Whether a chat trains a model is a setting, not a given

Start here

Before pasting something into a chatbot, it helps to separate three things a service can do with your words: hold them in the current conversation, store them in your account history, and use them later to retrain the underlying model. Only the last one is what vendors mean by training, and it is controlled by a setting that differs by product and by whether the account is personal or business.

What the documents say

OpenAI's help page on data use, as retrieved on 16 September 2026, states that for individual services such as ChatGPT, the company 'may use your content to train our models' by default, and that a user opts out by turning off 'Improve the model for everyone' in Settings, or by selecting 'Do not train on my content' in its privacy portal. The same page states that business products, including ChatGPT Business, ChatGPT Enterprise and the API, are not trained on by default. Anthropic's equivalent page describes the opposite default for consumer plans: Claude Free, Pro and Max chats are used to improve Claude only if the person opts in, a conversation is flagged for safety review, or through an explicit program such as a trusted-tester scheme. Google's Gemini Apps Privacy Hub ties training to a setting called Gemini Apps Activity, but adds a separate channel: human reviewers 'review some of the data we collect,' and the page tells users not to 'enter confidential information that you wouldn't want a reviewer to see.' Three starting points: opt-out, opt-in, and a setting paired with a review process it does not fully cover.

Check this

Open the specific control each page names, not a general privacy menu: Settings then Data Controls in ChatGPT, the model-improvement setting in Claude, or Gemini Apps Activity at myactivity.google.com. Then ask what the toggle excludes. OpenAI's page notes that feedback such as a thumbs-up can still send a conversation into training even after opting out. Anthropic's page states that turning off model improvement stops future chats from being used, but data already inside a training run is not withdrawn. Google's human-review channel runs regardless of the training setting. A toggle changes what happens going forward, not what already happened.

What holds and what fails

The distinction holds cleanly between consumer and business accounts: an enterprise agreement changes the default by contract, not only by a setting a person could miss. It fails as a privacy guarantee: none of these settings stops a reviewer reading a flagged conversation, and none retroactively erases content already used. This is an editorial reading of three documents, not a claim that any vendor is misleading users; each page states its own scope plainly.

• Find the exact training control your product names, not a generic privacy page.
• Check whether your plan's default is opt-in or opt-out before assuming either.
• Look for what the page says about feedback, flagged content and past conversations.

A single switch answers a narrower question than most people assume. Reading which conversations it covers, and which it does not, is more useful than treating it as a one-time privacy decision.