Three separate settings decide whether your conversations with OpenAI's ChatGPT, Claude, and Gemini get folded into the next version of each model, and none of the three works quite the way most people assume. Here's where each toggle actually lives, what turning it off does and doesn't cover, and the exceptions all three companies still carve out for feedback ratings and safety review — current as of August 18, 2026, because these settings menus move without much notice.
What "off" actually means
Turning any of these three off only protects what you send afterward — none of them undoes training that already happened on a past conversation, and none is a universal "stop using my data" switch. Each company frames it the same way: a forward-looking preference, not a retroactive deletion. Anthropic's own version of this setting didn't exist before August 2025 — before that, its consumer plans trained on chats by default with no opt-out at all, and turning it off today still only drops retention from up to 5 years down to 30 days, not to zero. Google's own buffer is shorter but not zero either: Gemini keeps a new chat for 72 hours regardless of the setting, just to run the service. What each setting specifically excludes differs enough between the three that reading one company's help page tells you nothing reliable about the other two.
Where each setting lives, and what it actually covers
| ChatGPT OpenAI | Claude Anthropic | Gemini | |
|---|---|---|---|
| Setting name | Improve the model for everyone | Help improve our AI models | Gemini Apps Activity |
| Where it lives | Settings → Data Controls | Settings → Privacy | Profile → Gemini Apps Activity |
| On by default (consumer plans) | Yes — Free, Plus, Pro | Yes — Free, Pro, Max | Yes — most personal accounts |
| Already excluded | Business, Enterprise, API | Team, Enterprise, API | Workspace accounts (admin-controlled) |
| Retention once switched off | Not stated as a fixed window on this setting | 30 days, down from up to 5 years | New chats kept 72 hours, not used to train |
Flipping each one off takes under a minute once you know where it actually lives — the paths below are current today, not a memory of an older menu. Google's Gemini is the one most likely to trip people up: the setting is named Activity, not training, and it's easy to assume it only controls your visible chat history.
Turn off training in whichever of these you actually use
- Profile picture → Settings → Data Controls → switch off "Improve the model for everyone." It's on by default for Free, Plus, and Pro; Business, Enterprise, and API accounts are already excluded. Temporary Chat, from the icon top-right, skips both history and training for that one conversation.
- Your name at the bottom of the sidebar → Settings → Privacy → switch it off. This also drops your retention window from up to five years back down to the standard 30 days.
- Profile picture → Gemini Apps Activity → Turn off, or go straight to myactivity.google.com/product/gemini. On by default for most personal accounts.
Where the opt-out doesn't reach
The setting in each product is real, but every company reserves at least one carve-out that survives it no matter how you have it configured:
What the toggle doesn't cover
None of this applies if you're already on a paid Team, Enterprise, or API-level plan — those tiers already exclude that data from training by default at all three companies, which is still worth confirming on your own plan's terms page rather than assuming from this guide. If none of the three toggles feels like enough control, the underlying alternative is running a model yourself instead of renting someone else's — see our guide to deciding between self-hosting and an API. And if "training data" itself is the term that's unclear, the dictionary has the one-paragraph version.
- ChatGPT, Claude, and Gemini each have a separate setting to stop training on your chats.
- Turning any of them off only protects chats sent afterward — nothing already trained gets undone.
- All three still carve out an exception for feedback ratings and safety-flagged conversations.
- Business, Team, Enterprise, and API accounts are already excluded by default at all three companies.
- Caveat: these settings are new and move — Anthropic's own toggle didn't exist before August 2025.