Cambridge Commentary on EU General-Purpose AI Law

Commentary

1. General remarks

1.1. Introduction

1Article 53 AI Act² sets out the key obligations for providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. of general-purpose AI (“GPAI”) models. These are models that display significant generality, can perform a wide range of distinct tasks, and can be integrated into a variety of downstream systems and applications (Article 3(63)).³ Like the other provisions in Chapter V of the AI Act, Article 53 was introduced at a later stage of the drafting process in response to the growing prominence of large language models.⁴ As the only provision that contains substantive obligations for GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. whose models do not present systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. ,⁵ it serves several distinct goals. First, it seeks to promote safe and trustworthy AI innovation in the European Union,⁶ for example by empowering regulators to request information on a GPAI model’s performance.⁷ These provisions are of key importance in determining what information the Commission, AI Office Article 3(47) AI Act: ‘AI Office’ means the Commission’s function of contributing to the implementation, monitoring and supervision of AI systems and general-purpose AI models, and AI governance, provided for in Commission Decision of 24 January 2024; references in this Regulation to the AI Office shall be construed as references to the Commission. and national competent authorities Article 3(48) AI Act: ‘national competent authority’ means a notifying authority or a market surveillance authority; as regards AI systems put into service or used by Union institutions, agencies, offices and bodies, references to national competent authorities or market surveillance authorities in this Regulation shall be construed as references to the European Data Protection Supervisor. may request from GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. – and, conversely, which information the latter may elect to retain for themselves.

2Second, it aims to ensure that downstream AI system Article 3(1) AI Act: ‘AI system’ means a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. who intend to integrate the GPAI model in their AI system Article 3(1) AI Act: ‘AI system’ means a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. have access to relevant information concerning its performance and compatibility.⁸ Third, it seeks to secure compliance with Union copyright law by requiring GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. to make publicly available the information used to train their models, thereby enabling affected individuals and rights holders to better safeguard their legal interests,⁹ and by arguably extending the scope of Union copyright law.¹⁰

3A preliminary general observation is that Recital 109 suggests that the relevant obligations should be applied more stringently to larger providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. . Particularly for obligations not explicitly scaled by provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. size, this implies that Article 53’s substantive provisions ought to be enforced more rigorously for such providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. . In practical terms, they may therefore be expected to produce the relevant documentation with greater detail and thoroughness.

1.2. Structure & overview

4This contribution broadly follows the structure of Article 53. It begins by situating the Article within the broader context of the AI Act before proceeding to a paragraph-by-paragraph analysis. Article 53(1) arguably contains the most substantive provision, outlining various documentation obligations for GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. . These include internal documentation – to be submitted to the AI Office Article 3(47) AI Act: ‘AI Office’ means the Commission’s function of contributing to the implementation, monitoring and supervision of AI systems and general-purpose AI models, and AI governance, provided for in Commission Decision of 24 January 2024; references in this Regulation to the AI Office shall be construed as references to the Commission. and national competent authorities Article 3(48) AI Act: ‘national competent authority’ means a notifying authority or a market surveillance authority; as regards AI systems put into service or used by Union institutions, agencies, offices and bodies, references to national competent authorities or market surveillance authorities in this Regulation shall be construed as references to the European Data Protection Supervisor. upon request – concerning the model’s training and testing (Article 53(1)(a)). In addition, Article 53(1)(b) contains requirements regarding information to be shared with downstream system providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. intending to incorporate the model into their systems.

5Further, Article 53(1) addresses significant aspects related to copyright, requiring GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. to develop a copyright policy that underscores compliance with Union copyright law (Article 53(1)(c)) and disclosure of information on the content used for model training (Article 53(1)(d)). The analysis then turns to Article 53(2), which introduces a partial exception from certain documentation and information requirements for providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. of GPAI models that do not pose systemic risks Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. . The Article 53(3) discussion considers the obligation to cooperate with the Commission and national competent authorities Article 3(48) AI Act: ‘national competent authority’ means a notifying authority or a market surveillance authority; as regards AI systems put into service or used by Union institutions, agencies, offices and bodies, references to national competent authorities or market surveillance authorities in this Regulation shall be construed as references to the European Data Protection Supervisor. .

6The next three sections examine compliance mechanisms, including codes of practice (Article 53(4)) and the use of delegated acts (Articles 53(5) and (6)). The discussion concludes with Article 53(7), which refers to the confidentiality obligations set out in Article 78 in relation to any information communicated pursuant to Article 53.

2. Substance

2.1. Article 53(1): Information & documentation

7Article 53(1) sets out the substantive obligations for GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. , distinguishing four main categories of information that must be provided or made accessible. First, GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. must, upon request, supply certain documentation to the AI Office Article 3(47) AI Act: ‘AI Office’ means the Commission’s function of contributing to the implementation, monitoring and supervision of AI systems and general-purpose AI models, and AI governance, provided for in Commission Decision of 24 January 2024; references in this Regulation to the AI Office shall be construed as references to the Commission. and national competent authorities Article 3(48) AI Act: ‘national competent authority’ means a notifying authority or a market surveillance authority; as regards AI systems put into service or used by Union institutions, agencies, offices and bodies, references to national competent authorities or market surveillance authorities in this Regulation shall be construed as references to the European Data Protection Supervisor. (Article 53 (1)(a)). Second, they are required to provide specific information and documentation to AI system Article 3(1) AI Act: ‘AI system’ means a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. intending to integrate the GPAI model into their systems ((1)(b)). Third, they must establish a copyright compliance policy ((1)(c)). Fourth and finally, they are obliged to make publicly available a summary of the data used to train the model ((1)(d)). As discussed further below, all such information must be kept up to date.

8Recital 109 indicates that persons who develop or use GPAI models for non-professional or scientific research purposes are encouraged rather than obliged to comply with the relevant documentation and transparency rules for GPAI models.¹¹ This exclusion is, however, not reiterated in the text of Article 53, but it could be read to follow from Article 2.¹² In any case, it is key to underscore that this exception – as well as the open-source exception (Article 53(2)) discussed below¹³ – do not go as far as to exempt the provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. from some of the underlying requirements, including the need to respect Union copyright law, which features its own independent exceptions.¹⁴

2.1.1. Article 53(1)(a): Documentation for AI Office Article 3(47) AI Act: ‘AI Office’ means the Commission’s function of contributing to the implementation, monitoring and supervision of AI systems and general-purpose AI models, and AI governance, provided for in Commission Decision of 24 January 2024; references in this Regulation to the AI Office shall be construed as references to the Commission. and national competent authorities Article 3(48) AI Act: ‘national competent authority’ means a notifying authority or a market surveillance authority; as regards AI systems put into service or used by Union institutions, agencies, offices and bodies, references to national competent authorities or market surveillance authorities in this Regulation shall be construed as references to the European Data Protection Supervisor.

