Stop Claude and ChatGPT training on your chats
Most people assume their chatbot conversations disappear into the void. They don’t. On the consumer plans of both Claude and ChatGPT, your chats are used to train future models unless you go into settings and say no. It’s a five-minute fix — but three things almost every guide skips are the parts that actually matter.
Step by step
First, work out which account you’re on
This is the fork in the road. Consumer accounts train on you by default — that’s Claude Free, Pro and Max, and ChatGPT Free, Plus and Pro. Business accounts don’t — Claude for Work, Team, Enterprise, Education and the API, plus ChatGPT Business, Enterprise, Edu and the API.
Rule of thumb: if your employer pays for it, you’re probably already covered. If you pay for it yourself, you’re probably not. A personal Pro subscription is a consumer account no matter how professionally you use it — which catches a lot of freelancers and small business owners.
Turn it off in Claude
Log in, click your profile name (bottom left) → Settings → Privacy, find the model-improvement setting “Help improve Claude” and toggle it off.
New and resumed conversations stop being used for training, and your retention window drops. Anthropic keeps training-eligible data for up to five years; with the setting off, you’re back to the standard 30 days.
Heads up: Anthropic changed its consumer terms in August 2025. Existing users got a pop-up with a big Accept button and a toggle already switched on, so a lot of people accepted without reading — worth double-checking even if you don’t remember opting in.
Turn it off in ChatGPT
Log in, click your profile icon → Settings → Data controls, find “Improve the model for everyone” and toggle it off. It syncs across devices, so doing it once on any device covers your whole account.
While you’re there, check two more things. Voice and video recordings are separate toggles — if voice mode is on, your audio can be used for training and a human can listen to it, so turn them off. Memory (under Personalization) decides whether ChatGPT remembers things across chats; turning it off stops new memories but doesn’t delete old ones — you have to clear those yourself.
None of this changes your experience. Your history stays, nothing degrades. You just stop feeding the machine.
Never thumbs-up anything personal
Here’s the trap nobody mentions: feedback overrides your opt-out, on both platforms. Even with training off, a single thumbs up or thumbs down can send the entire conversation off to be used for training. OpenAI states this directly; Anthropic’s policy says the same, and adds that the whole related conversation gets stored for up to five years, de-linked from your user ID.
The logic: the toggle is you deciding not to share; the thumbs button is you deciding to share this one. So if Claude or ChatGPT nails a work problem, a health question or something about your family and you instinctively tap thumbs up to say thanks — you just handed over the conversation. Rate the boring stuff, not the personal stuff.
Know the toggle’s ceiling — and the past
The opt-out isn’t total. If an automated system flags your conversation for safety review, it can still be used regardless of your setting. Flagged content can be kept for around two years and reviewed by humans, and neither company publishes what triggers a flag or tells you when one’s raised.
It’s also not retroactive. You can’t un-train a model. If your chats were already inside a training run when you flipped the switch, they stay there — the same is true of deleting a chat. The opt-out is forward-looking on both platforms, so the sooner you flip it, the less there is to worry about.
And two things no setting touches. Legal orders: in the New York Times lawsuit, a court ordered OpenAI to preserve user chats — including deleted ones — and OpenAI was later ordered to hand over a sample of 20 million de-identified conversations. Accumulation: even with everything off, the pattern of what you ask still builds a picture that lives in a corporate account. That’s a cloud problem, not just an AI one.
For anything sensitive, use a private chat
The account toggle is your baseline; private modes are for the specific conversation you don’t want anywhere near a training pipeline. Claude: Incognito chat — the ghost icon, top right. It’s never used to improve Claude, even if your model-improvement setting is on, and it doesn’t go into your history. ChatGPT: Temporary Chat — the button at the top right of a new chat. No history, no memories, not used for training, and deleted within 30 days.
Use these for the medical question, the legal draft, the client contract, the salary negotiation. But be clear what “private” means here: not used for training and not saved to your history — not that the message never left your device. It did. It went to a server, and a copy sits in short-term retention for abuse monitoring. Think incognito browser window, not Signal.
No setting is your last line of defence — you are. Don’t paste passwords, full account numbers, or someone else’s private information (like medical records) into any chatbot. That stuff shouldn’t be in a cloud service at all, opt-out or not.