Which AI companies train on your data? We read 15 apps' policies

The Morse teamUpdated October 23, 20269 min read

Most AI companies train on your data by default. Of the 15 consumer AI apps whose own policy and help pages we read on October 19, 2026, 11 train on your chats, including ChatGPT, Claude, Gemini, Grok, Meta AI, Perplexity, and DeepSeek. Two of those 11 document no way out for US users: Meta AI and Character.AI. Qwen Chat's way out is an email to privacy@qwen.com. Two say they don't train on your chats (Proton's Lumo and DuckDuckGo's Duck.ai), the new Microsoft Copilot app says your prompts aren't used to train foundation models, and Venice's privacy policy doesn't mention training.

We went app by app, read the company's own pages that day, and wrote down four things: whether training is on by default, the exact name of the switch and where it is, whether people at the company may read your chats, and whether business plans work differently. Every cell in the table links to the page it came from. Where a company's page didn't answer a question, the cell says so instead of guessing. For a wider look at retention, ads, and courts across the five biggest apps, see whether AI chats are private.

Which AI apps train on your chats by default?

Personal accounts only, free and paid. Read October 19, 2026.

The count is 11 yes, 2 no, one that depends on which version of the app you have, and one that doesn't say. Policies change often, so treat this as a snapshot of October 19, 2026.

How do you stop an AI app from training on your chats?

Find the switch and turn it off before your next sensitive chat. The paths below are each company's own wording.

• ChatGPT: account menu, Settings, Data controls, then turn off "Improve the model for everyone" (OpenAI). Our ChatGPT training guide covers temporary chats and the Privacy Portal route.
• Claude: your name, Settings, Privacy, then toggle off "Help Improve our AI models" (Anthropic). Our page on whether Claude is private explains the five-year retention that comes with leaving it on.
• Gemini: on gemini.google.com, Settings & help, Activity, then "Turn off" or "Turn off and delete activity" (Google). This also turns off your chat history, and Google still keeps new chats for 72 hours (Google).
• Grok: in the app, Settings, Data Controls, "Improve the model"; on grok.com, Settings, Data, "Improve the Model" (xAI). Our page on whether Grok is private has the rest.
• Perplexity: Account settings, Preferences, then the "AI data retention" toggle (Perplexity). Another of its help pages calls the same switch "AI Data Usage" (Perplexity).
• Mistral Vibe: on the web, the Admin panel, Vibe, Privacy, then disable "Allow your interactions to be used to train our models"; in the mobile app, Settings, Data & Account Controls, then clear "Enable data sharing" (Mistral).
• Older Copilot app: on copilot.com, profile icon, profile name, Privacy, then the two "Training on…" switches (Microsoft). Microsoft says this switch doesn't cover "other general product or system improvements" or "advertising."
• Character.AI (EEA and UK): in the app, Settings, "Data & Privacy", then clear "Improve the Model for Everyone" (Character.AI).
• DeepSeek: its terms tell you to turn off "Improve the model for everyone" but give no menu path; the other route is privacy@deepseek.com (DeepSeek).

Does opting out stop all training?

No. Six of the apps say feedback can still be trained on after you switch off. OpenAI says a thumbs up or down means "the entire conversation associated with that feedback may be used to train our models" (OpenAI). Anthropic keeps a rated conversation for up to five years (Anthropic), and also trains on chats "flagged for safety review" after you opt out (Anthropic). xAI says feedback "may be used for training purposes" even after an opt-out (xAI). With Keep Activity off, sending Google feedback brings in "the last 24 hours of your chats" (Google). Mistral says a rating with a comment authorizes training "whichever subscription plan you adhere to" (Mistral). Qwen's terms say feedback "will still be used for model training" after you opt out (Qwen).

The switch also only looks forward. Perplexity says "opt-outs only apply to data collected after the opt-out date" and "previously collected training data cannot be deleted or removed" (Perplexity). xAI covers "new conversations" (xAI), Character.AI covers "new content" (Character.AI), and Meta says deleted data "won't be used for future training" (Meta).