2.1.1.1. All GPAI models

9Article 53(1)(a) requires GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. to document and maintain up-to-date technical information concerning the model, its training and testing processes, and model evaluation. This documentation must, at a minimum, include the information specified in Annex XI and must be made available, upon request, to the AI Office Article 3(47) AI Act: ‘AI Office’ means the Commission’s function of contributing to the implementation, monitoring and supervision of AI systems and general-purpose AI models, and AI governance, provided for in Commission Decision of 24 January 2024; references in this Regulation to the AI Office shall be construed as references to the Commission. and/or the national competent authorities Article 3(48) AI Act: ‘national competent authority’ means a notifying authority or a market surveillance authority; as regards AI systems put into service or used by Union institutions, agencies, offices and bodies, references to national competent authorities or market surveillance authorities in this Regulation shall be construed as references to the European Data Protection Supervisor. . The latter refers to the authorities designated by Member States pursuant to Article 70 of the AI Act. We will briefly reconsider the procedural implications for national competent authorities Article 3(48) AI Act: ‘national competent authority’ means a notifying authority or a market surveillance authority; as regards AI systems put into service or used by Union institutions, agencies, offices and bodies, references to national competent authorities or market surveillance authorities in this Regulation shall be construed as references to the European Data Protection Supervisor. , the AI Office Article 3(47) AI Act: ‘AI Office’ means the Commission’s function of contributing to the implementation, monitoring and supervision of AI systems and general-purpose AI models, and AI governance, provided for in Commission Decision of 24 January 2024; references in this Regulation to the AI Office shall be construed as references to the Commission. and the Commission below.¹⁵

10Annex XI provides further detail on the required information, dividing it into two sections. The first outlines the information applicable to all GPAI models, while the second sets out additional requirements for models presenting systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. .

11In examining these obligations, we will frequently reference the corresponding provisions of the relevant code of practice.¹⁶ Although adherence to such codes does not, in and of itself, constitute conclusive evidence of compliance with the AI Act in general,¹⁷ such codes do provide useful guidance as to how the AI Office Article 3(47) AI Act: ‘AI Office’ means the Commission’s function of contributing to the implementation, monitoring and supervision of AI systems and general-purpose AI models, and AI governance, provided for in Commission Decision of 24 January 2024; references in this Regulation to the AI Office shall be construed as references to the Commission. may interpret the relevant provisions.¹⁸ Furthermore, once the AI Office Article 3(47) AI Act: ‘AI Office’ means the Commission’s function of contributing to the implementation, monitoring and supervision of AI systems and general-purpose AI models, and AI governance, provided for in Commission Decision of 24 January 2024; references in this Regulation to the AI Office shall be construed as references to the Commission. and the Board deem a code of practice adequate (Article 56(6) AI Act), such a code of practice can be used to demonstrate compliance with the relevant obligations.¹⁹ It is thus key to note that the European Commission²⁰ and AI Board²¹ have confirmed the adequacy of the 2025 Code of Practice on 1 August 2025.

12As per Section 1. of Annex XI, all GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. are required to document at least ‘A general description of the general-purpose AI model Article 3(63) AI Act: ‘general-purpose AI model’ means an AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications, except AI models that are used for research, development or prototyping activities before they are placed on the market. including:’

‘(a) the tasks that the model is intended to perform and the type and nature of AI systems Article 3(1) AI Act: ‘AI system’ means a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. in which it can be integrated;’

13The tasks the model is intended to perform – ‘intended uses’ in the Code of Practice Transparency form²² – refers to the specific functions the model is designed to carry out, such as ‘productivity enhancement, translation, creative content generation, data analysis, data visualisation, programming assistance, scheduling, customer support, variety of natural language tasks’.²³ The type and nature of AI systems Article 3(1) AI Act: ‘AI system’ means a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. in which the model can be integrated refers to the category and characteristics of those systems – such as ‘autonomous systems, conversational assistants, decision support systems, creative AI systems Article 3(1) AI Act: ‘AI system’ means a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. , predictive systems, cybersecurity, surveillance, or human-AI collaboration’.²⁴ While it has been argued that the tasks of a model will generally correspond to the types of systems into which it may be integrated,²⁵ the Code of Practice form does distinguish between the two.²⁶ While the wording of Article 53(1)(a) appears to require only a relatively abstract description,²⁷ the Code of Practice calls for a more detailed approach, suggesting a recommended length of 200 words for intended uses and 300 words for the type and nature of AI systems Article 3(1) AI Act: ‘AI system’ means a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. in which the GPAI model can be integrated.²⁸ Interestingly, it suggests that those elements do not need to be described positively but may also be adhered to negatively by describing the restricted or prohibited uses or the type of nature of AI systems Article 3(1) AI Act: ‘AI system’ means a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. in which the GPAI model should not be integrated, respectively.²⁹

‘(b) the acceptable use policies applicable’

14The relevant use policies are likely to concern restrictions on the use of the GPAI model, for example to prevent its deployment in the commission of criminal offences³⁰ or in breach of copyright law. Such policies may also impose limitations on the use of the model’s outputs. GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. may, for instance, restrict use to non-commercial purposes unless a specific licence is obtained. The model form annexed to the Code of Practice suggests that providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. should indicate whether such a policy exists. This, together with the use of the term ‘applicable’, may be taken to imply that the absence of an acceptable use policy is permissible.

