AI Privacy Compared

10:36 Most AI assistants can use what you type to train their models unless you change a setting, as of Nov 6, 2026. ChatGPT trains on Free, Go, Plus and Pro chats unless you turn off "Improve the model for everyone"; Microsoft Copilot, Perplexity, Mistral's Vibe, Grok and DeepSeek also use consumer chats by default. Gemini trains on your chats only while Keep Activity is on, and Claude only if you choose to allow it. Work plans (ChatGPT Business and Enterprise, Microsoft 365 Copilot, Perplexity Enterprise, Vibe Enterprise) do not train on your data by default. For the most privacy, use a temporary or incognito chat, or run an open model on your own computer (Run AI Locally).

Training, opt-out and retention by assistant

Sources: OpenAI, Anthropic training and retention, Google, Microsoft, Perplexity, SpaceXSI, Mistral, DeepSeek and Meta, all read Nov 6, 2026. "Not stated" means the page read did not say.

Each assistant in detail

ChatGPT. On Free, Plus and Pro, OpenAI may use your content to train its models; turning off "Improve the model for everyone" or choosing "Do not train on my content" in the Privacy Portal stops it for new chats and Codex tasks. A thumbs up or down still lets OpenAI use that whole conversation, even after you opt out. Temporary chats are not used for training. ChatGPT Business, Enterprise, Edu and the API are not used for training by default. OpenAI's pricing page says the same plan by plan.

Claude. Anthropic uses consumer chats (Free, Pro, Max) for training only if you turn on Model Improvement, if a chat is flagged for a safety review, or if you join a program such as its Trusted Tester Program. With it on, de-identified chats may be kept up to 5 years; deleted chats leave back-end storage within 30 days; incognito chats are never used, and feedback is kept 5 years. Team, Enterprise and API use are covered by Anthropic's separate commercial terms.

Gemini. With Keep Activity on, chats are saved and used to improve Google services, including training generative models, and a subset is read by human reviewers after being disconnected from your account; reviewed chats are kept up to three years. Turning Keep Activity off keeps chats for 72 hours and leaves them out of training unless you send feedback. Audio and Gemini Live video are not used for improvement by default. The page does not say whether Keep Activity starts on. Work and school accounts follow Google Workspace terms.

Microsoft Copilot. Microsoft uses consumer Copilot conversations, including uploaded files and images, for training except for users who opt out and certain groups, including users in Brazil, mainland China, Israel, Nigeria, South Korea and Vietnam. Conversations are stored for 18 months by default. Anyone signed in with a work account is excluded; see Microsoft 365 Copilot, whose chats fall under Enterprise Data Protection.

Perplexity. Perplexity uses searches and feedback to improve its service unless you turn off AI Data Usage in Settings. Its contracts with OpenAI, Anthropic and other model makers forbid them from training on Perplexity data, and Enterprise Pro and Max never train on your data.

Grok. SpaceXSI (formerly xAI) lists training its models among the uses of what you type, and uses public X posts as well; Private Chat conversations and deleted chats are removed within 30 days unless kept for legal or safety reasons (privacy policy, updated Oct 27, 2026). Grok used inside X follows X's privacy policy instead.

Vibe (formerly Le Chat). Mistral trains on standard Vibe plans by default, with an opt-out; Enterprise customers start opted out. Rating a reply lets Mistral use it whatever your plan.

DeepSeek. DeepSeek's policy says it uses your prompts, files and chats to train its models and stores them in the People's Republic of China; you can request an opt-out by email. Running DeepSeek's open weights on your own hardware sends nothing to the company.

Meta AI. Since Dec 16, 2025, Meta uses your conversations with its AI to personalize the posts, reels and ads you see on its apps, except conversations about sensitive topics such as health, religion or politics; Meta describes no direct opt-out, only its ad and feed controls. For its Muse agent, Meta says conversations are not shared with its ad systems and you can opt out of training.

Work and meeting tools

Which AI is most private?

Most private: a local model. Open models run with Ollama or LM Studio keep everything on your computer.

7:20 Most private by default among cloud assistants: Claude. Anthropic does not train on consumer chats unless you allow it, and incognito chats are never used. A Feb 9, 2026 video by Business Reform, which scored 30 cloud assistants on their published policies, ranked privacy-focused services such as Proton Lumo and DuckDuckGo AI above all the big names, put Claude, ChatGPT, Perplexity, Grok and Copilot in the middle third, and Gemini and Meta AI in the bottom third.

Easy to make private: ChatGPT and Gemini. One setting each (Improve the model for everyone; Keep Activity), plus temporary chats.

Least control: DeepSeek and Meta AI. DeepSeek's opt-out is by email and its data sits in China; Meta describes no opt-out from AI-based ad personalization.

How we chose

Each row comes from the maker's own privacy policy or help page, read on Nov 6, 2026 and linked. We describe what the pages say, not how the companies behave in practice, and leave a cell "Not stated" where the page did not say. Settings change; check the linked page before sharing sensitive material.

Frequently asked questions

Does ChatGPT train on my data? Yes, on Free, Go, Plus and Pro unless you turn off "Improve the model for everyone" in Settings, Data controls, as of Nov 6, 2026; ChatGPT Business, Enterprise and the API are not used for training by default.

Does Claude train on my chats? Only if you allow it: Anthropic uses Claude Free, Pro and Max chats for training when Model Improvement is on, and never uses incognito chats, as of Nov 6, 2026.

Does Gemini use my conversations for training? Yes, while Keep Activity is on; with it off, Google keeps Gemini chats for 72 hours and does not use them for training unless you send feedback.

Which AI is the most private? An open model run on your own computer is the most private; among cloud assistants, Claude does not train on consumer chats unless you opt in, and its incognito chats are never used, as of Nov 6, 2026.

Does Microsoft Copilot train on my data? Microsoft's consumer Copilot uses conversations for training unless you opt out, except in Brazil, mainland China, Israel, Nigeria, South Korea and Vietnam; work accounts are excluded.

Is DeepSeek safe for private information? DeepSeek stores data in China and trains on your prompts unless you request an opt-out by email, so avoid sharing sensitive information; its open weights can be run locally instead.

Sources

• How your data is used to improve model performance, OpenAI Help Center, read Nov 6, 2026
• Plans and pricing, ChatGPT, read Nov 6, 2026
• Is my data used for model training?, Anthropic Privacy Center, Mar 16, 2026
• How long do you store my data?, Anthropic Privacy Center, Jul 1, 2026
• Gemini Apps Privacy Hub, Google, Oct 26, 2026
• Privacy FAQ for Microsoft Copilot, Microsoft Support, read Nov 6, 2026
• What is Microsoft Copilot?, Microsoft Learn, Nov 5, 2026
• What data does Perplexity collect about me?, Perplexity Help Center, read Nov 6, 2026
• Are third-party model providers training on my data?, Perplexity Help Center, read Nov 6, 2026
• Enterprise pricing, Perplexity, read Nov 6, 2026
• Privacy policy, SpaceXSI, Oct 27, 2026
• Do you use my user data to train your AI models?, Mistral Help Center, read Nov 6, 2026
• Privacy policy, DeepSeek, Feb 10, 2026
• Improving your recommendations on our apps with AI at Meta, Meta, Oct 1, 2025
• Introducing Muse, Meta, Sep 8, 2026
• Security, Granola, read Nov 6, 2026
• Pricing, Notion, read Nov 6, 2026
• Lawsuit claims AI service Otter secretly records private chats, NPR, Aug 15, 2025
• Most Private AI in 2026? I Scored 30 Platforms, Business Reform, Feb 9, 2026 (video)