Choose the privacy outcome before you touch a switch. “Do not train on this,” “do not show this in history,” “do not remember this later,” and “delete what I already shared” are four different requests. A product may offer one without the others.
The practical rule is still blunt: do not paste a secret merely because a privacy toggle exists. But current controls can reduce exposure when the task genuinely needs an AI chat. The useful question is not whether a product has a privacy page. It is which future use, visible record, memory, and retention window the selected control changes.
Pick the outcome, not the label
| Your goal | Control to inspect | What to verify afterward |
|---|---|---|
| Keep ChatGPT history but stop future training use | Settings → Data Controls → turn off Improve the model for everyone | A new normal chat still appears in history; the account-wide training setting remains off |
| Use ChatGPT once without history or memory | Start a Temporary Chat | The chat is absent from history and no memory was created |
| Stop future Gemini chats from appearing in Activity or training models | Turn off Keep Activity | A new chat is absent from Gemini Apps Activity; note any Connected App that became unavailable |
| Remove a past visible record | Delete the chat or activity in that product | Check memory, connected services, shared links, and copied output separately |
A training choice is usually forward-looking. It does not promise to pull old text out of a model, erase a saved memory, recall a shared link, or delete a copy sent to another service. Treat each of those as its own cleanup job.
ChatGPT separates training from history
OpenAI's current Data Controls FAQ says signed-in users can turn off Improve the model for everyone under Settings → Data Controls. New conversations can remain in chat history while no longer being used to train ChatGPT. The choice syncs across the account's web and mobile use.
- Open your profile menu, then Settings and Data Controls.
- Turn off Improve the model for everyone.
- Start a harmless new chat and confirm it still appears in history if that is the outcome you wanted.
- Open Data Controls again and confirm the setting stayed off.
Temporary Chat is a different control. OpenAI says those chats do not appear in history, do not create memories, and are not used for training. They may be reviewed for abuse and are deleted from OpenAI's systems after 30 days. “Temporary” therefore means a defined service window, not that the text vanishes the instant the tab closes.
Gemini ties more behavior to Keep Activity
Google's current Gemini Apps Privacy Hub makes the tradeoff more visible. Keep Activity affects saved chat activity, model improvement, personalization, and access to some Connected Apps.
| State | Future chat handling | Tradeoff or exception |
|---|---|---|
| Keep Activity on | Chats and shared content are saved in Activity and may be used to improve services, including generative AI models | Activity auto-deletes after 18 months by default; Google offers 3 months, 36 months, or no auto-delete |
| Keep Activity off | Future chats do not appear in Activity and are not used to train models unless you submit feedback | Chats are retained with the account for 72 hours; some Connected Apps are unavailable |
| Temporary Chat | The chat is not used to train Google's AI models and is retained with the account for 72 hours | Google may still process it to respond and protect users and systems |
| Delete Activity | Removes the activity tied to the account | Material already reviewed and disconnected from the account may be retained for up to three years |
Turning off Keep Activity does not turn off unrelated Google settings such as Web Activity or Location History, and deleting Gemini Apps Activity does not delete records held by another Google service. A single Google account can contain several overlapping histories. Check the one that actually received the data.
Do not let feedback undo the choice
Feedback is a separate data path. Google says that when Keep Activity is off, choosing to send feedback can include the feedback, related conversation context, uploads, and Connected App data. OpenAI also distinguishes ordinary training controls from abuse and safety review.
If a conversation contains material you deliberately kept out of training, do not attach that conversation to a thumbs-up, thumbs-down, bug report, or support ticket without reading the disclosure beside the submission control. Describe the problem with invented details when the original content is not needed.
Run a one-minute pre-paste check
- Name the account. Confirm whether it is personal, work, school, or an API workspace. Similar product names can carry different data terms.
- Name the data. Remove passwords, recovery codes, private links, account numbers, full customer records, and details that identify another person.
- Name the outcome. Decide whether you need history, personalization, connected apps, or only a one-off answer.
- Set the narrow control. Use a training opt-out when you want history; use a temporary mode when you do not need a lasting chat; disconnect apps when the task does not need them.
- Test with harmless text. Create a short invented prompt, then check history, Activity, and memory before using real material.
Redaction is not just removing a name. A job title, rare diagnosis, exact date, workplace, and quoted sentence can identify a person together. Summarize only the facts the answer needs. When a complete document is unnecessary, do not upload it.
Delete with a map
| Possible record | Where to check |
|---|---|
| Visible conversation | The chat list or product Activity page |
| Saved memory or personalization fact | The product's memory or personalization settings |
| Connected-service copy | The connected drive, mailbox, calendar, notes app, or other destination |
| Shared link or collaborator copy | Sharing controls and the recipient's copy |
| Exported or pasted output | The document, ticket, message, or local file that received it |
Deleting your chat cannot recall text someone copied or erase a file created in another service. If the original material included a password, token, recovery code, or private link, rotate or revoke it. Deletion is housekeeping; replacement is what makes an exposed credential stop working.
Work and school accounts need their own check
Consumer help pages do not automatically describe business, enterprise, education, or API products. An employer or school may set retention, logging, connected-app, and model-improvement rules centrally. Confirm the workspace name and its administrator policy before treating a personal-account toggle as the governing promise.
When the task involves customer data, employee records, unreleased work, or another person's private information, use the organization's approved tool and data path. A temporary chat in the wrong account is still the wrong account.
Close the loop after the task
- Delete test chats and any unnecessary real chat from the correct history or Activity page.
- Review memory or personalization separately and remove facts that should not carry into later chats.
- Disconnect apps the task no longer needs and inspect the destination service for copied files or drafts.
- Recheck the training or Activity setting after product updates and before the next sensitive task.
- Keep a short note of the account, control, date checked, and remaining copies. Do not keep the sensitive prompt itself as the record.
The safest privacy setting is a smaller input. The next safest is a control chosen for the exact outcome you want, followed by a check that the product behaved that way.