15The question of whether the modification of the model can be restricted in such use policies is discussed elsewhere in this commentary.³¹

‘(c) the date of release and methods of distribution’

16The notion of ‘release’ differs from the concept of ‘ placing on the market Article 3(9) AI Act: ‘placing on the market’ means the first making available of an AI system or a general-purpose AI model on the Union market. ’ as defined in Article 3(9)³² in that release does not require the model to be made available on the Union market.³³ In line with arguments by some authors that model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. should document the release date of the model following each modification or technical change,³⁴ the form provided in the Code of Practice requires GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. to indicate the release date of the current model, the release date on the Union market,³⁵ and any ‘model dependencies’ – meaning an overview of previous versions of the model and their respective release dates. The term ‘Union market release date’ arguably corresponds more closely to the notion of ‘ placing on the market Article 3(9) AI Act: ‘placing on the market’ means the first making available of an AI system or a general-purpose AI model on the Union market. ’ in Article 3(9).

17 Providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. must also disclose the ‘methods of distribution’. The model form in the Code of Practice lists several examples – ‘e.g. enterprise or subscription-based access through existing software suites or enterprise-specific solutions; public or subscription-based access through an API; public or proprietary access through integrated development environments, device-specific applications or firmware, open-source repositories’.³⁶ For each distinct method of distribution, a link (where available) to information about how the model can be accessed should be provided, along with a brief description of the level of the access (e.g. ‘weights-level access’ or ‘black-box access’).³⁷

‘(d) the architecture and number of parameters’

18With respect to the architecture, a brief description is required – the Code of Practice form recommends a length of approximately 20 words (e.g. ‘a transformer architecture’).³⁸ The model provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. must also disclose the total number of parameters. The form mandates the use of at least two significant figures (e.g. ‘7.3×10¹⁰’) and further requires the provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. to indicate the range within which this number falls, selecting from a set of predefined options.³⁹

‘(e) the modality (e.g. text, image) and format of inputs and outputs’

19The information concerning the modality and format of inputs and outputs refers to the types of data a model can process or generate – for example, text, images, audio, video, or other types. It should also include the maximum input and output file sizes,⁴⁰ where such limits are defined.⁴¹

‘(f) the licence’

20The information concerning the licence should, naturally, specify the licence under which access to the model is granted, or, alternatively, indicate that no such licence exists.

21Annex XI further requires the model developer to provide ‘relevant information on the development process, including the following elements:’

‘(a) the technical means (e.g. instructions of use, infrastructure, tools) required for the general-purpose AI model Article 3(63) AI Act: ‘general-purpose AI model’ means an AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications, except AI models that are used for research, development or prototyping activities before they are placed on the market. to be integrated in AI systems Article 3(1) AI Act: ‘AI system’ means a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. ’

22The model provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. should specify the infrastructure requirements and necessary tools, along with any relevant operating instructions needed for the system provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. to integrate the GPAI model successfully within the intended system. This includes a description of the required hardware (if any – none may be needed if the model is accessed via an API) and software.

‘(b) the design specifications of the model and training process, including training methodologies and techniques, the key design choices including the rationale and assumptions made; what the model is designed to optimise for and the relevance of the different parameters, as applicable’

23The design specifications of the training process should be described in reasonable detail, covering the elements listed. The Code of Practice offers an illustrative example: ‘the model is initialized with randomly selected weights and optimised using gradient-based optimisation via the Adam optimiser in two steps. First, the model is trained to predict the next word on a large pre-training corpus using the cross-entropy loss, passing over the data for a single epoch. Second, the model is post-trained on a dataset of human preferences for 10 epochs to align the model with human values and make it more useful in responding to user prompts.’⁴² This level of detail goes beyond a mere characterisation of the training method (e.g. ‘supervised learning’) and should also include a description of the key design choices made during model training, accompanied by a rationale for their adoption.

‘(c) information on the data used for training, testing and validation, where applicable, including the type and provenance of data and curation methodologies (e.g. cleaning, filtering, etc.), the number of data points, their scope and main characteristics; how the data was obtained and selected as well as all other measures to detect the unsuitability of data sources and methods to detect identifiable biases, where applicable’

24The information regarding the data used for training, testing and validation should include the type of data (e.g. text, image, video or audio) and how it was obtained (e.g. via web-crawling, publicly available sources, synthetic data, user data), as well as the specific means and criteria used for its collection and selection. This includes the methods and resources employed for data annotation as well as any tools or techniques used to generate synthetic data. Where data has been sourced from third parties, the GPAI model provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. should explain how the rights to use such data were acquired, unless this information has already been made publicly available under Article 53(1)(d).⁴³

25The number of data points used in training, testing and validation should be indicated, along with the relevant unit (‘e.g. tokens, documents, images, hours of video or frames’⁴⁴). The Code of Practice requires this information to be reported with at least one or two significant figures (e.g. ‘3×10¹³ tokens’), depending on whether the data is submitted to national competent authorities Article 3(48) AI Act: ‘national competent authority’ means a notifying authority or a market surveillance authority; as regards AI systems put into service or used by Union institutions, agencies, offices and bodies, references to national competent authorities or market surveillance authorities in this Regulation shall be construed as references to the European Data Protection Supervisor. or the AI Office Article 3(47) AI Act: ‘AI Office’ means the Commission’s function of contributing to the implementation, monitoring and supervision of AI systems and general-purpose AI models, and AI governance, provided for in Commission Decision of 24 January 2024; references in this Regulation to the AI Office shall be construed as references to the Commission. .⁴⁵

26The scope and principal characteristics of the data should also be described. This includes the domain (e.g. ‘healthcare, science, law’ or scientific data), geographical origin (e.g. European, global, or US-based data), and the language(s) of the data (in the case of text, audio or video).⁴⁶ If applicable, the description should also include the modality coverage of the dataset.⁴⁷ Additionally, the information must detail the measures taken to identify potential biases – specifically methods used during data acquisition or processing – and describe how the provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. assessed the potential unsuitability of the data sources. This latter obligation extends beyond legal compliance (e.g. exclusion of unlawfully processed personal data Article 3(50) AI Act: ‘personal data’ means personal data as defined in Article 4, point (1), of Regulation (EU) 2016/679. or non-consensual intimate imagery) and also encompasses data that may be unsuitable for robust model training given the intended use of the model.⁴⁸

‘(d) the computational resources used to train the model (e.g. number of floating point operations), training time, and other relevant details related to the training’

