For press and research
Trained On is a public record of what AI companies' own terms say about training on your prompts, quoted and dated, and of every time those words changed.
It reads Open Terms Archive's recorded history of each company's terms and privacy policy, keeps only the changes to the sentences about training, and shows each one old against new. It quotes; it does not judge any company, and it is not legal advice.
What this record does
- The old words next to the new. Every change shows the clause before and after, with the removed and added words marked, and links the archive captures it comes from, so anyone can check it.
- Nothing is published on one reading. Three AI readers from three companies (Anthropic, OpenAI and Google) read each change on their own. It is published only when all three agree it is real, under the most cautious of their labels; anything else goes to a person. Every change page says which of the two decided it. How it works.
- Free and open. No account and no paywall. The registry and every change download as CSV, JSON and Atom under ODC-By 1.0, and the code that builds the site is public.
- Built on Open Terms Archive. The document history is theirs, recorded by its contributors. This site separates the changes to the training clause from the noise around them.
The numbers
These figures update with every nightly rebuild.
How to cite
In an article, one line is enough: Trained On (trainedon.me), a public record of what AI companies' terms say about training on your prompts, built on Open Terms Archive.
The record as a whole:
Trained On, "Is this AI training on my prompts?", https://trainedon.me/. Derived from Open Terms Archive, ODC-By 1.0.
The dataset:
Trained On, "The registry and the change record" (dataset), https://trainedon.me/data. Derived from Open Terms Archive, ODC-By 1.0.
A single change: every change page carries its own, under Cite this change. Its address is built from the date, the vendor and the document, and it is permanent.
Changes of position on record
- Cognition now says it "may use Customer Data for model training purposes"; "paid Service Tier" customers "may opt out" (Teams administrator only). Change of position · Windsurf (Codeium), Terms of Service
- xAI's enterprise terms dropped automatic 30-day deletion of User Content and claimed ownership of de-identified data for any lawful purpose. Change of position · Grok, Commercial Terms
- GitHub now says it may train on “Your Content” and Affiliates may use Inputs/Outputs “Unless you opt out,” replacing “if you allow in your settings.” Change of position · GitHub Copilot, Terms of Service
- DeepSeek now states that "certain entities within our corporate group" process your provided, automatically collected and third-party data for "foundation model training and optimization"; the opt-out remains. Change of position · DeepSeek, Privacy Policy
- DeepSeek still may train on Personal Data, but now grants “the right to opt-out of using your Personal Data for training our models or optimizing our technologies.” Change of position · DeepSeek, Privacy Policy
- Jasper still may use Customer Property for “modifying, improving, and enhancing artificial intelligence models,” but EU/UK users “may have the right to object.” Change of position · Jasper, Terms of Service
- Anthropic's Claude.ai now uses Materials for "training our models, unless you opt out of training through your account settings," with feedback and safety-flagged content trained on regardless. Change of position · Claude.ai, Terms of Service
- Claude.ai (Anthropic) now says it “may use your Inputs and Outputs to train our models…unless you opt out through your account settings,” with safety-review/reported-materials exceptions. Change of position · Claude.ai, Privacy Policy
- xAI dropped its limit on training "only with the Usage Data you voluntarily choose to provide to us." Change of position · Grok, Privacy Policy
- Grok (xAI) now says logged-in users “can select whether or not” User Content trains models; logged-out use “where permitted” grants “full rights” for training. Change of position · Grok, Terms of Service
- Cursor replaced its Privacy Mode clause with: Anysphere will not train, or let any third party train, AI models on your content unless you explicitly agree. Change of position · Cursor, Terms of Service
Every change, including scope changes and new disclosures.
Follow it
Every future change, as a feed. Each vendor page has a feed of its own, and the data page has the files.
Contact
[email protected]. I am glad to answer questions about the record and how it is made.