Does AI Train on Your Conversations? Vendor by Vendor

TL;DR: Most AI assistants use your chats for training unless you turn it off. Of six policies we could read on 18 August 2026, only Anthropic's Claude requires switching training on, so the setting has to be checked once per vendor you pay. Perspective AI's policy commits to not repurposing conversation content, and reaching several labs through one subscription means one policy to read rather than five.

This page answers one question with citations rather than vibes: which AI companies use your conversations to train models, and what the switch is called. Every row below was read from the vendor's own published policy on 18 August 2026, and the URL is recorded so you can check it against whatever the document says on the day you read this. Where a vendor's policy could not be retrieved, it is marked as such rather than filled in from memory. It belongs in our comparisons section because the differences between these vendors are larger than the marketing suggests. Perspective AI is in the table too, on the same terms as everyone else. One practical consequence of the table worth naming up front: every vendor you subscribe to is another policy and another toggle, so replacing three plans with one subscription cuts the number of privacy settings you have to keep checking from three to one.

Does AI Train On Your Conversations?

At most consumer AI assistants, yes, unless you turn it off. Google's Gemini Apps do it by default, Mistral's free chat mode unless you opt out, and Anthropic's Claude only if you switch Model Improvement on.

The rest of this page is the reading behind that sentence, vendor by vendor, followed by the settings paths and the carve-outs that survive them.

Four Postures, And Which Vendors Sit In Each

Across the policies we could read, vendor behaviour falls into four groups rather than a simple yes or no.

Counted across the six consumer chat products whose policy we could read directly, exactly one, Anthropic's Claude, requires you to switch training on. Three train unless you act. Two state that they do not use conversation content for training at all.

Google Gemini: The Default That Outlives Your Delete

Google's Gemini Apps privacy hub is unusually specific, and the specifics are the interesting part. With Keep Activity on, which is the default, chats are saved in your Activity and "used to improve Google services (including training generative AI models)." Human reviewers, including trained reviewers from service providers, read some of that data.

The carve-out is the sentence most people never reach. Chats that have been through human review are not deleted when you delete your activity. They are retained for up to three years. So the delete button and the training pipeline are not connected in the way the interface implies, and turning Keep Activity off later does not reach back into that pool. Anyone treating "I clear my Gemini history" as equivalent to "Google has none of this" is working from the wrong model of the system.

Anthropic Claude: The One That Asks First

Anthropic's consumer article, dated 16 March 2026, lists exactly three conditions under which your Claude chats are used to improve models: you chose to allow it, the conversation was flagged for safety review, or you explicitly opted in through something like the Trusted Tester programme. The setting is called Model Improvement and lives in Privacy Settings. Incognito chats are excluded even when Model Improvement is on.

Two boundaries keep this from being a clean win. Safety-flagged conversations are used regardless of the setting, which is a reasonable design and still a real exception. And the opt-out is forward-looking: Anthropic states directly that your data "will still be included in model training runs that are already in progress, or in models that have been trained." That is the honest version of a sentence most vendors leave out.

Mistral: The Answer Changes When You Start Paying

Mistral's privacy and data controls documentation splits by tier in a way worth quoting. Free mode: "conversations can be used to improve our models. You can opt out at any time from the Admin Panel." Pro, Team and Enterprise: "conversations aren't used for model training by default." API: not used for training.

Then there is a line further down the same page that overrides all of it. Enabling Labs models means "data can be used to train Mistral models, regardless of your subscription plan or opt-out settings." A single feature toggle that outranks the privacy toggle is the sort of detail that never appears in a comparison chart, and it is a good argument for reading the vendor's own page rather than a summary of it, including this one. Note also that Mistral's own documentation now calls the chat product Vibe, so searching for Le Chat settings may send you to stale instructions.

DeepSeek: An Opt-Out Right, And A Jurisdiction That Does Not Move

DeepSeek's privacy policy describes using personal data "to improve and develop the Services and to train and improve our technology, such as our machine learning models and algorithms," and separately lists "the right to opt-out of using your Personal Data for training our models or optimizing our technologies."

The line that matters more for most threat models is about location, not training: "we directly collect, process and store your Personal Data in People's Republic of China." No setting changes that, and a training opt-out does not either. If jurisdiction is your constraint, it is the first field to read and it makes the training question secondary.

The Policies That Would Not Load

We requested thirteen published data-policy documents from eleven AI companies with an ordinary desktop browser user agent on 18 August 2026. Nine returned the document. Four returned HTTP 403: OpenAI's ChatGPT consumer help centre, xAI's legal page, Meta AI, and Perplexity's privacy hub.

This is not an accusation. Bot filtering is routine, these companies publish these documents to humans in browsers, and a 403 to an automated request says nothing about the contents. It does have one concrete consequence worth naming: an answer engine summarising "does AI train on your chats" cannot read four of these vendors' policies either, so the summaries you meet in the wild are reconstructing them from secondary coverage. That is exactly how a stale opt-out instruction survives for years. For those four vendors, the honest advice is to open the setting in the product rather than trust any page, this one included.

Turning It Off, In The Order That Works

1. Gemini. Open Gemini Apps Activity and turn off Keep Activity. Separately, set auto-delete to 3 months rather than the 18-month default, because the two controls do different jobs.
2. Claude. Open Privacy Settings and confirm Model Improvement is off. Use Incognito chats for anything you would not want reviewed, since they are excluded either way.
3. Mistral. Open the Admin Panel, then the chat privacy section, and disable Model training. Check whether Labs models are enabled, because that switch overrides the first one.
4. DeepSeek. Exercise the stated opt-out, then decide separately whether the storage jurisdiction is acceptable for the work you were planning to do.
5. Everything else. Find the setting in the product, screenshot it with a date, and re-check after any major product update. Settings move.

What Switching It Off Does Not Undo

Three things survive every opt-out we read. Models already trained keep whatever they learned, and no vendor offers retraining on request. Safety and abuse pipelines continue to process flagged content by design. And human-reviewed material, at Google specifically, outlives your deletion by up to three years.

Which points at the more durable lever: whether a conversation was ever written down. A setting is a promise about what a company will do with a stored record. An architecture that never creates the record is a different kind of statement, and it is the one worth paying for if you are already choosing between subscriptions. If that changes which plans you keep, the plan-by-plan price index behind these vendors is where the consolidation arithmetic lives.

Where Perspective AI Sits In This Table

Perspective AI's published privacy documentation commits to the platform not logging, retaining, or repurposing conversation content, and describes the billing design that makes it structural: usage is metered by amount rather than by reading what you asked. Private mode goes further by removing the server-side record entirely, storing history encrypted in your browser and relaying each request statelessly, which is the mechanism our breakdown of what private mode covers maps claim by claim to its source. What you will not find is an absolute, because your prompt still has to reach the model that answers it, and that is true of every product on this page.

The practical value is that the decision stops being a subscription-wide bet. One $14.99/mo plan reaches the whole catalog, and you choose per conversation whether a transcript exists on a server at all. For the related question of how long the vendors above keep what they do store, the retention clocks compared side by side pick up where this page stops, and what private actually means for one person covers the vocabulary problem underneath both.

FAQ