27The information regarding the computational resources used to train the model should include the total duration of the training process. The Code of Practice distinguishes between national competent authorities Article 3(48) AI Act: ‘national competent authority’ means a notifying authority or a market surveillance authority; as regards AI systems put into service or used by Union institutions, agencies, offices and bodies, references to national competent authorities or market surveillance authorities in this Regulation shall be construed as references to the European Data Protection Supervisor. and the AI Office Article 3(47) AI Act: ‘AI Office’ means the Commission’s function of contributing to the implementation, monitoring and supervision of AI systems and general-purpose AI models, and AI governance, provided for in Commission Decision of 24 January 2024; references in this Regulation to the AI Office shall be construed as references to the Commission. with respect to the required level of detail – for example, indicating the number of months for the former and specifying the duration in wall-clock days for the latter.⁴⁹ The Code of Practice indicates that a similar distinction applies to the reporting of computational resources used: for national competent authorities Article 3(48) AI Act: ‘national competent authority’ means a notifying authority or a market surveillance authority; as regards AI systems put into service or used by Union institutions, agencies, offices and bodies, references to national competent authorities or market surveillance authorities in this Regulation shall be construed as references to the European Data Protection Supervisor. , this should be recorded in floating point operations (FLOPs) to the correct order of magnitude, whereas for the AI Office Article 3(47) AI Act: ‘AI Office’ means the Commission’s function of contributing to the implementation, monitoring and supervision of AI systems and general-purpose AI models, and AI governance, provided for in Commission Decision of 24 January 2024; references in this Regulation to the AI Office shall be construed as references to the Commission. , the figure must be provided with at least two significant figures.⁵⁰

28The Code of Practice further clarifies that a description of the methodology used to measure or estimate the volume of computation should be included, in the absence of a delegated act adopted under Article 53(5).⁵¹

29To support compliance with these requirements, the Commission is empowered to adopt delegated acts pursuant to Article 97 of the AI Act. These acts may establish harmonised methods for calculating and measuring computational resources, training duration, and other relevant aspects of model training.⁵² Such delegated acts can supplement or amend non-essential elements of the AI Act and are thus binding on those subject to its requirements, including GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. .⁵³

‘(e) known or estimated energy consumption of the model’

30Lastly, the model provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. should provide information about the energy consumption of the model. Annex XI clarifies that ‘where the energy consumption of the model is unknown, the energy consumption may be based on information about computational resources used.’ While the past tense of that provision might be interpreted to imply that this energy use only refers to the training of the model, some authors argued that it should also include the model’s energy consumption during inference.⁵⁴ In any case, such an interpretation cannot fully be ruled out as the energy consumption for training (past tense) may be indicative of that during (future) inference.⁵⁵

31The Code of Practice takes a clearer position, as it provides that the energy consumption of the model does not only relate to the energy used during model deployment in a potential AI system Article 3(1) AI Act: ‘AI system’ means a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. (‘energy consumption during inference’), but also to the amount of energy that was used for model training (‘energy used for training’),⁵⁶ as had been previously recommended by some authors. For the former, the model provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. should detail the amount of computation used for inference measured in FLOP with at least two significant figures, per the Code of Practice.⁵⁷ Moreover, in the absence of a delegated act in accordance with Article 53(5),⁵⁸ the model provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. should provide a description of the computational tasks and the hardware used to measure or estimate the inference-time computation, and that used for any estimations of the energy consumption of the model.⁵⁹

32The amount of energy used for training should be estimated and reported in megawatt-hours with at least two significant figures according to the Code of Practice, though the provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. can also indicate that they do not have the necessary information (e.g. due to unknowns related to hardware provided by an external provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. ) to make a reasonable assessment.⁶⁰ Here too, the Code of Practice requires the provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. to clarify the methodology used for their measurements.⁶¹ In the absence of a delegated act as per Article 53(5), the provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. should estimate this on the basis of computational resources required for model training.⁶² As before, it is permissible to indicate that no assessment could be made due to missing information, though the provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. should specify the information that they lack.⁶³ If the model results from the modification or fine-tuning of another GPAI model, it is reasonable to estimate its energy consumption using known estimates or information about the parent model, in the absence of more specific data.

33As for the computational resources used to train the model,⁶⁴ the Commission is empowered to adopt delegated acts in application of Article 97 AI Act⁶⁵ to facilitate compliance.⁶⁶ These can support the calculation and measurement methods to support comparable and verifiable documentation of the energy consumption of the model. An approach might be for the Commission to establish that the environmental impact of GPAI models should be estimated using the relevant servers’ or data centres’ energy use and their efficiency ratings (the power usage effectiveness, PUE⁶⁷).⁶⁸

34Interestingly, the extent of these obligations and the information required should be ‘appropriate’ to the ‘size and risk Article 3(2) AI Act: ‘risk’ means the combination of the probability of an occurrence of harm and the severity of that harm. profile of the model’.⁶⁹ Notably, this is the only explicit reference to the ‘size’ of the model in the AI Act. While ‘size’ most likely denotes the number of parameters of the model, it is notable that the AI Act explicitly refers to the ‘number of parameters’ elsewhere – for example, in Recital (98), Annex XI(1)(d), and Annex XII(1)(f) – without using the term ‘size’, which could imply that ‘size’ should be read differently, referring more broadly to, for example, the volume of training data Article 3(29) AI Act: ‘training data’ means data used for training an AI system through fitting its learnable parameters. ,⁷⁰ the computational resources required,⁷¹ the model’s complexity⁷² or the breadth of its capabilities⁷³.

35Likewise, the AI Act neither defines the notion of a ‘ risk Article 3(2) AI Act: ‘risk’ means the combination of the probability of an occurrence of harm and the severity of that harm. profile’ nor employs the term elsewhere in its text. The Act does not clarify whether this should be understood, in the context of GPAI models, as referring to the (potential) systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. presented by the model or, rather, whether it relates to the likelihood that the model will be incorporated into high-risk systems within the meaning of Article 6. The latter interpretation may be defensible, given the broader relevance of various Annex XII elements for high-risk systems, discussed below.⁷⁴ However, the most natural reading is likely the more general one, which would have ‘ risk Article 3(2) AI Act: ‘risk’ means the combination of the probability of an occurrence of harm and the severity of that harm. profile’ refer to the risk Article 3(2) AI Act: ‘risk’ means the combination of the probability of an occurrence of harm and the severity of that harm. definition set out in Article 3(2).⁷⁵ This definition is not confined to the categorisation requirements of high-risk systems or systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. models, but rather entails a contextualised assessment of the model’s potential for misuse or harm, as well as the likelihood of such incidents, taking into account the scale and scope of deployment, the model’s capabilities, the vulnerability of affected parties, and any available information about past incidents or vulnerabilities.⁷⁶

