Why This Is the AI Version of Yesterday's Check
Authenticator was a build number. Outlook was About Outlook. ChatGPT is a settings path most people skip because the product feels like a search box. OpenAI's September 21, 2026 release notes rolled a Privacy Center to signed-in Free, Go, Plus, and Pro - a map of the same controls, not a replacement for Settings.
Two different levers, easy to mix up:
- Improve the model for everyone = may your new chats train OpenAI's models. Off does not delete old chats. OpenAI says data sharing is on by default on personal Free, Plus, and Pro.
- Memory = may ChatGPT personalize later replies from saved memories, chat history, custom instructions, files, and (if you connected it) apps such as Gmail. Turning Memory off does not wipe the sidebar. Deleting a chat does not always delete a separate saved memory.
Honest limit: this is not Microsoft MSRC. There is no CVSS. There is no "exploited in the wild" line. The value is the same as Authenticator: one sitting, three clicks, you know what is on.
Who This Hits
| What you use | This article? | What OpenAI documents |
|---|---|---|
| chatgpt.com or the ChatGPT iOS/Android app, personal Free / Plus / Pro | Yes | Training default on; Memory and Data controls in Settings |
| ChatGPT Go (where ads exist) | Yes, plus ads toggles | Separate ad personalization controls; paid Plus and up have no ads |
| Signed-out chatgpt.com in a browser | Partial | Settings still has Improve the model for everyone; choice is saved in that browser |
| ChatGPT Business, Enterprise, Edu, Healthcare workspace | No as the training default | OpenAI does not train on that workspace content by default. Ask your admin about Memory. |
| Microsoft Copilot, Google Gemini, Claude | No | Different products. Copilot on Windows is not this path. |
| API / Codex Cloud environments | Only the ChatGPT toggle overlap | Codex has a separate "Include environments" training switch that ChatGPT Data controls do not change |
The Two-Minute Path
- Training. Profile → Settings → Data controls → Improve the model for everyone → Off → Done. Mobile: sidebar → profile → Data controls. One account covers web and the apps when you are signed in.
- Memory list. Settings → Personalization → Memory. Read the summary. Delete junk. OpenAI: "Don't mention this again" reduces future mentions; it does not delete the original chat. "Delete and turn off memory" clears the summary items shown and turns Memory off - still not your old chats.
- Ask it. In a normal chat: "What do you remember about me?" Then delete what should not be there. Sources under a reply (when shown) can point at a past chat, a saved memory, a file, or a connected email.
- Shared links. Data controls → Shared links. Revoke anything you ever generated from a chat that had work text in it.
- Apps. If Gmail or another connector is listed, disconnect what you do not use. Disconnect stops future access; it does not rewind chats that already used that data.
Temporary chat: start it from the chat UI, not from Settings. While it stays temporary it is out of history, does not create memories, and is not used for training. OpenAI may retain it up to 30 days for safety. If you save it, it becomes a normal chat and follows your Memory and training switches. August 2026 notes: you can choose to personalize a temporary chat; leave that off for secrets.
What This Does Not Do
Turning training off does not unsay 2024 chats already used for training. OpenAI's help is forward-looking for new conversations. Deleting chats is a separate control (Data controls → Delete all chats, or delete one from the sidebar). OpenAI describes deleted content leaving systems on a delay (commonly described as within 30 days, with legal exceptions). This site will not invent a "secure wipe."
If a chat already leaked a password, rotate the password. If it leaked a mailbox, use the 15-minute mailbox audit. Fake "ChatGPT support, send crypto" popups are still scams: fake virus popups.
Windows and Office remain separate checks: Outlook Reading Pane build, Authenticator Android build.
Takeaways
- Personal ChatGPT: training is on by default. One Data controls toggle turns it off for new chats.
- Memory is a second switch. Off ≠ deleted chats. Deleted chat ≠ deleted saved memory.
- Temporary chat: no history, no new memories, no training; up to 30 days for safety.
- Work ChatGPT (Business/Enterprise/Edu): training off by default. Do not apply this article's panic to a managed workspace.
- Not a CVE. Two minutes. Then stop.
FAQ
I do not see Memory. OpenAI: features vary by plan, region, platform, and workspace. If the row is missing, you are not on that rollout yet. Data controls can still be there.
Teen account. A parent can manage supported Memory controls in Parental controls. Do not "fix" a child's account by guessing.
Voice mode. OpenAI documents a separate training-share control for Voice. Check Data controls / Privacy Portal if you use Voice daily.
Will this make ChatGPT dumber? It will make it less personalized. That is the point for tax IDs and HR drafts. Use a normal chat for recipes.
Gemini / Copilot instead. Different settings trees. This URL is ChatGPT only so the clicks stay true.
Sources
- OpenAI Help - Memory in ChatGPT
- OpenAI Help - Data controls in ChatGPT
- OpenAI Help - ChatGPT release notes (Privacy Center, 21 Sep 2026)
- OpenAI Help - Temporary chat in ChatGPT
- OpenAI Privacy Portal
Changelog. October 29, 2026 - Initial publication: two-minute ChatGPT training + Memory check from OpenAI Help; personal vs workspace; no CVE claim.
Get US tech fixes weekly
No spam. Unsubscribe anytime. Tested on Windows 11 24H2.
Comments
Comments
Post a Comment