Short answer. A GPAI training-data summary must follow the AI Office template published on 24 July 2025, and it has three mandatory sections: general information about the provider and model, a list of data sources, and relevant data-processing aspects. The template is compulsory, not advisory, and it applies to every provider placing a general-purpose AI model on the Union market, including models released under free and open-source licences.
Article 53(1)(d) of the EU AI Act is short. It tells providers of general-purpose AI models to draw up and make publicly available a sufficiently detailed summary about the content used for training, and to do it according to a template provided by the AI Office. Everything that makes the obligation real sits in that template, which the AI Office published on 24 July 2025 alongside an explanatory notice.
The obligation has been applicable since 2 August 2025. The part most teams get wrong is what happens next: the AI Office can take enforcement action from 2 August 2026, and models already on the market before 2 August 2025 have until 2 August 2027. Those are three different dates doing three different jobs, and confusing them is the most common planning error in this area.
Key takeaways
- The template is mandatory, not a suggested format. Article 53(1)(d) requires its use.
- It has three sections: general information, list of data sources, and relevant data-processing aspects.
- It covers every provider placing a GPAI model on the Union market, including providers releasing under free and open-source licences, and including models with systemic risk.
- Scraped-data disclosure is the sharpest requirement: crawler names, collection period, a description of the content, and the top 10% of all domains scraped — or for SMEs, the top 5% or 1,000 domains, whichever is lower.
- The summary must be updated at six-month intervals, or sooner if new training data makes a materially significant difference.
- Non-compliance may draw fines of up to 3% of total worldwide annual turnover or 15,000,000 euros, whichever is higher.
Quick comparison
| Date | What happens | Legal basis |
|---|---|---|
| 24 July 2025 | AI Office publishes the template and explanatory notice | Article 53(1)(d) |
| 2 August 2025 | Obligation becomes applicable; a new model needs its summary at market placement | Article 113(b) |
| 2 August 2026 | AI Office can begin enforcement action | Commission guidance |
| 2 August 2027 | Deadline for models placed on the market before 2 August 2025 | Article 111(3) |
| Every six months | Summary reviewed and updated, or sooner on material change | Template requirement |
What exactly are the three sections?
The template sets a common minimal baseline split into general information, a list of data sources, and relevant data-processing aspects. Each section also invites voluntary additional detail, so a provider can say more than the floor requires but never less.
General information identifies the provider and the model, then describes the training content by modality — text, image, video, audio — with the size per modality given within broad ranges rather than exact counts, plus the general characteristics of the data.
List of data sources is where the work is. It asks separately about publicly available datasets, private datasets obtained from third parties, data scraped from online sources, data collected from users, and synthetic data. The template sets different levels of detail per source type, which is how the Commission balanced transparency against trade secrets.
Relevant data-processing aspects covers the processing that matters to people exercising rights under Union law — copyright above all — and includes what the provider did about removing illegal content.
How much must you disclose about scraped data?
Scraped data draws the most specific requirements in the whole template, and they are more granular than most teams expect. A provider must name the crawlers used, state the period of collection, describe the scraped content comprehensively, and publish a list of the most-scraped domains.
The domain list is the provision worth reading twice. The requirement is the top 10% of all domains scraped from the internet. For small and medium-sized enterprises the threshold relaxes to the top 5% or 1,000 domains, whichever is lower. This is the clause that turns a vague gesture at publicly available web data into something a rightsholder can check against their own site.
It connects directly to Article 53(1)(c), which separately requires a policy to comply with Union copyright law, including identifying and respecting a reservation of rights expressed under Article 4(3) of the Copyright in the Digital Single Market Directive. The summary tells rightsholders what was taken; the copyright policy tells them whether their opt-out was honoured. Read together, they are an audit trail.
What about user data and personal data?
The template requires a provider to disclose whether the model was trained on data collected through user interactions with its services and products, including interactions with its own AI models. It asks for the modalities of that user data and a description of the related services, and it does not require disclosure of personal information.
That boundary is deliberate and it is often misread in both directions. The summary is not a privacy notice, and the Commission's position is that further detail on the use of personal data belongs in the provider's privacy statements. But the existence of user-interaction training data is not optional to disclose. Training on conversations is a fact about your data sources, and the template asks for it.
Does this apply to open-source models?
Yes. The obligation covers any provider of a general-purpose AI model placing it on the Union market, and the Commission states explicitly that it also applies to providers releasing models under free and open-source licences.
This surprises people because other parts of the AI Act carve out open-source releases. Article 53(1)(d) does not. A permissive licence changes nothing about who has to publish a training-content summary, and neither does systemic-risk classification — providers of models with systemic risk sit inside the same requirement, with additional obligations elsewhere in the Act.
What if the data records simply do not exist?
For a model placed on the market before 2 August 2025, a provider that cannot supply some required information — because it is unavailable, or retrieving it would impose a disproportionate burden — must clearly state and justify those gaps in the published summary.
This is a narrow and conditional allowance, and it is worth being precise about its limits. It applies to legacy models, it requires best efforts first, and it does not permit silence: the gap has to be declared and reasoned in the document itself. For a model placed on the market after 2 August 2025 there is no equivalent relief.
That is the operational lesson for anyone sourcing data now. Record the licence, the collection period, the consent basis and the source for each dataset while you are acquiring it, not afterwards. A team that treats provenance as a delivery artefact finds this obligation to be paperwork. A team that treats it as an afterthought finds it to be archaeology, and the clause above is the description of that situation rather than an excuse for it.