Users report OpenAI’s training opt-out turning back on in Codex. How to check yours

What’s being reported

On November 3, 2026, a post on Hacker News reached the front page under the title “OpenAI seems to be silently re-enabling model training on Codex chats.” Its author wrote that their company turns off data sharing as a rule, and that “yesterday my coworkers and I noticed that the setting had been silently reenabled in Codex. None of us knowingly opted back in.” The setting is Improve the model for everyone. The post doesn’t name their plan.

The claim isn’t new. A thread from September 10, with 487 points, said much the same: “I’ve reset this more than once and the last time I made a careful note of when I did it and to my surprise I found it re-enabled when I checked just now.” A December 2024 post on OpenAI’s own forum said it “became enabled again without my awareness,” possibly when the iOS app updated.

The evidence is thin in both directions. In the new thread, one commenter writes “can confirm,” and another says “We’ve noticed this occasionally.” Two others say “Unchanged for me” and “mine wasn’t reenabled.” In September, several people said theirs had stayed off for months or years. We read no report with evidence anyone else can check, such as a record of when the switch changed. Treat it as unconfirmed, and as cheap to check.

What OpenAI’s pages do say

Whether OpenAI may train on your Codex work depends on how you sign in:

How you use Codex	Trained on by default?	What controls it
ChatGPT Free, Plus or Pro	Yes: “data sharing is enabled by default”	Improve the model for everyone, or Do not train on my content in the Privacy Portal
ChatGPT Business, Enterprise or Edu	No	Your organization’s agreement with OpenAI
An API key	No	Your API organization’s data-sharing settings, where an owner can opt in

Five details are easy to miss:

• The ChatGPT setting does cover Codex. “Your ChatGPT training data controls apply to content processed through Codex, including screenshots taken by Computer Use,” says OpenAI’s Codex article.
• One opt-out is enough, OpenAI says. “Either option is sufficient for ChatGPT conversations and Codex tasks. You don’t need to do both,” says its training page. In September, OpenAI’s Ted Sanders wrote “If you opt out on either location, we’ll respect it,” adding “I work at OpenAI, but not on privacy. I am not their spokesperson.” He didn’t address the setting changing.
• Codex Cloud has its own switch. Include environments, in Codex settings, controls “whether additional context from your Codex environments can be used to train OpenAI models. Changing your ChatGPT setting or opting out through the Privacy Portal does not change that setting.” OpenAI’s pages don’t state its default.
• Feedback overrides the opt-out. On a personal plan, if you select thumbs up or thumbs down, “the entire conversation associated with that feedback may be used to train OpenAI models.” Codex’s /feedback sends “logs to the Codex maintainers,” and OpenAI’s pages don’t say how those are treated, so don’t send it from a session you wouldn’t share.
• Daybreak is an exception. “During Daybreak sessions on your personal account, we may use content from your interactions to improve our models even when ‘Improve the model for everyone’ is off.” An email to Daybreak Blue users, as posted on Hacker News, dates that from November 29.

What OpenAI hasn’t said

• Whether it has happened. We found no statement from OpenAI on these reports. Its release notes, whose latest entry is October 30, don’t mention them, and no one has identified themselves as speaking for OpenAI in the new thread.
• What could cause it. Commenters offer guesses: a lapsed plan (one says it came on “when the credit card failed to renew and it switched back to free”), an app update, accepted new terms, or a change that never saved. None is confirmed.
• What happens to work sent while it was on. The page promises only that “we won’t use your new conversations” after you opt out. Turning training off doesn’t delete chats, and OpenAI’s pages don’t say whether deleting one removes it from training. In another case in September, Sanders wrote that for Codex prompts sent when “training was not opted out, there’s a chance they made their way into our training pipeline in some form.”
• Whether you’d be told. OpenAI’s pages don’t describe any notice when this setting changes, or say what happens if it and the portal disagree.

How to check yours

1. Open the setting. In ChatGPT, go to Settings > Data controls. Turn off Improve the model for everyone, then select Done. It applies to your account on the web and on your phone.
2. Reload and look again. That rules out a change that never saved, one of the guesses above.
3. Add the portal request. Select Do not train on my content in the Privacy Portal. OpenAI says you don’t need both, but it takes a minute, and your choice is then recorded twice.
4. Check Codex Cloud, if you use it. Look at Include environments in Codex settings.
5. Look again now and then, especially after a plan change, a failed payment or an app update. Codealong, a free, open-source app for Mac and Linux from one of the thread’s commenters, marks any account whose provider may train on what you send.

There may also be a quick tell. An open Codex issue and a forum report say the thumbs buttons in ChatGPT on the web disappear when the setting is off and return when it’s on, though OpenAI’s page says you can still give feedback after opting out. That’s a user report, not a documented signal, so use it as a prompt to check, not as proof.

If a switch isn’t good enough

If your company promises customers that their code isn’t trained on, a setting on each person’s personal account is a weak thing to rest that on. In order of how much they change:

1. Turn on Advanced Account Security. OpenAI says “Conversations are not used to train OpenAI models while Advanced Account Security is enabled.” The page says conversations and doesn’t name Codex. It also needs passkeys or security keys, and turns off password sign-in, email and SMS codes, and email recovery.
2. Sign in to Codex with an API key. The desktop app, CLI and IDE extension support it. You pay by the token, usage “follows your API organization’s retention and data-sharing settings,” and Codex Cloud still requires ChatGPT sign-in. Check that your organization’s owner hasn’t opted in to sharing.
3. Move work to a Business or Enterprise workspace. “By default, OpenAI does not use any inputs or outputs from our products for business users, including ChatGPT Business, ChatGPT Enterprise, and the API, to improve our models.” That’s the workspace’s default, not a switch on each person’s account.
4. Run a local model. Nothing goes to OpenAI at all, at the cost of the frontier models. See our guide to local coding agents and Flint, which runs on a laptop.

How other agents compare

The same question applies to the others. Their personal plans also leave this to you, as of November 3, 2026:

Agent, on a personal plan	Trained on?	Where to turn it off
Claude Code on Free, Pro, Max or Ultra	“when this setting is on”	claude.ai/settings/data-privacy-controls
GitHub Copilot on Free, Pro, Pro+ or Max	“may use your interactions” unless you opt out	Copilot settings: “Allow GitHub to use my data for AI model training”
Cursor	“never” with Privacy Mode on; with it off, Cursor “may”	Cursor Settings > General > Privacy Mode
Antigravity	“We use Interactions to evaluate, develop, and improve” Google’s products and machine learning	The Settings panel

Anthropic keeps your Claude Code data for 5 years if you allow training and 30 days if you don’t. Claude’s Team and Enterprise plans and API, and Copilot Business and Enterprise, aren’t trained on by default. Google’s terms ask you not to “submit confidential information or any data you wouldn’t want a reviewer to see.” One commenter in the September thread said the Claude setting had come back on too, which is also unconfirmed.

Where Enjoy fits

Enjoy has no AI model of its own and never sees your conversations. It hands them straight to Codex on your computer, and what Codex works on goes to OpenAI under your account settings there (what AI providers see). So everything above applies in Enjoy, and Enjoy doesn’t watch or change this setting. Two things it does can help:

• Keep work on the work account. Enjoy can give each project its own account, so a company project stays on its Business seat and a side project on your personal plan.
• Keep some work on your computer. OpenCode with a local model runs in the same conversation, and nothing leaves your computer.

We’ll update this memo if OpenAI explains the reports or changes how the setting works.