Who it applies to
Providers of general-purpose AI models must draw up and publish a sufficiently detailed summary of the content used to train the model, using a template from the AI Office (Article 53(1)(d)). The Commission’s guidelines give an indicative criterion for what counts as a general-purpose model: training compute above 1023 FLOP and the ability to generate language (text or audio), text-to-image or text-to-video. Open-source models with public weights are exempt from the documentation duties, unless they have systemic risk, but not from the copyright policy or the training summary.
What the template asks for
- General information: the provider, the model, its modalities, and the overall size and characteristics of the training data.
- List of data sources: public data sets, private data sets from third parties (licensed or other), data crawled and scraped from the web (crawler name, purpose, behaviour such as whether it respects robots.txt, collection period and a description of the content), user data, synthetic data and other sources.
- Data processing aspects: how rights reservations against text and data mining were respected, whether you signed the GPAI Code of Practice, and how illegal content was removed.
Level of detail
The summary should be comprehensive in scope rather than technically detailed, for example by listing the main data collections and explaining other sources in plain text. Trade secrets are taken into account.
When to publish and update
Publish the summary at the latest when the model is placed on the market, on your website and in every distribution channel, and update it every six months or earlier after a significant change. Models placed on the market before 2 August 2025 have until 2 August 2027. The AI Office has been able to enforce the model rules since 2 August 2026.
If you fine-tune someone else’s model
If a modification makes you a provider, you report only the content used for the modification and name the original model. See fine-tuning and the provider role.
Start with an inventory
The summary is much easier if you already keep a list of every data source, how it was collected, and what you cleaned or filtered out. See how to document a dataset. This guide is general information, not legal advice.