Your ChatGPT training setting may have quietly switched itself back on. That’s what a wave of developers reported this week, and it’s the second time this year the same thing happened. If you turned off “Improve the model for everyone” at any point, it takes about two minutes to make sure it’s still off.
What actually happened
On November 3, a post hit the front page of Hacker News claiming OpenAI is silently re-enabling model training on Codex chats. It pulled in over 460 upvotes and more than 140 comments in a day. The person who posted it said they and their coworkers noticed the toggle flipped back on without any notification.
Then the comments filled up with similar stories. One developer wrote that their ChatGPT training setting was re-enabled even though their account history showed two earlier opt-out requests that had been granted. Another said they spotted it while digging through Codex settings for something unrelated. One commenter who builds monitoring tools at their company called it a serious problem, because their customers were told none of their data gets trained on.
So far, OpenAI hasn’t publicly responded. No statement, no bug report acknowledgment, nothing official as of November 4.
Here’s the part that gives this story legs: it happened before. Back in September, a “Tell HN” thread claimed the same toggle kept re-enabling itself, and that one also hit the front page with nearly 500 points. So either there’s a bug that never really got fixed, or something changed on OpenAI’s side. Nobody outside the company knows which yet, and I’m not going to pretend otherwise.
What the ChatGPT training setting actually does
The toggle in question is called “Improve the model for everyone.” You’ll find it under Settings, then Data controls, at chatgpt.com. When it’s on, OpenAI can use your conversations to train its models. That includes your chats, and on personal plans it also covers tasks you run through Codex, OpenAI’s coding agent.
A few facts worth knowing, straight from OpenAI’s own help center:
- On personal plans (Free, Go, Plus, and Pro), data sharing is enabled by default. If you never touched this setting, your chats are already in the training pool.
- Turning it off keeps your chat history intact. Nothing gets deleted, your conversations just stop feeding future models.
- Business, Enterprise, Edu, and API accounts don’t get used for training by default, so this mainly affects personal users.
For most people reading this, that first bullet is the whole story. The default is on, and plenty of users have no idea.
The two-minute check
This takes longer to read than to do. Open chatgpt.com, click your profile icon, open Settings, and select Data controls. Look for “Improve the model for everyone,” which is the ChatGPT training setting this whole story is about. If it’s off, you’re done. If it’s on and you expected it to be off, flip it back.
On mobile, tap your profile icon in the sidebar, then Data controls, then the same toggle.
There’s a second route too. OpenAI runs a Privacy Portal at privacy.openai.com with a “Do not train on my content” option. Either one is enough. You don’t need to do both.
Two things to keep in mind while you’re in there:
- The opt-out applies to new conversations going forward. It doesn’t reach back and pull your old chats out of anything.
- If you have a workplace account through ChatGPT Business or Enterprise, workspace rules govern training and the default there is already no-training. This check is for your personal account.
The Codex setting people keep missing
Now the part that trips up even careful users. Codex has its own separate switch called “Include environments,” tucked away in Codex settings. It controls whether extra context from your coding environments can be used for training. And here’s the catch: the ChatGPT toggle does not touch it. Turning off “Improve the model for everyone” leaves “Include environments” exactly where it was.
This matters because some of the developers in the Hacker News thread may have been looking at Codex settings, not the main ChatGPT toggle. OpenAI’s own docs confirm the two controls are independent, which means a complete opt-out means checking both places.
One theory floating around the thread is that at least some of these “it re-enabled itself” reports are actually people who turned off the ChatGPT setting months ago, assumed it covered Codex, and only just now noticed the Codex-side switch sitting at its default. That would explain some of the reports without any silent flipping. But it can’t explain all of them, because the original poster insists the main ChatGPT setting itself was back on.
Either way, the practical takeaway is identical. Check both switches. Here’s where everything lives:
| Setting | Where it lives | What it controls | Changed by the ChatGPT toggle? |
|---|---|---|---|
| Improve the model for everyone | chatgpt.com, Settings, Data controls | Your chats and Codex tasks on personal plans | It is the toggle itself |
| Do not train on my content | privacy.openai.com (Privacy Portal) | Same coverage as the main toggle | Alternative route, either one works |
| Include environments | Codex settings (chatgpt.com/codex) | Context from your coding environments | No, it’s independent |
| Temporary Chat | Chat icon in any conversation | Not used for training, deleted after 30 days | Not a setting, a chat mode |
The fine print nobody mentions
OpenAI’s help pages hide a few details that change how you should use the opt-out.
Feedback overrides the opt-out. If you give a response a thumbs up or thumbs down, the entire conversation attached to that rating becomes eligible for training, even with the ChatGPT training setting turned off. OpenAI documents this openly. So if a chat contains anything you wouldn’t want in a training set, skip the rating buttons on it entirely.
Temporary Chats are the exception. They’re not used for training, they don’t create memories, and they’re deleted from OpenAI’s systems after 30 days. For one-off questions about anything sensitive, that’s the mode to use.
Voice has its own toggles. There are separate switches for audio and video recordings under Data controls. If you use voice mode and care about this stuff, glance at those while you’re in the menu.
And if you opted out long ago through an old support request or privacy form, OpenAI says those requests are still honored. Which is exactly why this week’s reports are strange: some affected users had granted opt-outs on record that got overridden anyway.
Does this affect the blog tools you already use?
If you just use ChatGPT for drafting emails, brainstorming, and writing help, the training toggle is a privacy preference more than a crisis. Your texts go into future models unless you say no. Some people genuinely don’t mind that trade for a free tool.
It gets real when you paste in things with consequences: client work, unreleased product details, code from a private repo, financial info, anything covered by an NDA. That’s the material that shouldn’t sit in a training corpus, and it’s exactly the material Codex users handle daily. This is also why the enterprise angle got attention; companies make contractual promises about customer data, and a toggle that flips itself undermines those promises.
For a broader privacy setup, we’ve covered related ground before. There’s a guide on hiding conversations with ChatGPT’s locked chats PIN feature, a breakdown of how ChatGPT ad targeting uses your prompts, and a walkthrough for stopping Google from using your Gmail for ads. Each one takes minutes to apply.
Takeaway
Spend two minutes on this today: open Settings, then Data controls, and confirm your ChatGPT training setting is actually off. If you use Codex, open the Codex settings too and check “Include environments.” Then note the date you checked, because if OpenAI’s toggle really is flipping itself back on, you’ll want proof of what your setting was supposed to be.