If you buy through our links, we may earn a commission. It never affects our verdicts or scores — how that works. As an Amazon Associate I earn from qualifying purchases.

In the consumer app, yes, by default — and there is one switch that stops it. In the developer API, no, by default, and has not since 2023. Same company, opposite starting positions, which is most of why the internet cannot agree on the answer.

The switch, and its exact name

OpenAI’s help centre gives the path. On the web, signed in: profile icon, Settings, Data Controls, then turn off “Improve the model for everyone”.

That is the whole control. Two details from the same page matter more than the path.

First, it syncs: “Once you turn off model training, the setting applies to your entire account. It doesn’t matter which device you’re using.” You do not need to repeat it on your phone.

Second, and this is where people go wrong: your conversations still appear in your chat history. OpenAI’s wording is that they “won’t be used to train ChatGPT” — not that they stop existing. Turning off training is not deletion, and it was never meant to be.

There is also a control for signed-out users — the ? icon at the bottom right, then Settings — which is worth knowing if you use it without an account and assumed that was already private.

The API has the opposite default

If you reach OpenAI through code rather than the chat window, the position reverses. Its platform documentation states that “As of March 1, 2023, data sent to the OpenAI API is not used to train or improve OpenAI models” unless you explicitly opt in.

That is why a developer will tell you your data is not used for training and a ChatGPT user will tell you it is, and both are reporting their own product accurately.

The API is not a black hole either. The same page describes abuse monitoring logs that “may contain certain customer content, such as prompts and responses”, generated for all API usage and retained for up to 30 days unless law or harm-prevention requires longer.

Temporary Chat is not the same as private

This is the other thing people reach for, and it does a narrower job than the name suggests. Per OpenAI, Temporary Chats “are deleted from our systems after 30 days”, “aren’t used to train our models”, “may be reviewed only to monitor for abuse”, and “don’t get saved in your history and don’t create memories”.

Read the third clause. Not trained on is not the same as not read, and thirty days is not immediately. Temporary Chat is the right tool for keeping something out of your history and out of the model’s memory. It is not a private channel.

What each control actually does

If you want to…	Use	What it does not do
Stop chats improving the model	Turn off “Improve the model for everyone”	Does not delete anything, or hide history
Keep a chat out of history and memory	Temporary Chat	Still kept 30 days, still reviewable for abuse
Remove a conversation	Delete the chat	Does not retroactively undo past training
Build on OpenAI without training	Use the API	Abuse logs still retained up to 30 days

The row that catches people is the last column of the first: they turn the setting off, see every old conversation still sitting in the sidebar, and conclude it did not work. It did. It is just not a delete button.

What to actually do

Turn off “Improve the model for everyone” now if you want it off. It takes ten seconds, it applies account-wide, and nothing about the product gets worse — OpenAI does not gate features on it.
Do it before the conversation you care about, not after. The setting is not retroactive, and neither is deleting a chat afterwards.
Use Temporary Chat for the one-off you would rather not keep, understanding it is a thirty-day window and an abuse-review exception, not a vault.
Do not assume this transfers. Every other assistant sets its own defaults, and several have changed them recently. Check the one you actually use.
If it is genuinely sensitive, do not type it into someone else’s computer. The strongest version of this control is the text you never send — and if the reason is that you want a model with no third party in the loop at all, what it takes to run one locally is the honest comparison.