Do AI Chat Apps Train on Your Data? Settings for 9 Apps

Of the 29 plan rows collected here across nine AI chat apps, ten use your conversations for model training by default, twelve do not, two make you choose at signup, and five do not say. The ten that train are all consumer or signed-out tiers. The twelve that do not include every business or enterprise tier the vendors sell except Mistral Vibe Teams, where Mistral does not state the default. That split, consumer trains and business does not, is the single pattern in the data, and the exceptions to it are where the detail matters.

This is a reference table, not an argument. Each row quotes the vendor's own wording for the default, names the setting and its menu path as the vendor documents it, and records the policy page and its date. Where a vendor describes training and an opt-out without saying which way the switch starts, the row is marked unclear, and it stays unclear until the vendor says otherwise.

The scope is chat products that a person or a team signs up for. Developer APIs and agent platforms, where the defaults and the retention rules are different, are covered in our API and agent-platform retention census. For how to use customer questions in AI-assisted work without exposing them, see our note on customer questions and AI content.

Read down the default column and the business tiers are monotonous: ChatGPT Business and Enterprise, Claude Team and Enterprise, Gemini in Workspace, Copilot for work, Grok Business, Perplexity Enterprise, Mistral Vibe Enterprise. Each vendor states no training by default, in words close enough to each other that the market has clearly converged. OpenAI's phrasing is typical: "By default, we don't use inputs or outputs from ChatGPT Business, ChatGPT Enterprise, ChatGPT Edu, or our API to improve our models."

The consumer tiers are nearly as uniform in the other direction. OpenAI says that for its services for individuals "we may use your content to train our models". Google says that for adults "Keep Activity is on by default" and that activity is used for "training generative AI models". Perplexity says its AI Data Retention setting "is enabled by default" for Free, Pro and Max. Mistral says input and output "are used by default to train" its models. Meta says it "uses your interactions with AIs to improve AI at Meta" and documents no training opt-out on the pages we read.

The exceptions are the useful part. Anthropic requires a choice rather than setting a default, and does not publish which option is pre-selected. Microsoft's new Copilot app, released August 18, 2026, states that prompts and responses "aren't used to train foundation models" and its privacy-controls page carries no training toggle, while the FAQ for the older app, still live and marked as applying only to that version, describes training unless the user opts out. And four plans, Grok consumer, Grok on X, Mistral Vibe Teams and DeepSeek, describe both training and an opt-out without stating the starting position.

The default column uses four values. Yes means the vendor states that training is on or that content is used by default. No means the vendor states no training by default. User choice means the vendor requires an explicit choice and does not publish a pre-selected position. Unclear means the vendor describes training and an opt-out but does not state the default. The quoted text in the evidence column is what each value rests on. Retention and temporary-chat behaviour are in the next section's table, one row per app, to keep this one readable.

Two reading notes. The ChatGPT rows cover the consumer product; OpenAI's sentence groups "ChatGPT and Codex" together as services for individuals, so the same default applies to Codex on a personal plan. The Gemini paid-plan row inherits the personal-account rule because Google's privacy hub mentions collecting subscription information but states no separate training rule for paid plans; we neither confirmed nor ruled out an exception.

Training and retention are separate questions, and vendors answer them on different pages. This table collapses each app to one row: the retention statements that apply to its consumer and business tiers, how temporary or incognito chats are handled, and the policy pages with their dates. Where a vendor publishes no period, the row says so.

Five of the nine vendors state that rating a response is a separate consent from the training setting. OpenAI is explicit: "If you proactively choose to submit feedback, such as by selecting thumbs up or thumbs down on a response, the entire conversation associated with that feedback may be used to improve our models, even if you've opted out." Note the scope: the entire conversation, not the rated message.

xAI's FAQ says that "even if you opt out of model training, when you subsequently decide to voluntarily provide feedback, that feedback may be used for training purposes". Anthropic's privacy centre says a rated conversation is stored for up to five years and may be used to "train our AI models"; Mistral applies the feedback rule to every subscription plan; Google says chats made with Keep Activity off "won't be used to train our AI models, unless you choose to send Google feedback", and keeps human-reviewed chats, with related feedback, for up to three years. On Claude Team and Enterprise, an owner can switch the rating buttons off for the whole organisation, which is the one place in the data where the override itself has a switch.

The example that catches teams: a staff member opts out, then rates a few answers while working through a document that contains customer details. Under OpenAI's wording the whole conversation is now eligible. The fix is a policy sentence, not a setting: do not rate responses in conversations that contain anything you would not paste into a public form.

None of the nine vendors' pages we read offers to remove a conversation from a model it has already trained, and two say so directly. Perplexity's help centre: "Previously collected training data cannot be deleted or removed." Anthropic's privacy centre says data already used stays "in model training runs that are already in progress, or in models that have been trained". The same Perplexity help page adds that opt-outs "only apply to data collected after the opt-out date".

That is not a criticism of the vendors; it is how training works. It does change the order of operations for anyone setting up an account. The switch is set before the first sensitive conversation, not after a review notices the default. For a team that has been on a consumer plan for months, the realistic step is to change the setting now, delete history where the vendor allows it, and accept that past conversations are past.

ChatGPT, Claude, Gemini and Grok each say their temporary, incognito or private mode is not used for training (xAI adds "where available"). All four also publish a retention period for those chats, and it is not zero. The list below is the vendors' own figures. Perplexity is the outlier in the other direction, stating that incognito searches are never stored.

The failure case is treating an ephemeral mode as a substitute for the account setting. A temporary chat is a per-conversation choice that the user has to remember to make, and in ChatGPT's case, saving it afterwards applies the account's training setting to it. The account setting, or the plan, is the control that holds.

For a business, the table reduces to one decision. Every vendor that sells a business tier states no training by default on it (on Mistral Vibe, for Enterprise; the Teams default is not stated), and most add admin-set retention. Meta AI and DeepSeek have no business chat plan on the pages we read, which is itself an answer for work use. The router below is our reading of the vendors' statements; it is not legal advice, and a regulated business should check the contract terms, not the help centre.

One more distinction the pages make. Not training on your data is not the same as not retaining it, and not retaining it is not the same as not reviewing it. Google's Workspace page states no training without permission and, separately, retention of up to 36 months as admins set it. Microsoft's work page states no foundation-model training and, separately, that prompts and responses are logged and auditable by admins. A vendor-risk review should check all three, which is the structure of our AI vendor risk assessment template.

A table like this is only as good as its gaps are honest. Nine items could not be settled from the vendors' pages, and each is recorded here rather than resolved by guesswork.

If your team needs these controls set and documented across the apps it actually uses, with the vendor pages recorded as evidence, that is part of what our AI transformation engagements cover.