2.1.1.2. GPAI Models presenting systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain.

36Specifically for providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. of GPAI models with systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. , who are also subject to Article 55,⁷⁷ Section 2. of Annex XI lists additional information that providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. of GPAI models with systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. must document and provide to the AI Office Article 3(47) AI Act: ‘AI Office’ means the Commission’s function of contributing to the implementation, monitoring and supervision of AI systems and general-purpose AI models, and AI governance, provided for in Commission Decision of 24 January 2024; references in this Regulation to the AI Office shall be construed as references to the Commission. and national competent authorities Article 3(48) AI Act: ‘national competent authority’ means a notifying authority or a market surveillance authority; as regards AI systems put into service or used by Union institutions, agencies, offices and bodies, references to national competent authorities or market surveillance authorities in this Regulation shall be construed as references to the European Data Protection Supervisor. upon request. This information largely concerns the methods used to evaluate the model’s performance, the measures taken to identify and address model vulnerabilities, and the interactions between various software components.

37For the interpretation of these requirements, it is useful to highlight that the AI Act does not explicitly refer to Annex XI outside Articles 53 and 54, with the latter only imposing an obligation to verify the relevant documentation.⁷⁸ This is remarkable, as some of the obligations found in Article 55(1)(a), (b), and (d) could,⁷⁹ seemingly, be interpreted to closely relate to Annex XI Section 2’s documentation requirements. This applies, for example, to the requirement that model developers should ‘perform model evaluation in accordance with standardised protocols and tools reflecting the state-of-the-art, including conducting and documenting adversarial testing of the model with a view to identifying and mitigating systemic risks Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. ’ (Article 55(1)(a)), which arguably relates to ‘a detailed description of the evaluation strategies’ discussed below.

38A more extensive interpretation of Annex XI Section 2. would strengthen the link with Article 55 and thus capture Article 55(1)’s obligations, resulting in more extensive documentation under that Section of Annex XI. An important policy implication of that interpretation is that it would require providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. of GPAI models with systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. to document the assessments made under Article 55 as well – some of which seem to be captured more explicitly by the wording of Annex XI Section 2. As a result, that information could be requested both by the AI Office Article 3(47) AI Act: ‘AI Office’ means the Commission’s function of contributing to the implementation, monitoring and supervision of AI systems and general-purpose AI models, and AI governance, provided for in Commission Decision of 24 January 2024; references in this Regulation to the AI Office shall be construed as references to the Commission. and the national competent authorities Article 3(48) AI Act: ‘national competent authority’ means a notifying authority or a market surveillance authority; as regards AI systems put into service or used by Union institutions, agencies, offices and bodies, references to national competent authorities or market surveillance authorities in this Regulation shall be construed as references to the European Data Protection Supervisor. on the basis of Article 53(1)(a)’s reference to Annex XI.

39This interpretation appears most consistent with the objectives of the AI Act in relation to these documentation requirements – namely, that this documentation should inform the AI Office Article 3(47) AI Act: ‘AI Office’ means the Commission’s function of contributing to the implementation, monitoring and supervision of AI systems and general-purpose AI models, and AI governance, provided for in Commission Decision of 24 January 2024; references in this Regulation to the AI Office shall be construed as references to the Commission. (and national competent authorities Article 3(48) AI Act: ‘national competent authority’ means a notifying authority or a market surveillance authority; as regards AI systems put into service or used by Union institutions, agencies, offices and bodies, references to national competent authorities or market surveillance authorities in this Regulation shall be construed as references to the European Data Protection Supervisor. ), upon request, regarding the potential risks Article 3(2) AI Act: ‘risk’ means the combination of the probability of an occurrence of harm and the severity of that harm. and capabilities of these models. Nevertheless, the need for legal certainty could be raised as a counterargument, as the Act does not explicitly set out this link between Annex XI Section 2. and Article 55(1). That argument would, however, be largely countered by the broad language in Annex XI Section 2., together with Article 55 and the AI Office Article 3(47) AI Act: ‘AI Office’ means the Commission’s function of contributing to the implementation, monitoring and supervision of AI systems and general-purpose AI models, and AI governance, provided for in Commission Decision of 24 January 2024; references in this Regulation to the AI Office shall be construed as references to the Commission. ’s authority under Article 91 to request relevant information, which together provide a sufficiently clear legal basis for requiring disclosure. In other words, even without an explicit textual cross-reference between Article 55 and Annex XI Section 2., the combined provisions establish a sufficiently predictable and clear framework.

40If, however, Section 2. of Annex XI were to be given a more restrictive interpretation – which would exclude any or some elements found in Article 55 that Annex XI does not explicitly reference – the direct result could be that Article 55 obliges providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. of GPAI models with systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. to carry out certain assessments, such as an assessment of ‘ systemic risks Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. at Union level’, without requiring them to document those assessments. One could argue, in such a scenario, that national competent authorities Article 3(48) AI Act: ‘national competent authority’ means a notifying authority or a market surveillance authority; as regards AI systems put into service or used by Union institutions, agencies, offices and bodies, references to national competent authorities or market surveillance authorities in this Regulation shall be construed as references to the European Data Protection Supervisor. might not be able to request that information, as such authority is not expressly provided for in the text of Article 55. The AI Office Article 3(47) AI Act: ‘AI Office’ means the Commission’s function of contributing to the implementation, monitoring and supervision of AI systems and general-purpose AI models, and AI governance, provided for in Commission Decision of 24 January 2024; references in this Regulation to the AI Office shall be construed as references to the Commission. would still be able to request information on the basis of Article 91(1), though arguably, in this interpretation, there might be little to request, since Article 55 does not explicitly require providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. of GPAI models with systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. to document those assessments.

