In short. The AI Act requires providers of general-purpose AI models to publish a training content summary using the European Commission template.

Definition: An AI Act training content summary is a public account of the content used to train a general-purpose AI model. The European Commission's official template, current as of 17 October 2026, sets the required disclosure format for providers placing these models on the Union market.

Key facts

Article 53(1)(d) of the AI Act requires providers to publish a sufficiently detailed summary following the AI Office's template.
The Commission's template FAQ states that using the template is mandatory; adherence to the General-Purpose AI Code of Practice is voluntary.
The Explanatory Notice covers training from pre-training through post-training, including alignment and fine-tuning.

What does the official template contain?

The Commission's template and explanatory notice organise disclosure into general information, data sources and relevant processing aspects. The following outline reproduces selected headings from the annex:

1. General information
1.1. Provider identification
1.2. Model identification
2. List of data sources
2.1. Publicly available datasets
2.3. Data crawled and scraped from online sources

General information includes model identifiers, training data types, size ranges and linguistic characteristics. A modality is a type of data, such as text, image or audio. The data-source section distinguishes public datasets, private datasets, directly scraped online content, user data and synthetic data, meaning content generated by another AI model. Processing disclosures include measures concerning illegal content.

How does the summary fit into the legal obligation?

Article 53 of Regulation (EU) 2024/1689 separates the public summary from technical documentation supplied to the AI Office and national authorities upon request. It also requires a copyright compliance policy, including identifying and respecting reservations of rights under the EU text and data mining rules.

The summary gives rightsholders information to exercise their rights. Recital 107 calls for comprehensive coverage in narrative form while taking account of trade secrets and confidential business information. Its examples include the main data collections, large databases and data archives used for training.

What does it reveal about website content?

The Commission's FAQ on training summaries specifies crawler names, collection periods and descriptions of scraped content. It also requires the top 10% of scraped domains; for small and medium-sized enterprises, the threshold is the top 5% or 1,000 domains, whichever is lower.

The template's online-source section ranks domains by the size of content scraped and used for training, with representative coverage across modalities where applicable. Its example domain format is example.com. A domain identifies a website; an individual page URL can identify a particular work. Disclosure of individual works and their URLs is optional. For a website publisher, the required domain list therefore provides a starting point for checking the provider's reported sources.

Where and when is the summary available?

The Commission's publication guidance requires public availability by the time the model enters the Union market. It calls for a visible, accessible summary on the provider's official website and alongside the model across public distribution channels, clearly identifying the covered models and versions.

The obligation applies from 2 August 2025. Models already on the market before that date have a compliance deadline of 2 August 2027. Further training that changes the summary calls for updates at six-month intervals, or sooner when the change is materially significant, with the update date included.

FAQ

Do open-source model providers have to publish a summary?

Yes. The Commission's FAQ expressly includes models released under free and open-source licences in the publication requirement.

Does the summary cover pages retrieved during an answer?

The Explanatory Notice excludes pages supplied as input during model operation from mandatory disclosure unless the model actively learns from them.

Sources

Check how a page reads to answer engines. Open the GEO score analyzer