EU AI Act Training Data Summary: Documenting Datasets for GPAI Compliance

Modern General-Purpose AI (GPAI) models are trained on extremely large and diverse data sets from open, licensed, and other sources. The scale and complexity of data collection and processing make it difficult to determine their origin, legal status, and compliance with legal requirements, particularly in the field of copyright.

To ensure an appropriate level of transparency and accountability, the European Union adopted the Regulation on Artificial Intelligence (EU AI Act), which establishes a comprehensive approach to regulating artificial intelligence systems based on the level of risk. A separate block of requirements is dedicated to providers of general-purpose AI models, for which a number of special obligations have been introduced regarding the documentation of the model development process and the training data used.

Regulatory basis for the requirement for a Training Data Summary

The legal basis for preparing a Training Data Summary is section 53 of the AI Act, which sets out the obligations of providers of general-purpose artificial intelligence models. In particular, in accordance with paragraph 1(d) of this section, providers must prepare and publish a sufficiently detailed summary of the content used to train the model, using a template developed by the AI ​​Office.

A significant part of modern GPAI models is trained on large arrays of text, graphics, audio, and video materials obtained from open or licensed sources. Therefore, the preparation of a Training Data Summary aims to ensure greater transparency regarding the origin of such data and to create conditions for the exercise of the rights of authors, publishers, and other rightholders, in particular in cases of the application of exceptions for text and data mining in accordance with EU law.

At the same time, the AI ​​Act does not directly establish detailed requirements for the structure or content of such a document in the Regulation's text. The specification of these requirements is entrusted to the AI ​​Office, which develops a standardized Training Data Summary template and relevant methodological recommendations. The Code is voluntary; its provisions can serve as evidence of proper implementation of the AI Act requirements and provide predictability during regulatory oversight.

Training Data Summary Objective

• Providing transparency — providing sufficient information about the content used to train a GPAI model, without having to disclose each individual dataset or work. This allows stakeholders to gain a general understanding of the origin and nature of the training data.
• Promoting compliance with copyright law — ensuring that copyright holders can understand which categories of content and sources may have been used to train a model, thus facilitating the exercise of their rights under European Union law.
• Balancing transparency and trade secret protection — The AI ​​Act does not require publication of a complete list of training data or the algorithms used to select them. The summary should be sufficiently informative to meet the objectives of transparency, while not revealing confidential information or trade secrets of the developer.
• Increasing accountability of GPAI model providers — the obligation to document training data encourages the implementation of appropriate procedures for data management, classification, accounting, and control at all stages of the model lifecycle.
• Facilitating effective regulatory oversight – the standardized Training Data Summary format makes it easier for the AI ​​Office and other competent authorities to assess compliance with the AI ​​Act requirements without analyzing each individual element of the training dataset.
• Enhancing trust in AI systems – openness about the sources and categories of training data helps build trust among users, businesses, government, and society at large, aligning with the overall aim of the AI ​​Act to develop safe, reliable, and human-centric artificial intelligence.

Legal requirements for the content of the Training Data Summary

In accordance with section 53(1)(d) of the AI ​​Act and the AI ​​Office recommendations, the Training Data Summary must contain information that allows for a general idea of ​​the content used to train the GPAI model. At the same time, the document is not a complete list of training data, but is of a generalized nature.

The main requirements for the content of the Training Data Summary are:

• Description of the main categories of training data — it is necessary to indicate what types of content were used to train the model (text materials, images, audio recordings, video, program code, structured data, etc.).
• Characteristics of data sources — information should be provided about the main sources of the training data, in particular licensed datasets, open databases, publicly available web resources, the developer's own data, or data obtained from third parties.
• A general description of data collection methods — the document should include information on the methods used to obtain the educational content, such as the use of open repositories, licensing agreements, web scraping, or other legal mechanisms for collecting information.
• Information on compliance with copyright law — it is appropriate to indicate that the requirements of European Union copyright law were taken into account when creating the educational dataset, including rules on text and data mining and mechanisms for accounting for rights reservations, if applicable.
• Sufficient level of detail — the information should be detailed enough to achieve transparency goals, but not so detailed as to require the disclosure of each individual work, dataset, or confidential elements of the educational process.
• Protection of trade secrets and confidential information — when preparing the Summary, it is necessary to ensure a balance between fulfilling the requirements of the AI ​​Act and protecting the legitimate interests of the provider, including non-disclosure of trade secrets, confidential information, and information, the disclosure of which could adversely affect the security or competitiveness of the model.
• Use of a standardized template — The Training Data Summary should be prepared in accordance with the template developed by the AI ​​Office, which provides a unified approach to fulfilling the requirements of the AI ​​Act and simplifies the assessment of the document by competent authorities.

Key Challenges in Preparing a Training Data Summary

FAQ

What are GPAI's obligations in relation to training data documentation?

GPAI obligations under the EU AI Act require providers to prepare a sufficiently detailed summary of the training content used for general-purpose AI models. This summary must provide transparency about the nature and origin of training data without requiring full disclosure of all datasets.

What is the role of the AI Office template in training data disclosure?

The AI Office template establishes a standardized structure for the Training Data Summary. It ensures that providers present information in a consistent format, facilitating regulatory oversight and comparability across different GPAI models.

What type of information is typically included in a Training Data Summary?

The summary generally includes categories of training data, primary data sources, and high-level descriptions of data collection methods. It focuses on aggregated information rather than detailed inventories of individual datasets.

How does training data disclosure relate to transparency requirements?

Training data disclosure provides a general understanding of how a GPAI model has been trained. It supports transparency by describing the composition and origin of training content in a way that is accessible to regulators and stakeholders.

In what way does the EU AI Act address copyright considerations?

The EU AI Act links training data transparency to copyright compliance by requiring a summary that helps identify the categories and sources of content used for training. This facilitates the assessment of compliance with EU copyright rules, including text and data mining provisions.

What level of detail is required for training data disclosure?

The required level of detail is aggregated and descriptive rather than exhaustive. The AI Act does not require listing every dataset or individual work; instead, it focuses on providing a sufficiently informative overview.

How is sensitive or confidential information treated in Training Data Summary preparation?

The framework allows for aggregated descriptions that avoid revealing trade secrets or proprietary datasets. Providers are expected to balance transparency obligations with the need to protect confidential business information.

What are common operational challenges in preparing Training Data Summaries?

Challenges include managing large-scale datasets, harmonizing heterogeneous data sources, maintaining accurate documentation across the training pipeline, and updating summaries after model modifications.

What functions does the AI Office perform regarding GPAI documentation?

The AI Office develops templates and guidance for implementing GPAI obligations, including the Training Data Summary. It also supports harmonized interpretation and application of the EU AI Act across member states.

How does Training Data Summary contribute to regulatory oversight?

The Training Data Summary provides regulators with structured information on the composition and sourcing of training data. This supports efficient supervision without requiring access to full datasets or internal training pipelines.