41The Code of Practice seems to support the more extensive ‘linked’ interpretation, as it exclusively considers Annex XI Section 2’s obligations together with the Article 55(1) obligations. Moreover, the Code of Practice imposes documentation for various of the obligations found in Article 55⁸⁰ – with some exceptions.⁸¹ Interestingly, the Code of Practice also indicates that the information that should be provided on the basis of Article 53, and thus Annex XI Section 2., is more detailed than the information required under Article 55 alone.⁸² While the text of the AI Act leaves open whether the evaluations and tests to be documented under Section 2. of Annex XI necessarily pertain to the systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. assessments mandated by Article 55(1), given that Annex XI does not mention systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. , that interpretation is strongly implied by the Code of Practice, which treats the Annex’s obligations (to be read as part of Article 53(1)) in conjunction with those arising under Article 55(1).

42The information that should be documented pursuant to Annex XI Section 2. entails:

‘1. A detailed description of the evaluation strategies, including evaluation results, on the basis of available public evaluation protocols and tools or otherwise of other evaluation methodologies. Evaluation strategies shall include evaluation criteria, metrics and the methodology on the identification of limitations.’

43The provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. of a GPAI model with systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. should describe the benchmarks used to evaluate the model’s performance. Interestingly, Annex XI does not go so far as to explicitly require testing for specific use cases or capabilities. Various public evaluation tools are available⁸³ that can be used to assess model performance across a range of aspects and applications. Examples include benchmarks for question answering⁸⁴, reasoning puzzles⁸⁵, coding challenges⁸⁶, and safety evaluations⁸⁷.

44 Providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. are not, however, obliged to use established or publicly available benchmarks under Section 2. of Annex XI. Arguably, though, they should document additional information regarding their evaluation procedure if unconventional or non-public benchmarks are employed. Particularly in such cases, it is important to elaborate on the evaluation criteria, metrics and methodology for identifying limitations. Evaluation criteria might include the accuracy on question answering⁸⁸, the extent of bias displayed, or the frequency of safety refusals.⁸⁹ Relevant metrics could then include accuracy percentages, bias scores, or refusal rates. The provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. should also set out how specific limitations in the model’s capabilities were identified through the evaluation process, and the procedures used to uncover them.

45A key observation, based on reporting practices prior to the application and enforcement of the AI Act, is that there is little consistency in what providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. disclose and that their reports tend to lack sufficient information – sometimes using vague claims such as ‘above human average’ – rather than reporting in-depth testing results.⁹⁰ Moreover, providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. might be inclined to withhold certain information for competitive or marketing reasons, or may selectively highlight other data for similar purposes. Nevertheless, as discussed further below,⁹¹ the requirements under Article 53 and Annex XI should arguably be interpreted as requiring extensive documentation, as such an interpretation most closely aligns with the rationale of these requirements.

‘2. Where applicable, a detailed description of the measures put in place for the purpose of conducting internal and/or external adversarial testing (e.g. red-teaming), model adaptations, including alignment and fine-tuning.’

46There is a similar requirement to document adversarial testing and model adaptations. Here, too, the provision does not appear to go so far as to prescribe such testing – given the phrase ‘where applicable’ – but merely requires its documentation if it has taken place, for example on the basis of an obligation found in Article 55(1)(a).⁹²

47Adversarial testing refers to exercises where designated users deliberately attempt to elicit harmful, unsafe or otherwise undesirable outputs from the model in order to uncover risks Article 3(2) AI Act: ‘risk’ means the combination of the probability of an occurrence of harm and the severity of that harm. and weaknesses.⁹³ For internal adversarial testing, this involves in-house teams red-teaming or attempting to break the model using challenging inputs. Depending on the level of detail required, which arguably depends on the size of the provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. ,⁹⁴ this documentation should include which scenarios these teams tested (such as prompts designed to elicit biased responses or to obtain confidential or harmful outputs), as well as the findings. For external adversarial testing, external experts or third parties are engaged to perform similar evaluations. Here, the documentation should arguably specify who was involved and their relevant expertise and should detail the scope and results of the testing.⁹⁵

48Model adaptations,⁹⁶ including alignment and fine-tuning, refer to modifications made to the model following its initial testing. It is common practice to first train a GPAI model using a large dataset before applying further guidelines or restrictions to better align the model’s output with human values or improve safety.⁹⁷ Techniques that may be employed – and should be described if so – include reinforcement learning with human feedback, or additional training with specialised or curated data.⁹⁸

49While Annex XI Section 2(2) only mandates these descriptions ‘where applicable’, the implementation of such measures is widespread.⁹⁹ One could also argue that they are required by Article 55(1).¹⁰⁰ Furthermore, if the model is intended for deployment in a high-risk system, it could be argued that such modification is generally required (for example in the context of Article 9 AI Act) to enable a downstream system provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. to meet their obligations.¹⁰¹ Nevertheless, it is conceivable that model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. could restrict themselves to their obligations under Article 55 and delegate any additional fine-tuning to providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. intending to implement the model in their high-risk system, thus limiting the information that needs to be documented according to Annex XI Section 2..

50Here, too, these elements should be described in detail, rather than superficially. This is discussed in more detail below.¹⁰²

‘3. Where applicable, a detailed description of the system architecture explaining how software components build or feed into each other and integrate into the overall processing.’

51The third part of Section 2. similarly employs the terminology ‘where applicable’. Notably, the Code of Practice subjects Article 3(58) AI Act: ‘subject’, for the purpose of real-world testing, means a natural person who participates in testing in real-world conditions. this requirement to an alternatively phrased condition – namely, ‘insofar as the Signatory is aware of such information’.¹⁰³ Thus, it can generally be assumed that this information is required, except where the provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. of the GPAI model with systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. is entirely unaware of it – which would, as discussed more extensively below,¹⁰⁴ significantly restrict their ability to market the model.

52It is interesting that Annex XI Section 2. refers to the ‘system architecture’ of the GPAI model with systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. , rather than the ‘model architecture’.¹⁰⁵ While the Code of Practice does discuss the need to document the ‘model’s architecture’,¹⁰⁶ it subsequently shifts the focus to AI systems Article 3(1) AI Act: ‘AI system’ means a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. by requiring ‘a detailed description of how the model is integrated into AI systems Article 3(1) AI Act: ‘AI system’ means a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. , explaining how software components build or feed into each other and integrate into the overall processing […]’,¹⁰⁷ which seems to indicate that Annex XI Section 2(3) should be understood as largely referring to systems implementing the model, rather than to the model itself. This interpretation is more consistent with the phrases ‘where applicable’ and ‘insofar as the Signatory is aware of such information’, as there may either be no system implementation (yet), or the provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. of the GPAI model with systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. may not be aware of the specific way in which the model was implemented in a system. It is also more consistent with Annex XI, Section 1(1)(d), which requires documentation of the model architecture, as it would be illogical for both provisions to mandate identical information.

