What happened with Codex’s training opt-out?
A November 3 Hacker News post says the author and coworkers discovered that “Improve the model for everyone” had been re-enabled, despite previously disabling data sharing. The author says none knowingly opted back in and asks whether the issue is widespread.
The post is a firsthand allegation, not an OpenAI announcement. Replies include both similar reports and users saying their setting remains off. An earlier HN thread describes a similar complaint; neither thread establishes the cause, affected account types, or whether any content was actually used for training.
We have not independently reproduced the reported reset. We found no OpenAI confirmation or explanation in the official sources checked for this article. An enabled toggle alone does not prove that a particular repository entered a training dataset.
Which Codex data controls should you check?
OpenAI’s data-controls documentation distinguishes conversation sharing from a separate Codex environment setting. Its model-improvement policy says either the ChatGPT opt-out or the Privacy Portal opt-out is sufficient for ChatGPT conversations and Codex tasks; neither changes Include environments.
| Control or route | What OpenAI documents | What to verify |
|---|---|---|
| Improve the model for everyone | Controls eligible conversation and task training | Your selected account and workspace |
| Codex: Include environments | Separate permission for training on full environments | The setting in Codex itself |
| Privacy Portal opt-out | Alternative opt-out for conversations and Codex tasks | Does not change Include environments |
| Business, Enterprise and Edu workspaces | Inputs and outputs are not used for training by default | You are actually in the managed workspace |
| API usage | Inputs and outputs are not used for training by default | Organization data-sharing choices |
Policy sources checked November 3, 2026. These are documented rules, not verification of an individual account’s behavior.
What should you do now?
- Confirm your account and workspace. A personal login used for company work is not automatically covered by a managed workspace’s policy.
- Open ChatGPT → Settings → Data controls. Check Improve the model for everyone and switch it off if you do not want eligible content used for training. OpenAI says the signed-in setting applies across web and mobile.
- Open Codex settings. Review Include environments separately. Do not assume the ChatGPT switch or a Privacy Portal request changes it.
- If a setting appears to have changed, preserve evidence. Record the time, account/workspace, client version and redacted screenshots, then contact OpenAI Support and your workspace administrator. Exclude code, credentials and customer information from the report.
- Pause sensitive submissions until your team’s policy is satisfied. For API workflows, ask the organization owner to review data-sharing settings. Do not assume changing a setting now reverses earlier processing.
Two qualifications matter. OpenAI says voluntarily submitting response feedback on individual plans may allow the associated conversation to be used for training even after opting out. Its data-controls page also documents a separate personal-account Trusted Access/Daybreak exception; users of that program should read its terms rather than assume the general toggle covers every session.
Training opt-out is not zero data retention, local-only inference, or deletion. Turning off training does not delete saved chats. Review retention, access and connected-service controls separately.
Does this change what you should pay for Codex?
This alert establishes no new rate card, model release, or subscription fee change. There is no verified before/after price to report, and buying a more expensive personal subscription is not a substitute for checking its data terms.
For teams, compare the required workspace protections, administrator controls and operational fit before changing products. OpenAI’s documented business/API training default differs from its individual-service policy, but that is not proof a paid upgrade is necessary for every developer—or that the reported issue affected managed workspaces.
Use our ChatGPT subscription tracker for maintained plan details and OpenAI pricing for API rates. If you are evaluating another provider, consult Anthropic pricing and verify that provider’s data policy separately. The token calculator and local, API and subscription cost guide help estimate a different workflow’s usage costs; neither establishes privacy equivalence.
FAQ
Did OpenAI confirm that Codex training was silently re-enabled?
Not in the sources checked for this article. The trigger is a user report, with mixed replies, not a confirmed platform-wide policy change.
Is turning off Improve the model for everyone enough?
OpenAI says that setting or the Privacy Portal opt-out covers ChatGPT conversations and Codex tasks. Codex’s separate Include environments setting must be reviewed independently, and documented exceptions still matter.
Does opting out remove code already sent to Codex?
No: OpenAI describes the opt-out in terms of new conversations, and turning off training does not delete saved chats. The HN report does not establish whether previously submitted code was used for training.