Notes on coding agents and the pages they leave behind, on moving between notes apps, and on the plain mechanics of keeping what you think. Written and published from a Sumibako vault.
Does ChatGPT, Codex, Claude or Gemini train on your chats? Where the switch is, and what it misses
Which AI trains on your chats: ChatGPT, Codex, Claude and Gemini, where each switch is, what turning it off costs, and what no switch covers.
Prashant Shaw
Founder, Sumibako
8 min read
On 3 November a developer told Hacker News that the setting which lets OpenAI train on their Codex chats had switched itself back on. The next day somebody else found that Gemini trains on a paying subscriber's chats unless they give up their chat history. In between, Postman's desktop app was caught installing itself into people's coding agents without asking. It was a good week for reading settings pages, so I read them.
This post is what ChatGPT, Codex, Claude and Gemini each do with your chats by default, where the switch is, what turning it off costs you, and the catches no switch covers. Everything was checked on 6 November 2026 against each company's own pages, with one exception I flag where it comes up.
🦜 The short answer
You use
Trains on your chats?
The switch
What turning it off costs
ChatGPT Free, Plus or Pro
Yes, until you turn it off
Settings, Data controls, Improve the model for everyone
Nothing you will notice
Codex on a personal ChatGPT plan
Follows the same switch
The same switch, plus a separate one in Codex for your environments
Claude Free, Pro, Max or Ultra, and Claude Code on those
Only if you allowed it when asked
Settings, Privacy, Help Improve our AI models
Nothing. Retention drops from five years to 30 days
Gemini on a personal Google account, paid or not
Yes, while Keep Activity is on
Turn Keep Activity off, or use temporary chats
Your chat history, and connected apps like Workspace
A business, team, enterprise or API plan from any of them
Not by default
Your admin
Not your problem, for once
The rest is the detail behind each row, because every row has a footnote somebody would rather you skipped.
♊ Gemini: paying is not opting out
The Hacker News post quoted Google's own pages, and they say what it said. From Google's Keep Activity page: "Gemini Apps activity, which includes data from apps you connect, is used to improve Google services for everyone, including by training generative AI models, if your Keep Activity setting is on."
There is no separate training switch. Keep Activity is the switch, and Keep Activity is also your history. Turn it off in Gemini Apps Activity and Google "stores your chats with your account for 72 hours", then they are gone. The privacy hub adds the second cost: "At this time, other Connected Apps like Google Workspace are unavailable if you turn off Keep Activity". So you can have privacy, or you can have Gemini read your Gmail, which is a sentence I did not expect to type.
The poster put it better than I can: "I learned this too late, thinking that having a paid subscription = not being turned into the product." Paying buys you the model. It does not buy you out of the dataset.
Then there is human review. Per the same hub, "A subset of chats are reviewed by human reviewers (including Google’s trained service providers) to help improve Google services", and "Reviewed chats are retained for up to three years." Deleting your activity does not reach those: reviewed chats "are not deleted when you delete your activity."
To be fair to Google, three things go the other way:
"Temporary chats are not used to train Google’s AI models." Use one for anything you would not put in an email to a stranger.
"Gemini does not train generative AI models directly on your Gmail inbox, Drive, Contacts, Calendar or other Workspace apps". Directly is doing some work in that sentence, but it is there.
"If you have a work or school Google Account, your use of Gemini Apps may be subject to different data handling terms." Work accounts are a different deal.
🤖 ChatGPT and Codex: one switch, and reports that it moves
OpenAI's switch is under Settings, Data controls (chatgpt.com/settings/data-controls), and it is called "Improve the model for everyone". On a personal plan it covers Codex too. OpenAI's data controls article says the same setting applies to your Codex tasks, and that a do-not-train request in the Privacy Portal does the same job; either one is enough. The same article describes one more setting in Codex itself, which decides whether context from your Codex environments can be used for training; I could not confirm its exact label, so look for it under Codex's own settings.
Temporary Chat is OpenAI's version of Gemini's temporary chats: not used to train, kept for up to 30 days for safety, and turned into an ordinary chat if you save it. Business, Enterprise and Edu workspaces are out by default; OpenAI's own Work admin FAQ promises "no training on business data by default".
📝NOTE
help.openai.com refuses scripts, so I could not keep a copy of OpenAI's help articles to check quotes against. The OpenAI paragraphs above are paraphrased. Follow the links for OpenAI's exact words.
