Trained On AI training-data disclosure

About this register

What this is

Trained On records whether AI model providers have published the training-data disclosures that transparency law requires of them. It currently covers the EU obligation; California's AB 2013 asks a similar question of the same companies and is the intended second section.

Article 53 of Regulation (EU) 2024/1689 requires providers of general-purpose AI models to publish a sufficiently detailed summary of their training content and a policy for complying with EU copyright law. The obligation has applied since 2 August 2025. Enforcement powers — information requests, model access, and fines of up to €15 million or 3% of global turnover — took effect on 2 August 2026.

This register records, for each provider we track, whether those documents exist at the sources we monitor, with a link and the date we last looked.

What this is not

It is not a compliance assessment, a certification, or legal advice. An entry marked Not published means we did not find a public document at the sources we monitor. It does not mean a provider is breaking the law.

Which deadline applies

This distinction decides what a blank cell means, so it is worth stating precisely. Models placed on the EU market from 2 August 2025 had to publish a training-data summary from that date. Models already on the market before it have a transitional period and are not required to comply until 2 August 2027. The Commission's enforcement powers - information requests, model access, fines - began on 2 August 2026.

Most models on this register predate the cutoff. For them, “not published” means not yet due, not late. We record the release date where a provider states it, and leave it blank rather than guess, because a wrong date here would change which deadline the register appears to report.

How it is compiled

  1. We fetch a small set of published pages for each provider, respecting robots.txt.
  2. Pages are re-fetched on a schedule; unchanged pages are not re-read.
  3. Changed pages are read into structured fields, each carrying a source URL, a supporting quote, and a confidence.
  4. Quotes are checked against the source text. A quote that does not appear verbatim is rejected.
  5. Low-confidence readings are held for human review rather than published.
  6. Field-level changes are recorded in the change log.

Corrections

Send the URL of anything we have missed or misread. Corrections are applied within a week, and the change is recorded in the log like any other — including when we were the ones who were wrong.

Disclaimer

This tracker records whether providers have published specific documents, and when we last checked. It is not a legal assessment, and an entry marked "not published" means we did not find a public document at the sources we monitor - not that a provider is non-compliant. Providers may publish elsewhere or qualify for a later deadline. Every field links to its source.