Does where you live change the answer?

Yes, for several apps. Character.AI's opt-out page is written for users "in the European Economic Area ("EEA") or the United Kingdom" (Character.AI). xAI says that "in some regions (excluding the EU/UK)," people using Grok without logging in "won't have the option to opt out" (xAI). The older Copilot app trains on no user data in Brazil, China (excluding Hong Kong), Israel, Nigeria, South Korea, and Vietnam (Microsoft). Google applies its paid Gemini API terms, which don't use prompts to improve products, to free API use in the EEA, Switzerland, and the UK (Google). Meta publishes separate EU and UK AI terms, and the US pages we read describe no opt-out.

Do business plans train on your data?

Mostly no, by default. OpenAI, Anthropic, xAI, and Perplexity keep business customers out of training, and so does Mistral's Enterprise plan. Microsoft says work and school prompts aren't used to train foundation models, and Google says Workspace content isn't used for training "outside of your domain without permission." Two caveats from the companies' own pages: Anthropic and Mistral still train on business feedback you choose to send (Anthropic, Mistral), and DeepSeek's API terms don't say either way.

Which guides on this are out of date?

Three we found while checking:

• Copilot. Engadget's guide, published September 28, 2026, says Copilot's "Training on conversation activity" is "on by default" under Settings, Privacy (Engadget). Microsoft's new Copilot app came out on August 18, 2026, and its privacy controls page lists no training switch (Microsoft), and its activity page says prompts "aren't used to train foundation models" (Microsoft). The old switch now applies only to the older app. Microsoft's general privacy statement still says it "may use your data to develop, train, and fine-tune our AI models" (Microsoft), so "foundation models" is the narrower promise.
• Claude. Cloaked's April 2026 guide tells readers to look for "Help improve Claude" (Cloaked). Anthropic's help page names it "Help Improve our AI models" (Anthropic).
• Le Chat. Mistral renamed Le Chat to Vibe on August 12, 2026, and says "your account, plan, conversations, and saved content all move with you" (Mistral). Guides that send you to Le Chat's settings describe screens that now say Vibe, with the switch names in the table above.

How Morse keeps your chats out of training

Every switch above sits on a server that already holds your chats, and the company decides what the switch covers: new chats but not old ones, chats but not feedback, your country but not someone else's. Morse is built so there is nothing on our servers to train on. Morse doesn't train AI on your content, sell your data, or show ads, and your chat history is stored in your browser, encrypted, not on our servers (privacy policy).

On Morse, there's no default to check, because nothing is collected for training. There's no switch to hunt for, so no settings page can rename it or leave it out of a menu. And there are no feedback exceptions or country lists to read, because Morse doesn't train on your chats.

Each chat has a privacy mode, so you choose how far the message travels:

• Private, for chats you don't want any host to keep. It's the default for a new chat. Open-weight models like DeepSeek, Kimi, and GLM, plus Grok, run on hosts that agreed to keep nothing, under zero data retention. Where a host confirms that on each reply, Morse checks it and refuses a reply that doesn't. DeepSeek's models come here without the DeepSeek app's training terms.
• E2EE, for anything no one should read. Your browser checks the enclave's hardware proof before anything leaves your device, then encrypts the message so only that verified enclave can open it. Morse's servers pass along text they can't read. This rests on the enclave hardware and attestation working as designed, and every reply has a lock that shows each check that passed.
• Anonymous, for GPT, Claude, and Gemini without the provider knowing it's you. Our proxy removes your account, email, and IP address and sends the message under Morse's own account. The provider still sees the words and may keep them for a limited time to check for abuse, under its own policy, so leave out names and details that point to you.

In every mode, your history stays in your browser, locked when you sign out and opened with two keys, one of which never leaves your device. Deleting a chat or clearing your browser removes it for good, and Morse can't recover it, so keep a copy of anything you need. Because we don't store chats, we usually have very little to give authorities.

Try a chat from the Morse home page without an account, or start free. End-to-end encrypted chats come with Pro, at $19 a month.