Training-data summary
Article 53 requires GPAI model providers to draw up and make publicly available a sufficiently detailed summary of the content used to train the model, using a template provided by the AI Office.
The summary sits alongside the Article 53 duty to put in place a policy to comply with Union copyright law, including identifying and respecting text-and-data-mining reservations expressed under the Copyright Directive. It must be sufficiently detailed to be useful to parties with legitimate interests — copyright holders above all — without requiring disclosure of the full dataset or trade secrets. The AI Office provides the template, which standardises the level of detail across providers. Unlike most GPAI documentation duties, the summary and the copyright policy survive the free and open-source relief.
- Art. 53(1)(c)
- Art. 53(1)(d)
Training-data summary: why it matters
This is the Act's main lever on the AI-and-copyright question: it does not resolve whether training was lawful, but it makes the inputs visible enough for rights holders to ask. Because the summary is public and template-driven, it is comparable across providers in a way no other AI Act artefact is — which turns it into a de facto disclosure standard that enterprise buyers and journalists will read side by side. And since it survives the open-source carve-out, publishing weights openly does not avoid it.
Training-data summary in practice
A provider preparing its summary has to describe the main data collections used, the provenance categories — licensed corpora, publicly available web data, user data, synthetic data — the scale, and how TDM reservations were detected and honoured during crawling. The last part is the one that bites: a provider that scraped before implementing reservation detection cannot describe a process it did not run, and the gap is visible in a public document.
Training-data summary: the mistake people make
Reading "sufficiently detailed" as a licence to publish a paragraph of generalities. The summary is meant to help rights holders exercise their rights, and it follows an AI Office template precisely so that vague submissions are visibly thinner than compliant ones.
Training-data summary: common questions
- Does the training-data summary mean publishing the dataset?
- No. It requires a sufficiently detailed public summary of the content used for training, following the AI Office template — not disclosure of the dataset itself, and it is designed to work alongside trade-secret protection.
- Do open-source models have to publish a training-data summary?
- Yes. The free and open-source relief in Article 53 does not extend to the copyright policy or the public training-content summary, and it does not apply at all to models with systemic risk.
- What is the copyright policy that goes with it?
- Article 53(1)(c) requires GPAI providers to put in place a policy to comply with Union copyright law, including identifying and respecting text-and-data-mining rights reservations expressed under Article 4(3) of Directive (EU) 2019/790.
Training-data summary: terms that come up alongside it
Run the free 2-minute EU AI Act checker — no sign-up.