new conversations from being used to train OpenAI's models. It does not retroactively remove anything already used for training before you turned it off, and it does not delete your chat history, which is a separate setting.
Thumbs up / thumbs down feedback on responses → Use sparingly if you want training fully off
Rating a specific response, even with the toggle above turned off, gives OpenAI permission to use that individual conversation for training. This is easy to miss since the two settings feel unrelated.
Memory
Settings → Personalization → Reference saved memories → Off
Turns off the facts ChatGPT has explicitly stored about you. Turning this off also disables "Reference chat history" automatically, since the two are linked in that direction.
Settings → Personalization → Reference chat history → Off
Stops ChatGPT from drawing on patterns across your past conversations. You can turn this off independently while leaving saved memories on, if you want explicit facts remembered but not implicit pattern-matching.
Temporary Chat → Use for one-off or sensitive conversations
Available from the new-chat menu, Temporary Chat does not appear in your history, does not draw on or contribute to memory, and is not used to train models. OpenAI still retains temporary chats for up to 30 days for safety and abuse monitoring, so it is private in the everyday sense but not a fully ephemeral or encrypted mode.
Connected apps
Settings → Apps & Connectors → Review and remove unused connections
Each connected app, such as Gmail or Google Drive, is granted access you approved individually and can be disconnected at any time from this list. Access is scoped to what you authorized, but it's easy to accumulate connections you forgot you granted.
Data controls
Settings → Data Controls → Export data → Request a copy
Sends a link to download your full ChatGPT history and account data. It can take up to several days to arrive by email, and the download link itself expires within 24 hours of being sent.
Settings → Account → Delete account → Use if you want to leave entirely
This is permanent. OpenAI states that account data is deleted within 30 days of deletion, except where it is required or permitted to retain a limited set of data longer for legal reasons.
Voice
Data Controls → Include your audio recordings / Include your video recordings → Leave Off (off by default)
These separate toggles control whether raw voice and video clips, not just their transcripts, are used to train models. They are off by default; leaving them off means only the model-training toggle above governs whether transcripts are used.
Trade-offs
- Turning off model training does not make ChatGPT itself more private in the moment; your conversations are still processed to generate responses and are still stored unless you also manage history and retention.
- Rating responses with thumbs up or down is one of the more useful ways to give OpenAI direct feedback, but it overrides your training opt-out for that specific exchange, which is a genuine trade-off if you use feedback often.
- Turning off memory means ChatGPT stops adapting to your preferences over time, so you’ll repeat context (your job, your writing style, ongoing projects) in each new conversation.
- Using Temporary Chat by default for everything means losing the convenience of searchable history, and it isn’t truly ephemeral since a 30-day safety retention window still applies.
- Disconnecting apps you rarely use is low-cost, but reconnecting Gmail, Drive, or similar tools each time you need them adds friction if you use them often.
Who should not change these settings
- If you rely on ChatGPT remembering your writing style, ongoing projects, or recurring preferences, keeping memory on is a reasonable and common choice; the convenience is the point of the feature.
- If you actively want to help OpenAI improve its models and are comfortable with your conversations contributing to that, leaving “Improve the model for everyone” on is a valid position, not a mistake.
- If you use ChatGPT for work through a Team, Enterprise, Business, or Edu account, your organization’s data is already excluded from training by default; there is usually nothing to change on the training front, though you should still confirm which account type you have.
- If you depend on Gmail, Drive, or other connected apps daily, disconnecting them for privacy’s sake will break workflows you may prefer to keep intact; scoped, revocable access is a reasonable middle ground for many people.