Everyone Who Uses ChatGPT — Turn Off These 2 Settings NOW!
If you've never gone into ChatGPT's settings, two toggles are almost certainly working against you right now. The first quietly hands your conversations over to OpenAI's training pipeline. The second lets ChatGPT build a running profile of you from everything you've ever typed, then feed that profile back into future chats — including ones where you didn't expect it to know anything about you.
Neither one is a conspiracy. Both are documented, both are legal, and both are on by default for most personal accounts. That's exactly why they're worth two minutes of your time: the risk isn't that OpenAI is doing something sneaky, it's that almost nobody goes looking for a setting they didn't know existed.
Setting 1: "Improve the model for everyone"
This is OpenAI's model-training toggle, and it's the one most people have at least heard of. When it's on, the content of your conversations can be used to train and improve future versions of ChatGPT. It doesn't mean a human is reading your chats line by line for fun, but it does mean your prompts and OpenAI's responses become part of the raw material future models learn from.
Who has it on by default
- Free, Plus, and Pro accounts on a personal workspace: training is on unless you switch it off yourself.
- Business, Enterprise, Edu, and API workspaces: excluded from training by default, so this specific toggle matters less there.
- Temporary Chats: never used for training and never saved to your history, regardless of this setting.
What turning it off actually changes
Once you switch this off, new conversations stop being used to train OpenAI's models going forward. A common misconception is that this also wipes your chat history — it doesn't have to. According to OpenAI's own help documentation, you can keep your conversation history intact while opting new chats out of training, so you don't have to choose between privacy and a usable chat archive.
Where to find it
- Click your profile icon in the top-right corner of chatgpt.com.
- Select Settings, then Data Controls.
- Find "Improve the model for everyone" and switch it off.
Setting 2: "Reference chat history" (Memory)
This is the one people miss even after they've locked down training, because it lives under Personalization, not Data Controls, and most guides only mention the first setting. Memory in ChatGPT works in two connected layers. Saved memories are specific facts you've asked it to remember, or that it decided on its own were worth keeping — your job, a recurring project, a dietary preference. Reference chat history is broader: it lets ChatGPT draw on the general content of past conversations, even details you never explicitly told it to save, to make later chats feel more personalized.
The two are linked in one important way: turning off "Reference saved memories" also turns off "Reference chat history," but you can leave saved memories on and switch off chat history referencing by itself if you want a middle ground.
What changed in 2026
OpenAI has been rolling out a rebuilt memory system through 2026, moving from a flat list of saved items toward a written "memory summary" generated by a background process the company calls Dreaming. The underlying mechanics changed, but the two user-facing controls — saved memories and chat history referencing — are still the switches that determine whether any of it happens.
Where to find it
- Click your profile icon, then Settings.
- Open Personalization, then Memory.
- Turn off Reference saved memories to disable both layers at once, or turn off just Reference chat history to keep explicit saved memories while stopping the broader cross-chat referencing.
Why these settings default to "on" in the first place
It's worth understanding the incentive before you judge it too harshly. Training on real conversations genuinely makes a model better at understanding how people actually phrase questions, and a memory layer genuinely makes repeat use of an AI assistant less repetitive. Neither feature exists purely to extract data — both solve a real product problem. The issue isn't that these settings exist; it's that opt-out, rather than opt-in, is the default for personal accounts, and most people never open a settings menu unless something breaks. A setting that's technically disclosed in a help article isn't the same thing as a setting most users have actually seen or chosen.
That gap between "documented" and "understood" is exactly where privacy problems tend to live, not just with ChatGPT but with most consumer software. It's why checking Data Controls and Personalization takes two minutes but rarely happens on its own — nothing in the normal flow of using ChatGPT ever points you toward either menu.
What these two toggles do not do
Turning both settings off is not the same as making ChatGPT forget you existed. A few limits are worth knowing before you assume you're fully reset:
- Turning off training stops future conversations from being used — it doesn't retroactively pull your past chats out of any training that has already happened.
- Switching off "Reference chat history" triggers deletion of what it had learned from your past chats, but that deletion happens on OpenAI's servers over a window of roughly 30 days, not instantly.
- OpenAI states that disabling memory and personalization does not turn off the narrow, safety-related use of context in rare high-risk conversations — that's a separate system from personalization memory.
- Deleting a chat from your history does not automatically remove any memory that chat had already created.
Should you actually turn these off?
Not everyone needs both switches flipped. The honest answer depends on what you use ChatGPT for.
| Turn them off if… | Leave them on if… |
|---|---|
| You discuss client work, legal matters, health details, or anything confidential | You mainly use ChatGPT for generic, low-stakes questions |
| You want ChatGPT to treat every session the same way, with no drifting personalization | You like ChatGPT remembering your writing style, ongoing projects, or preferences |
| You share a device or account with other people | It's a private account only you use, for low-sensitivity tasks |
| You'd rather opt out and re-enable later than find out something sensitive was stored | You've already reviewed your memory summary and you're comfortable with what's there |
If any part of your work involves information you wouldn't want repeated back in a future session with someone else logged into the same tool — even accidentally, even in aggregate — that's the clearest signal to switch both off today rather than "eventually." For a broader look at locking down how much of your activity any online tool holds onto, our step-by-step privacy and settings guides cover the same kind of default-opt-in traps in other apps you're probably also using.
Frequently asked questions
No. You can disable training while keeping your chat history visible and searchable. The two are separate settings, even though some third-party guides conflate them.
The training toggle matters less there, since Business, Enterprise, and Edu content is excluded from training by default. The memory and personalization setting is separate and still worth checking regardless of plan.
Not instantly. Switching off chat history referencing starts the deletion of what it had learned, but OpenAI's own documentation describes that process completing within about 30 days rather than the moment you flip the switch.
Yes — use Temporary Chat for a single session. It skips your chat history, isn't used to build memory, and isn't used for training, without you needing to change your account-wide settings at all.
Yes, both settings can be switched on or off at any time with no restrictions, and turning saved memory off doesn't erase what was already remembered — it just stops ChatGPT from adding to or using it going forward.
It makes it less personalized, not less capable. Within a single conversation ChatGPT still has full context of everything you've typed in that session — you only lose the cross-session personalization, not the model's ability to follow a conversation.
Final verdict
Turn both off if you use ChatGPT for anything you wouldn't want summarized back to you — or influencing a future answer — by an AI that doesn't forget as easily as you'd assume. It costs two minutes and neither change is permanent.
Leave them as-is only if you've deliberately decided the personalization is worth it and you've never used the tool for anything sensitive. For most people who haven't actually checked, that second condition is the one that turns out to be false.