OpenAI appears to re-enable training on Codex chats, users report

A Hacker News post published on 2026-11-03 claims a Codex data-sharing setting was switched back on without users knowingly opting in. It matters because teams that use coding assistants for proprietary work depend on those controls remaining off.

According to Hacker News, the post by user052919 had reached 409 points and 131 comments at the time captured in the source. The poster said their company was cautious about leaking intellectual property through prompts and had therefore been careful to turn off data sharing in tools including Codex and Claude Code. The concern raised was not simply that a setting existed, but that, in the poster's account, it appeared to have changed without a deliberate action by the team.

What did the original poster report?

According to Hacker News, user052919 said coworkers had noticed the change the previous day. The poster said none of them knowingly opted back in, and asked whether the experience was isolated or more widespread. That framing matters: the source is a user report in a discussion thread, not a statement from OpenAI, and the article can only report what was claimed there.

The poster also added two edits. One pointed readers to an earlier Hacker News discussion that had noted a similar issue, and the other pointed to a separate OpenAI privacy page where, according to the poster, data sharing could also be turned off. Those edits suggest the poster was trying to give other users places to verify their own settings rather than asking readers to take the report on trust alone.

No response from OpenAI is included in the source text provided. Without such a response, it is not possible from this material to confirm why the setting changed, whether it affected a small number of accounts or a larger group, or whether any prompts were used for training as a result.

Where can users check the setting?

According to the original poster on Hacker News, the place to look is ChatGPT settings under data controls, for an option labelled "Improve the model for everyone". The poster presented that as the setting to inspect if someone wanted to see whether sharing had been re-enabled.

The poster also said data sharing could be turned off through an OpenAI privacy control page. For readers, the practical point is to check both places rather than assuming one toggle tells the whole story, especially where a product such as Codex is used alongside a chat interface.

For teams, a useful routine is to check the setting on a regular cadence, record the date and account checked, and repeat the check after product updates, workspace changes, or account migrations. That does not prove what happened in this case, but it creates a paper trail if a setting changes unexpectedly.

Is this an isolated report?

The source does not establish that it is isolated. According to Hacker News, the original poster linked to an earlier thread that had noted the issue before, which at least indicates the poster believed similar reports had circulated earlier. The earlier thread is referenced in the post, but its contents are not reproduced in the source text, so no further details can be drawn from it here.

Another commenter, identified on Hacker News as chriskanan, said their setting had also been re-enabled. That commenter said their history showed two previous requests to disable training that had been granted, followed by the setting being enabled again. The commenter described the situation as baffling and speculated about possible reasons, but that speculation is presented in the thread as a personal theory, not as confirmed information.

Taken together, the thread contains more than one user claiming a similar experience. It does not contain a denominator, a representative sample, or company confirmation, so it cannot support a conclusion about how widespread the issue was.

What wider concerns did commenters raise?

Much of the discussion moved beyond the specific setting into broader questions about retention, safety filtering, and trust. According to Hacker News, one commenter, AlanYx, argued that flagged conversations could be retained under certain conditions and that contractual protections might not apply in the way users expect, citing an Anthropic report and describing Azure-hosted filters and Microsoft contract terms. Those are claims made by a commenter in a forum thread, not findings established by the source, and they should be treated accordingly.

Other commenters raised regulatory and legal questions. AlanYx also referred to warnings from the US Federal Trade Commission about consent for user data handling and suggested that class action waivers could complicate private litigation. Again, these were assertions in discussion, offered without supporting documents in the source text beyond links mentioned in the comments.

Separate threads in the comments discussed whether inserting sensitive or taboo subjects into prompts could deter training use. Commenters disagreed, and some warned that such tactics could backfire by triggering additional scrutiny or retention. The Grafy News Desk view is that this kind of prompt-level workaround is not a substitute for checking account settings, using enterprise controls where available, and avoiding sending material that should not leave the organisation in the first place.

The discussion also included unrelated personal allegations and speculation that do not provide evidence about Codex settings. They are not relevant to determining whether a toggle changed, and they should not be treated as part of the factual record on the setting itself.

How can teams reduce exposure when using coding AI?

The most direct lesson from the Hacker News thread is procedural: treat data-sharing controls as something to verify, not something to set once and forget. Teams that handle sensitive code should assign an owner for AI tool settings, check controls after major updates, and keep a short record of when checks were performed and what was found.

In practice, that also means separating work by sensitivity. General questions, public examples, and throwaway snippets carry a different risk from proprietary algorithms, unreleased product plans, or customer data. A team can reserve its most sensitive material for environments and workflows it has reviewed, and use less sensitive tasks for experimentation.

From the Grafy News Desk, readers who want to explore the underlying trade-offs can use Grafy Chat to compare how different models respond to the same coding brief, use Grafy Code for AI-assisted coding work in a dedicated studio, or use Reader to read and question PDFs and papers about data governance and model policy before deciding what to send to any provider. Those tools do not answer whether OpenAI changed a setting in this case, but they can help a team think through model choice, prompt hygiene, and documentation before sensitive work is pasted into a chat window.

What the Hacker News thread does establish is that at least two users said they found the setting enabled again after previously disabling it. What it does not establish is a cause, a scope, or an OpenAI explanation. Readers should therefore treat the post as a prompt to check their own settings, not as proof of a broader policy change.

Frequently asked questions

Did the original poster say they knowingly opted back in?

No. According to Hacker News, user052919 said coworkers noticed the setting had been re-enabled and that none of them knowingly opted back in. The poster asked whether the issue was isolated or more widespread, and pointed readers to settings they could check themselves.

Where did the poster direct users to check the setting?

The poster directed users to ChatGPT settings under data controls, looking for "Improve the model for everyone". The poster also said a separate OpenAI privacy page allowed data sharing to be turned off, suggesting users check both locations.

Did another commenter report a similar experience?

Yes. According to Hacker News, commenter chriskanan said their setting had also been re-enabled. The commenter said their history showed two granted requests to disable training, followed by the setting being enabled again, and described the situation as baffling.

Source: Hacker News

---

Source: Hacker News — front page — OpenAI seems to be silently re-enabling model training on Codex chats