WhatsApp and Telegram +44 7563 115980
[email protected]
Every provider of a general-purpose AI model placed on the EU market must publish a summary of the content used to train it, on the European Commission's template. This guide covers who is caught, the dates, each section of the template and how licensed data is reported.
Since 2013 we have helped firms with due diligence and data services, in more than 60 countries.
The short answer
Article 53(1)(d) of the EU AI Act requires every provider of a general-purpose AI model to publish a sufficiently detailed summary of the content used to train it, using the European Commission's template. The duty has applied since 2 August 2025. Models placed on the market before that date must have a summary by 2 August 2027. The Commission's AI Office has been able to enforce it since 2 August 2026, with fines of up to 3% of worldwide annual turnover or EUR 15 million, whichever is higher. For data licensed from rights holders, the template asks only whether such agreements exist and which types of content they cover.
On this page
A general-purpose AI model is a model trained on broad data that can perform many different tasks and be built into many products. Article 53 of the EU AI Act sets the core duties for the companies that provide these models. Point (d) requires them to draw up and make publicly available a sufficiently detailed summary about the content used for training, according to a template provided by the AI Office (the part of the European Commission that supervises these models).
Recital 107 of the Act explains the purpose: transparency about training data, including text and data protected by copyright, so that rights holders and others with legitimate interests can exercise their rights. The summary should be generally comprehensive in scope rather than technically detailed, and should take account of the need to protect trade secrets.
Point (c) of the same Article sits alongside it: providers must have a policy to comply with EU copyright law, including identifying and respecting reservations of rights made under the text and data mining rules. Text and data mining means automated analysis of text and data to find patterns.
| Date | What happened or happens |
|---|---|
| 1 August 2024 | The EU AI Act entered into force |
| 24 July 2025 | The Commission published the Explanatory Notice and Template, later adopted as Communication C(2025) 8311 final of 5 December 2025 |
| 2 August 2025 | Obligations for general-purpose AI models began to apply, including the training data summary |
| 2 August 2026 | The AI Office's supervision and enforcement powers for general-purpose AI models began |
| 2 August 2027 | Deadline for summaries of models placed on the market before 2 August 2025 |
Fines for providers of general-purpose AI models can reach 3% of annual total worldwide turnover in the preceding financial year or EUR 15 million, whichever is higher. The consolidated text on the Commission's AI Act Service Desk, which reflects the Digital Omnibus on AI, shows no amendment to Article 53 or to the 2 August 2027 date for existing models.
Every dataset we supply arrives with the records a provider needs to complete the template, including the rights record for licensed data.
Speak to an expertAI training dataThe template has three parts: general information, a list of data sources and data processing aspects. It covers data used at every stage of training, from pre-training to fine-tuning and alignment (adjusting a model's behaviour to human preferences).
| Section | What the provider must disclose |
|---|---|
| 1.1 and 1.2 Provider and model | Provider details, authorised representative, model and version names, any model it was built from, and the date it was placed on the EU market |
| 1.3 Modalities and size | For text, images, audio and video, the training data size within broad ranges (for text: under 1 billion tokens, 1 billion to 10 trillion, or over 10 trillion), the types of content, languages and the latest date of data collection |
| 2.1 Publicly available datasets | Each large public dataset by name and link; a dataset is large if it exceeds 3% of all public data used for that modality. Others are described generally |
| 2.2.1 Licensed datasets | Whether commercial licensing agreements were concluded with rights holders or their representatives, and the modalities covered |
| 2.2.2 Other private datasets | Private datasets from third parties not licensed under 2.2.1, such as data from intermediaries: listed if publicly known, otherwise described generally |
| 2.3 Crawled and scraped data | Crawler names, purposes and behaviour, collection period, the type of content and sites, and the top 10% of domain names by size of content scraped. Smaller companies list the top 5% or 1,000 domains, whichever is lower |
| 2.4 User data | Whether data from users' interactions with the model or the provider's other services was used for training |
| 2.5 Synthetic data | Whether AI-generated data was created to train the model, and which models generated it |
| 2.6 Other sources | Any other data, such as offline sources, digitised material and human-labelled datasets commissioned by the provider |
| 3.1 Text and data mining opt-outs | Measures taken to respect rights reservations, and whether the provider signed the Code of Practice for general-purpose AI models |
| 3.2 Illegal content | General description of measures to avoid or remove illegal content from the training data |
The template asks less about licensed data than about any other source. Section 2.2.1 asks only two questions: has the provider concluded transactional commercial licensing agreements with rights holders or their representatives, and which modalities (text, image, video, audio, other) do they cover. The Commission explains why: the rights holders are parties to those agreements, so they already know their content was used, and limited disclosure protects the provider's commercial confidentiality.
Data obtained from other third parties without a licence from the rights holders falls into Section 2.2.2. There, datasets must be listed by name if publicly known, and otherwise described by modality, nature of content (for example personal data or copyright-protected content) and language. Scraped data in Section 2.3 requires the most disclosure of all, including domain names.
Licensed data therefore reduces what a provider has to publish, provided it can show the licence is genuine. That means a written agreement with the rights holder or its authorised representative, and a record of the chain of permission. Each dataset supplied through our AI training data service carries a rights record for exactly this purpose.
Providers can only complete the template accurately if each data supplier gives them the right information. Annex XI of the Act also requires technical documentation covering the type and provenance (origin) of training data, how it was obtained and selected, and curation methods.
Our guides to data provenance and rights clearance and AI training data due diligence explain how to collect these records before data is bought.
A public summary of the content used to train a general-purpose AI model, which Article 53(1)(d) of the EU AI Act requires providers to publish using the European Commission's template.
Every provider placing a general-purpose AI model on the EU market, including open-source providers and providers of models with systemic risk. A company that modifies a model enough to become its provider covers its own modification.
It has applied since 2 August 2025 for new models and must be available when the model is placed on the market. Models placed on the market before 2 August 2025 must have a summary by 2 August 2027.
Only whether the provider concluded commercial licensing agreements with rights holders or their representatives, and which modalities they cover. No list of licensed datasets is required.
No. They list the top 10% of domain names by size of content scraped. Small and medium-sized companies list the top 5% or 1,000 domains, whichever is lower.
Yes. The exemption for free and open-source models in Article 53(2) does not cover the training data summary.
Since 2 August 2026 the AI Office can enforce the duty and request corrective measures. Fines can reach 3% of worldwide annual turnover or EUR 15 million, whichever is higher.
If the provider further trains its own model on new data that changes the summary, it updates the summary. A company that modifies another provider's model and becomes its provider publishes a summary of its modification data.
Not for the mandatory sections. Data a model looks up when answering is excluded unless the model actively learns from it, though providers may disclose it voluntarily.
Each point of law and each reported figure on this page was checked against these sources on 27 October 2026. This page is general information, not legal advice.
Tell us about the model, its training data and your reporting deadline.