53As such, this provision requires the provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. of the GPAI model with systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. to provide information about the way in which the model is integrated in a specific system. That description must also address the interaction between these components, which should arguably cover what input is fed to the GPAI model with systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. and how the system processes that model’s output.

54Interestingly, the wording of Annex XI Section 2(3) mirrors that of Annex IV(2)(c),¹⁰⁸ which sets out the technical documentation required for high-risk systems (Article 11). In that sense, such information is likely to be required for most GPAI models, given the interaction with high-risk system requirements described below.¹⁰⁹

2.1.1.3. Modalities and exceptions

55Regarding the modalities of these obligations, a first aspect concerns how long providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. should preserve the relevant information. The Code of Practice chapter on safety and security indicates that GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. should retain the relevant information for ten years.¹¹⁰

56As discussed more extensively above regarding Annex XI Section 2. and GPAI models with systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. ,¹¹¹ there are some arguments that would support a broad interpretation of the relevant documentation requirements. A broad and thorough approach would best enable the AI Office Article 3(47) AI Act: ‘AI Office’ means the Commission’s function of contributing to the implementation, monitoring and supervision of AI systems and general-purpose AI models, and AI governance, provided for in Commission Decision of 24 January 2024; references in this Regulation to the AI Office shall be construed as references to the Commission. and national competent authorities Article 3(48) AI Act: ‘national competent authority’ means a notifying authority or a market surveillance authority; as regards AI systems put into service or used by Union institutions, agencies, offices and bodies, references to national competent authorities or market surveillance authorities in this Regulation shall be construed as references to the European Data Protection Supervisor. to exercise their competences under the AI Act. This interpretation is further supported – even under a restrictive reading – by the duty to cooperate, discussed below.¹¹² This is especially pertinent to the level of detail required in the documentation. It is therefore insufficient, for instance, to describe a model’s performance only in relative terms (‘below-human capabilities’) while omitting the precise test score. That said, it is equally clear that the extent of documentation and level of detail also depend on the size of the provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. of the GPAI model.¹¹³

57The Transparency¹¹⁴ and Safety and Security¹¹⁵ chapters of the Code of Practice further clarify that this information, if required, does not generally¹¹⁶ need to be shared proactively or made publicly available, but should only be provided to the AI Office Article 3(47) AI Act: ‘AI Office’ means the Commission’s function of contributing to the implementation, monitoring and supervision of AI systems and general-purpose AI models, and AI governance, provided for in Commission Decision of 24 January 2024; references in this Regulation to the AI Office shall be construed as references to the Commission. and national competent authorities Article 3(48) AI Act: ‘national competent authority’ means a notifying authority or a market surveillance authority; as regards AI systems put into service or used by Union institutions, agencies, offices and bodies, references to national competent authorities or market surveillance authorities in this Regulation shall be construed as references to the European Data Protection Supervisor. upon request.

58Lastly, it is important to note that these obligations do not apply to open-source models, as discussed below.¹¹⁷ This exception does not apply if the open-source GPAI model presents systemic risks Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. .¹¹⁸

2.1.2. Article 53(1)(b): Transparency for downstream AI system Article 3(1) AI Act: ‘AI system’ means a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge.

2.1.2.1. Context

59Article 53(1)(b) requires GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. to make information and documentation available to system providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. intending to integrate the model into their AI system Article 3(1) AI Act: ‘AI system’ means a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. . To properly understand this requirement, it is useful to briefly contextualise it: any AI system Article 3(1) AI Act: ‘AI system’ means a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. seeking to implement a GPAI model will themselves be subject to the AI Act’s provisions on AI system Article 3(1) AI Act: ‘AI system’ means a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. , including – if the relevant requirements are met – the provisions regarding prohibited practices and high-risk systems. Access to key information about the integrated GPAI model is therefore essential for system providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. to ensure compliance with these obligations.¹¹⁹ The associated penalties (Articles 99–101 AI Act) are sufficiently severe that one could reasonably expect system providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. to refrain from incorporating any GPAI model that would hinder their compliance and thereby their ability to market their system without incurring significant sanctions. Consequently, in the context of prohibited AI practices and high-risk systems, there exists a strong upstream incentive to provide the information necessary to assess whether incorporating a given GPAI model into the intended system configuration might result in such a prohibited practice or a (non-)compliant high-risk system. Specifically for those high-risk systems, the AI Act’s enforcement mechanism generates a robust upstream incentive to incorporate only those (GPAI) models that allow – one could even say facilitate – high-risk system providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. to comply with the AI Act’s requirements. We will discuss some of the relevant high-risk system requirements in more detail below.¹²⁰

60As a result, the explicit requirement to inform downstream AI system Article 3(1) AI Act: ‘AI system’ means a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. to enable compliance (Article 53(1)(b)(i)), and more generally about the GPAI model, may appear somewhat remarkable, particularly with respect to information that would be crucial for such providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. even absent a formal duty of disclosure. Admittedly, some of the information and documentation mandated by Article 53(1)(b) goes beyond what is directly relevant to assessing compliance with prohibited practices and high-risk system requirements, but the observation holds for those sections that do overlap.¹²¹ It also applies to certain information that is essential for the downstream system provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. to implement the model in the first place.¹²² As such, the inclusion of these provisions appears to assume failures¹²³ or frictions in information exchange.¹²⁴ While such an assumption is, to some extent, justified given the complexity, unpredictability and opacity of many AI models and systems¹²⁵ – particularly those deployed in GPAI contexts – it is given a remarkable interpretation here, as is discussed more extensively below.¹²⁶

61An interesting and arguably desirable consequence of this approach is that some of the concerns that typically dominate high-risk AI system Article 3(1) AI Act: ‘AI system’ means a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. compliance will be brought to the attention of low-risk AI system Article 3(1) AI Act: ‘AI system’ means a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. as well if they decide to implement a GPAI model in their system. As such, the mechanism of requiring the GPAI model provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. to share information with the integrating system provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. results in some effects akin to some of the high-risk requirements – increasing awareness among low-risk system providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. – without directly regulating those systems. Article 53 ensures that providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. of such low-risk systems have access to more information and records than they would be required to maintain had they not incorporated a GPAI model.¹²⁷

