Which AI Tools Train on Your Prompts? A Terms-Based Tracker

Summary

Trained On is a terms-based tracker of whether AI tools use users’ prompts, inputs, outputs, or customer content to train or improve models. It records the wording of companies’ own policies and dates the captured versions, while an archive of more than 1,000 document versions is condensed into a smaller set of substantive changes. The table distinguishes consumer, commercial, enterprise, and API offerings, because the answer often changes by plan. ChatGPT Free, Plus, and Pro are listed as training by default with opt-out controls, while ChatGPT Business, Enterprise, Edu, and API data is not used unless a customer opts in; temporary chats are excluded from training. Claude consumer plans are listed as training by default with an account-setting opt-out, while commercial services are listed as not training on customer content. Microsoft Copilot may train on conversations in certain markets, and GitHub Copilot is listed as training by default unless users opt out. Perplexity API and Enterprise are listed as not using customer content for model training, while Grok’s consumer setting leaves the initial choice unclear and logged-out use grants broader data rights; its API and enterprise terms prohibit training on user content, although de-identified data may have wider permitted uses. Cursor requires explicit agreement before training, whereas Windsurf’s free tier has no stated opt-out and paid tiers train by default with an opt-out. DeepSeek, Jasper, and free or individual Le Chat plans are also described as allowing training by default or under regional conditions, with exceptions for opt-outs, feedback, or moderated content. The change log records recent shifts in Windsurf, Claude, Grok, OpenAI, and GitHub’s stated positions, showing why users must check both the product plan and the current terms rather than rely on a general company-wide assumption.