Open any of the major AI chatbots and begin typing. Within seconds, your words are being processed, stored, and — unless you have taken deliberate steps to prevent it — potentially fed back into the very model you are speaking with. This is not a hypothetical concern buried in a terms-of-service document. It is the default behaviour of ChatGPT, Claude, Gemini, Copilot, and Grok alike, and it applies to the vast majority of users who have never thought to look for an opt-out.
The mechanics are straightforward enough. Each platform collects consumer conversations as training material, using them to refine future iterations of their models. The opt-out options exist, and on most platforms they take under a minute to activate. Crucially, disabling training does not affect personalisation or memory features, which are governed by separate settings entirely. The practical cost to the user is negligible; the benefit, in terms of data hygiene, is real.
There is, however, one significant exception — and it is a telling one.
The Platforms That Give You a Choice
OpenAI’s toggle is labelled Improve the model for everyone and sits under Settings, then Data Controls. All personal accounts are opted in by default. Switching it off applies across both web and mobile, and chat history continues to function normally. For conversations of a particularly sensitive nature, ChatGPT offers a Temporary Chats feature: these are never used for training and are purged by OpenAI after thirty days.
Anthropic’s approach to Claude is worth noting for its recent evolution. The company did not use consumer conversations for model training until August of last year, when it rewrote Claude’s terms and gave Free, Pro, and Max users a deadline of 28 September to accept or decline. The toggle — labelled Help improve our AI models — is found under Settings, then Privacy. Leaving it enabled means Anthropic retains conversations for up to five years; switching it off returns accounts to a thirty-day deletion window. Whatever data Anthropic trained on before the opt-out, however, it keeps.
Google’s equivalent control, recently renamed Keep Activity, is enabled by default on Gemini. When active, conversations may be used to improve Google’s models, and a subset of chats are reviewed by human annotators — with any reviewed conversation retained for up to three years, even after a user deletes their history. Disabling the setting stops both training use and the saving of interactions to Gemini’s chat history, though Google retains each conversation for seventy-two hours to operate the service, using the final twenty-four hours for contextual responses.
Microsoft and X: Variations on the Same Theme
Microsoft takes a slightly more granular approach with Copilot, offering two distinct controls: Training on conversation activity, which is on by default, and Training on voice conversations, which is off by default. Both are accessible via the profile icon, then Settings and Privacy. Opting out excludes future conversations from model training, though it does not prevent Microsoft from using chat data for advertising or broader product improvements — a distinction worth bearing in mind.
X, the platform formerly known as Twitter, enrols every account in Grok training by default, covering both public posts and direct conversations with the chatbot. The relevant toggle lives under Settings and privacy, then Privacy and safety, then Grok & Third-Party Collaborators. Unchecking the data-sharing option withdraws posts and Grok interactions from future training and fine-tuning runs.
Meta: Where the Opt-Out Does Not Exist
Meta occupies a category of its own. American users have no mechanism to opt out of their data being used for model training — none whatsoever. Users in the United Kingdom and European Union benefit from the protections afforded by the General Data Protection Regulation, which allows them to object to their information being used for AI training through a form in Meta’s privacy centre. The divergence is a pointed illustration of what robust data protection law can achieve when it is actually enforced.
The situation has grown more pointed since December of last year, when Meta began using text and voice conversations with Meta AI to target advertising across Facebook, Instagram, WhatsApp, and Messenger — again, without any opt-out for US users. For those in America, the only practical recourse is to decline to engage with Meta AI altogether. Given the company’s recent launch of its new AI agent, Muse, that advice may prove harder to follow as the product is pushed more aggressively across Meta’s properties.
The broader picture is one of asymmetric information and asymmetric power. The opt-outs exist, but they are not advertised. The labels — “Help improve our AI models,” “Improve the model for everyone” — are framed as acts of civic generosity rather than commercial data extraction. For users who do locate the relevant settings, the process is quick and the cost minimal. For those who do not, the default remains firmly in the platform’s favour. That is not an accident of design.