In short. The AI Act requires a public training content summary using the Commission template, which requires selected domains and makes individual page URLs optional.
Definition: No: AI Act Article 53(1)(d) requires a public summary of the content used to train a general-purpose AI model. The Commission's template dated 5 December 2025 requires dataset information and selected domain names, while individual page URLs are optional.
Key facts
- The AI Act, Article 53(1)(d), requires a sufficiently detailed summary using the AI Office's template.
- The Commission's template, Section 2.3, requires the top 10% of scraped domains, ranked by the size of content scraped.
- The Commission's template FAQ confirms that the publication obligation also covers models released under free and open-source licences.
What does Article 53 require?
A general-purpose AI model can perform a wide range of tasks and integrate into different systems or applications, according to Article 3(63) of the AI Act. Article 53(1)(d) places the public training summary obligation on its provider.
Recital 107 explains the intended detail: broad coverage of training content, with the main data collections identified and other sources described in narrative form. It also calls for protection of trade secrets and confidential business information.
Article 53 distinguishes audiences. Paragraph 1(a) concerns technical documentation supplied on request to the AI Office and national authorities; paragraph 1(b) concerns information for providers integrating the model into their systems. Paragraph 1(d) concerns the public summary.
Which domains and pages appear in the template?
The Commission's explanatory notice, paragraph 16, expressly excludes mandatory disclosure of details of specific works. Its template separates publicly available datasets from data crawled directly by the provider or on its behalf.
Section 2.3 requests crawler identifiers, collection periods, content descriptions and selected domain names. The template's example of the domain format is:
example.com
The top 10% rule concerns domain names ranked by scraped content size. Small and medium-sized enterprises, including start-ups, disclose the top 5% or 1,000 domains, whichever is lower. The list can be a downloadable file. Additional domains, individual URLs and sources of individual works belong to optional disclosures.
An absent domain therefore leaves open whether particular pages supplied training content. For domains outside the published list, paragraph 16 recommends a voluntary mechanism through which rightsholders can request information about scraping and training use.
How can a website owner inspect a summary?
The Commission's template FAQ directs providers to publish summaries visibly on their official websites and alongside the model through public distribution channels. Each summary should identify the models and versions it covers.
For a website, the relevant checks are the domain list, the collection period and any named public datasets. The FAQ explains that summaries also describe private datasets and measures concerning copyright reservations. Together, these fields provide context for a rightsholder assessing the reported data sources.
Related terms
- AI Act Article 53
- AI Act training content summary
- General-purpose AI model under the AI Act
- Text and data mining opt-out (EU)
FAQ
Can I demand a page-by-page list under this requirement?
Article 53(1)(d) requires a summary. The Commission's template treats individual URLs as voluntary additional information; its recommended domain enquiry mechanism is also voluntary.
Must every model already have a summary?
The Commission's template FAQ dates the obligation from 2 August 2025. For models placed on the market before that date, it gives providers until 2 August 2027 to make the corresponding summaries available.
Sources
- European Parliament and Council, "Regulation (EU) 2024/1689", 13 June 2024
- European Commission, "Explanatory Notice and Template for the Public Summary of Training Content for general-purpose AI models", 5 December 2025
- European Commission, "Template for general-purpose AI model providers to summarise their training content", updated 26 March 2026
Check how a page reads to answer engines. Open the GEO score analyzer