Your data in ChatGPT: what to change and what to avoid
Training, Memory, Temporary Chat, connected services, deletion, and practical data minimization.
ChatGPT guide · as of July 10, 2026 · 5 minutes read · details change — confirm current specs on chatgpt.com
ChatGPT is a cloud service. Anything you type, upload, photograph, record, connect, or allow it to access may be processed on systems you do not control.
Privacy settings reduce certain uses of data. They do not make ChatGPT local, invisible to OpenAI, or appropriate for every confidential task.
As of July 2026, review OpenAI's current consumer privacy page, Data Controls FAQ, and privacy policy.
The controls are separate
| Control | What it affects | What it does not necessarily do |
|---|---|---|
| Improve the model for everyone | Whether new consumer chats may be used to improve models | Stop cloud processing or all retention |
| Memory | Whether details may personalize later chats | Delete every old chat |
| Reference chat history | Whether prior conversations inform future responses | Turn a chat into a reliable database |
| Temporary Chat | Keeps the chat out of history and memory and out of training | Guarantee immediate deletion |
| Delete chat | Removes it from your view and starts deletion processes | Delete saved memory automatically |
| Delete memory | Removes the saved memory | Delete the original chat |
| Export data | Gives you a copy of account data | Create a perfect human-readable archive |
| Delete account | Starts account and data deletion | Undo data already sent to third parties |
OpenAI says Temporary Chats are not used for training and may be retained for up to 30 days for safety. Check current policy before relying on that window.
Training settings
For consumer ChatGPT, the “Improve the model for everyone” control determines whether new conversations may be used to improve models.
Turn it off if you do not want that use. The setting is not a confidentiality agreement. Authorized processing, abuse monitoring, security, legal obligations, and retention rules may still apply.
Business, Enterprise, Edu, and API products have different defaults and terms. OpenAI says organizational and API data is not used for model training by default, but administrators and developers still need to configure retention and access appropriately.
Memory deserves its own review
Memory can store details across chats. That may be convenient for preferences, projects, and recurring context. It also creates the chance that sensitive or outdated information appears later.
Ask ChatGPT what it remembers. Correct errors. Delete memories that are no longer useful. Turn Memory off or use Temporary Chat when the topic should not affect future conversations.
Remember:
- Deleting a chat does not necessarily delete a saved memory.
- Deleting a memory does not necessarily delete the chat where it appeared.
- Turning Memory off does not erase existing chats.
- Shared or work accounts may have additional administrator controls.
Connected services widen the boundary
A connection to email, cloud storage, calendars, health data, or another service may let ChatGPT retrieve information or take actions.
Before connecting, ask:
- What can it read?
- Can it write or delete?
- Is access limited to selected files or the whole account?
- Who else controls the workspace?
- What logs are kept?
- Can outside content manipulate the agent?
- How do I revoke access?
Use least privilege. Prefer read-only access and require confirmation for actions.
What not to paste
Do not put these into a normal ChatGPT conversation unless you have an approved reason and suitable account terms:
- Passwords and recovery codes.
- API keys, private keys, and authentication tokens.
- Full Social Security or government ID numbers.
- Unredacted medical, legal, employment, or client records.
- Confidential source code or business plans.
- Children's identifying information.
- Alarm codes, camera credentials, or detailed security layouts.
- Data your employer or contract prohibits sending externally.
Redaction helps, but names can often be inferred from context. Minimize the whole story, not just obvious identifiers.
Shared links and files
A shared chat link can expose the conversation to anyone with access to the link, depending on current product behavior. Review the entire chat before sharing, including earlier messages and uploaded material.
Files may remain associated with chats, Projects, GPT knowledge, or organizational workspaces. Deleting the local copy on your computer does not remove the cloud copy.
Keep the authoritative original somewhere you control.
Account security
Use a unique password or secure sign-in method, enable multifactor authentication, review active sessions, and protect the email account used for recovery.
| Risk | Practical control |
|---|---|
| Stolen password | Unique password and MFA |
| Exposed shared computer | Sign out and use device encryption |
| Sensitive lock-screen preview | Hide notification content |
| Malicious connected app | Review and revoke permissions |
| Overshared conversation | Inspect before creating a shared link |
| Account compromise | Review sessions and recovery methods |
The more Memory and connected services you use, the more valuable the account becomes to an attacker.
When this is the wrong tool
ChatGPT is the wrong tool when policy, law, contract, or personal risk requires data to remain entirely on your device or controlled network.
Use a local model, approved enterprise environment, redacted workflow, or conventional software. For legal compliance, obtain qualified advice rather than assuming a setting is enough.
The most effective privacy control is still data minimization: do not send what the task does not require.
