The EU Training Data Public Summary Template

Article 53(1)(d) requires providers of general-purpose AI models to draw up and make publicly available a sufficiently detailed summary about content used to train the model, according to a template provided by the AI Office. The Commission adopted and published the template in July 2025. The stated purpose is to improve transparency and help people with legitimate interests, including copyright holders, exercise rights under EU law. The template asks for a generally comprehensive account, not a technically exhaustive list of every training item. It includes information about data sources, collections, and other content used. The Act’s recital says the summary should account for trade secrets and confidential business information, should be broad in scope rather than technically detailed, and may identify major data collections plus a narrative description of other sources. The template does not require providers to reveal every record, model weight, or proprietary engineering detail. The summary is not proof that all training content was lawfully used, nor does it determine whether a particular copyright work appeared in a dataset. The AI Office’s role under the Act is to monitor whether the provider has fulfilled the obligation, not to conduct a work-by-work copyright assessment. A provider also has a distinct obligation to put in place a policy to comply with EU copyright law and respect rights reservations under the relevant directive. These duties should not be collapsed into a single statement. A provider should map training sources and collections, record provenance at an appropriate level, document how the template’s fields were answered, and publish a summary that accurately reflects the model’s training content. Keep a version history when training data or model scope changes. If a provider relies on a limited open-source transparency exemption, check its precise conditions and the retained copyright duties in the current consolidated Act.