Now the catch. On 10 September a Hacker News post said: "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." On 3 November another, from a team that had turned it off on purpose: "But yesterday my coworkers and I noticed that the setting had been silently reenabled in Codex. None of us knowingly opted back in." Other people in the same thread found theirs still off, and OpenAI has said nothing I could find, so call it two reports and not a confirmed bug.
My own approach is the belt and braces one. Turn the switch off and file the Privacy Portal request too. OpenAI says either is enough, but only one of them is a checkbox.
🧠 Claude and Claude Code: you were asked, once
Anthropic changed its consumer terms in August 2025, and the change was to ask. "These updates apply to users on our Claude Free, Pro, and Max plans, including when they use Claude Code from accounts associated with those plans." Existing users had to choose, and new ones "can pick your setting for model training during the signup process." Anthropic is the only company here that asked first. It is a low bar, and the others found a way to walk under it.
The Claude Code data page brings the list up to date: "We will train new models using data from Free, Pro, Max, and Ultra accounts when this setting is on (including when you use Claude Code from these accounts)." The switch is in Settings, Privacy, under the heading Help Improve our AI models, which Anthropic's privacy article walks through for desktop and phone.
What the choice changes is retention as much as training. "We are also extending data retention to five years, if you allow us to use your data for model training." Say no, and "you’ll continue with our existing 30-day data retention period."
Two Claude Code habits are worth knowing about:
/feedback, /bug and /share send the transcript, and those transcripts "are retained for 5 years".
The "How is Claude doing this session?" prompt only records a rating, unless you then say Yes to sharing the session. That uploads the transcript, though Anthropic says those transcripts "cannot be used to train our AI models".
Team, Enterprise, the API and the cloud platforms are under commercial terms, where "Anthropic does not train generative models using code or prompts sent to Claude Code under commercial terms".
🧾 What no switch covers
1️⃣ The thumbs. Rating a reply is an opt-in of its own. OpenAI's data controls article says a thumbs up or down can put the whole conversation into training. Anthropic: "The new five-year retention period will also apply to feedback you submit to us about Claude’s responses to prompts." Google's promise only holds "If Keep Activity is off and you don’t submit feedback". The thumbs-down on the chat with your API key in it is the one click that undoes the setting.
2️⃣ Safety review. Anthropic: "If our safety classifiers flag your conversations, they may still be used to improve our internal trust and safety models". Google: "Even if your Keep Activity setting is off or you use temporary chats, Google still uses your chats to respond to you and help protect Google, our users, and the public, including with help from human reviewers."
3️⃣ The past. Every switch works forwards. Anthropic is the plainest about it: "Your data will still be included in model training that has already started and in models that have already been trained".
4️⃣ Your other tools, which brings me to Postman.
🔌 Meanwhile, in your config files
On 25 October a Postman user opened an issue titled "[SEVERE] Desktop auto-installs Codex/Claude plugins; installer attempts to trust its own startup hook". They had opened Postman on a Mac and found a Postman plugin, a startup hook and a remote MCP connection set up in Codex and Claude Code. It reached the top of Hacker News on 3 November. Postman answered on 5 November that "we will be re-designing the experience with the right level of control", and: "In the meantime, we are reverting this change."
It is not a training story, but the lesson is the same. A setting you chose stays chosen only until something else edits it. After installing any developer tool, look at ~/.codex/config.toml and at the plugins Claude Code lists under /plugin.
2️⃣ Codex: in Codex's own settings, turn off the switch for your environments' context, the one OpenAI's data controls article describes.
3️⃣ Claude: Privacy settings, Help Improve our AI models off, if you said yes in 2025 and have changed your mind.
4️⃣ Gemini: Keep Activity off, or temporary chats for anything sensitive.
5️⃣ Do not rate a chat you would not send to the company.
6️⃣ Put a reminder in your calendar to look again, because one of these has been reported flipping back twice.
🗄️ Where the part worth keeping should live
Turn Gemini's history off and your chats are gone in 72 hours. Leave everything on and they live in somebody's training set for five years. Either way, a chat is the wrong home for the thing you will want next month: the plan, the decision, the command that finally worked.
That is the reason I built Sumibako. Your coding agent files those pieces as pages you own (here is how), and Claude can read and write them too. Our privacy page says we "do not use your notes to train AI models". There is no model here to train, which makes it an easy promise to keep, and I would still rather say it out loud than leave you to find the switch.
Security news, 25 October to 2 November 2026: a backdoored Claude Code extension, an unpatched Linux bug and Trigger.dev
A tampered Claude Code extension in Cursor's marketplace, CVE-2026-98432 still unpatched on Ubuntu, and Trigger.dev: what happened and what to check.
