Millions of people type sensitive questions into ChatGPT every day — medical symptoms, salary numbers, drafts of emails they haven’t sent yet — without checking a single settings menu first. Most never realize that, by default, OpenAI can use those conversations to train the very model answering them, and that paying for a subscription does not automatically turn that off.
The setting responsible sits a few taps deep in the account menu, and it behaves differently depending on which type of account is doing the typing.
The Default Sits Under One Toggle
OpenAI calls the setting “Improve the model for everyone.” When it is switched on, conversations may be used to help train future versions of the underlying model. According to OpenAI’s own Data Controls FAQ, the toggle lives in Settings under Data Controls on the web version of ChatGPT, reachable by clicking the profile icon in the corner of the screen, and in a similarly named menu on mobile. Turning it off does not erase or hide anything already in a person’s chat history — those conversations remain visible — but new conversations stop being used for training once the switch is flipped.
Which Plans Are Covered — and Which Are Not
The training default is not universal across every OpenAI product. OpenAI’s explanation of how data improves its models draws a clear line between “services for individuals” and “services for businesses.” ChatGPT and Codex, used through an individual account, may have their content used for training by default — and that includes paid personal-tier subscriptions like ChatGPT Plus and ChatGPT Pro, not just the free version. Business-facing products sit on the opposite side of that line: ChatGPT Business, ChatGPT Enterprise and the API platform do not train on inputs or outputs by default, and organizations using those tiers must explicitly opt in if they want their data used at all. The practical result is that a person paying for a premium personal subscription is treated the same as a free user for training purposes, while a company paying for a business or enterprise seat gets the opposite default.
What Opting Out Actually Changes
Switching off “Improve the model for everyone” stops new conversations from feeding into training, but it does not change the everyday experience of using ChatGPT. Chat history keeps working the same way, memory features continue functioning, and nothing about response quality is degraded by the change, according to OpenAI. A separate feature, Temporary Chat, goes a step further: conversations started that way are deleted from OpenAI’s systems after 30 days, are never used for training, and do not appear in chat history or generate memories at all. For anyone who wants to ask a single sensitive question without leaving a lasting trace, that mode offers a cleaner option than toggling the main setting on and off repeatedly.
Why the Individual/Business Split Exists
The divide between individual and business defaults is not incidental — it reflects two different products built for two different kinds of risk. Business, Enterprise and API customers frequently pass proprietary code, internal documents and confidential business data through OpenAI’s systems, and a default that trained on that material without explicit consent would be a nonstarter for corporate legal and security teams. Individual ChatGPT accounts, including paid Plus and Pro tiers, are treated as a different category entirely: the subscription buys higher usage limits, access to more capable models and additional features, but it does not, by itself, change how a person’s conversations are handled on the back end. API customers can still opt in to sharing data — for instance, by submitting feedback inside OpenAI’s Playground tool — but that sharing has to be a deliberate choice rather than an inherited default, the reverse of how the setting works for individual ChatGPT users.
Feedback Buttons Can Reopen the Door
There is one notable exception built into the opt-out. Even after disabling model training, a person who clicks a thumbs-up or thumbs-down rating on a response is opting that specific exchange back in — OpenAI’s documentation states that the entire conversation associated with a piece of feedback may then be used to train its models. Support conversations carry a similar caveat: if training is enabled in account settings, exchanges with OpenAI’s support team may also be used to improve its services. Neither exception is hidden, but both run counter to the assumption that flipping one toggle permanently seals off every channel by which a conversation could end up in a training set.
How to Turn It Off on Web and Mobile
On the web, signed-in users click their profile icon, open Settings, select Data Controls, and switch off “Improve the model for everyone.” Signed-out visitors can reach a simplified version of the same toggle through the help icon in the bottom corner of the ChatGPT screen. On mobile, the path runs through the sidebar menu, the profile icon, and a Data Controls screen. The setting syncs across devices tied to the same account, so a change made on a phone applies the next time the same account is opened on a browser. OpenAI also offers a separate privacy portal where users can submit a broader “do not train on my content” request, and the company says it will continue honoring opt-outs submitted in the past, including ones made by contacting support directly rather than through the in-app toggle.
This article was produced with the assistance of AI and reviewed by an editor.
More from Morning Overview