2.1.2.2. Listed requirements

62Article 53(1)(b) requires GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. to make information and documentation available to AI system Article 3(1) AI Act: ‘AI system’ means a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. intending to integrate the relevant GPAI model in an AI system Article 3(1) AI Act: ‘AI system’ means a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. . This requirement – described as ‘transparency’ in Annex XII – consists of two components. To quote Article 53, the information and documentation ‘shall:

(i) enable providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. of AI systems Article 3(1) AI Act: ‘AI system’ means a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. to have a good understanding of the capabilities and limitations of the general-purpose AI model Article 3(63) AI Act: ‘general-purpose AI model’ means an AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications, except AI models that are used for research, development or prototyping activities before they are placed on the market. and to comply with their obligations pursuant to this Regulation; and

(ii) contain, at a minimum, the elements set out in Annex XII’

63The first of these two components does not specify particular information to be provided or documented, but rather clarifies the objective of the documentation or information requirement. The second element refers to Annex XII and adds that the relevant information and documentation should ‘at a minimum’¹²⁸ contain:

‘1. A general description of the general-purpose AI model Article 3(63) AI Act: ‘general-purpose AI model’ means an AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications, except AI models that are used for research, development or prototyping activities before they are placed on the market. including:
(a) the tasks that the model is intended to perform and the type and nature of AI systems Article 3(1) AI Act: ‘AI system’ means a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. into which it can be integrated;
(b) the acceptable use policies applicable;
(c) the date of release and methods of distribution;
(d) how the model interacts, or can be used to interact, with hardware or software that is not part of the model itself, where applicable;
(e) the versions of relevant software related to the use of the general-purpose AI model Article 3(63) AI Act: ‘general-purpose AI model’ means an AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications, except AI models that are used for research, development or prototyping activities before they are placed on the market. , where applicable;
(f) the architecture and number of parameters;
(g) the modality (e.g. text, image) and format of inputs and outputs;
(h) the licence for the model.

2. A description of the elements of the model and of the process for its development, including:
(a) the technical means (e.g. instructions for use Article 3(15) AI Act: ‘instructions for use’ means the information provided by the provider to inform the deployer of, in particular, an AI system’s intended purpose and proper use. , infrastructure, tools) required for the general-purpose AI model Article 3(63) AI Act: ‘general-purpose AI model’ means an AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications, except AI models that are used for research, development or prototyping activities before they are placed on the market. to be integrated into AI systems Article 3(1) AI Act: ‘AI system’ means a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. ;
(b) the modality (e.g. text, image, etc.) and format of the inputs and outputs and their maximum size (e.g. context window length, etc.);
(c) information on the data used for training, testing and validation, where applicable, including the type and provenance of data and curation methodologies.’

64Many – though not all¹²⁹ – of these requirements largely mirror those found in Annex XI, discussed earlier,¹³⁰ albeit generally with a lower level of detail. The Code of Practice helps to illustrate some of these differences.¹³¹

65In comparison to Annex XI, Annex XII(1) adds two further elements:

‘(d) how the model interacts, or can be used to interact, with hardware or software that is not part of the model itself, where applicable’

66The model provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. should supply information regarding the relevant hardware required to use the model. The Code of Practice adopts a rather minimal interpretation of this provision, requiring the GPAI model provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. only to describe the technical means for model integration (such as ‘ instructions for use Article 3(15) AI Act: ‘instructions for use’ means the information provided by the provider to inform the deployer of, in particular, an AI system’s intended purpose and proper use. ’ or ‘infrastructure tools’), and, where applicable, the hardware and software required (including its version).¹³² As a result, this involves a rather limited level of detail compared to how this requirement is phrased in Annex XII(1). This is, in part, nuanced because certain elements discussed above,¹³³ such as a description of the type and nature of AI systems Article 3(1) AI Act: ‘AI system’ means a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. in which the GPAI model can be integrated, also need to be communicated to the downstream provider Article 3(68) AI Act: ‘downstream provider’ means a provider of an AI system, including a general-purpose AI system, which integrates an AI model, regardless of whether the AI model is provided by themselves and vertically integrated or provided by another entity based on contractual relations. .

67Nevertheless, a more extensive interpretation – more closely aligned with the literal wording of Annex XII – could be preferred by downstream providers Article 3(68) AI Act: ‘downstream provider’ means a provider of an AI system, including a general-purpose AI system, which integrates an AI model, regardless of whether the AI model is provided by themselves and vertically integrated or provided by another entity based on contractual relations. . For example, it is highly useful for them to know whether the model can (or must) connect to external software or hardware, such as being designed to send commands to external tools or applications. It is also important for downstream providers Article 3(68) AI Act: ‘downstream provider’ means a provider of an AI system, including a general-purpose AI system, which integrates an AI model, regardless of whether the AI model is provided by themselves and vertically integrated or provided by another entity based on contractual relations. to know whether the model can or should be integrated with third-party applications, such as email or text processing software, CRM systems or social media platforms. At the same time, they can likely be expected to obtain these details irrespective of the AI Act’s obligations; that is, the frictions characterizing information exchange in the AI context¹³⁴ are unlikely to affect these elements, as they do not directly pertain to the complexity of AI models.

‘(e) the versions of relevant software related to the use of the general-purpose AI model Article 3(63) AI Act: ‘general-purpose AI model’ means an AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications, except AI models that are used for research, development or prototyping activities before they are placed on the market. , where applicable’

68Annex XII(1) also requires GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. to inform downstream system providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. of the relevant software versions that might be related to the use of the GPAI model. The Code of Practice gives this provision a similarly minimal interpretation, requiring the model provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. only to submit information about the software – including the relevant version – necessary to use the model.¹³⁵ Similar considerations as for the model interactions, discussed above, apply.

69Interestingly, Annex XII does not impose explicit requirements regarding the model’s energy consumption, but it does require providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. to disclose the maximum size of the model’s output. Some authors argue that the hardware requirements discussed above should be interpreted even more broadly to encompass the model’s energy performance,¹³⁶ though this approach does not appear to be supported by the wording of the Annex or by the interpretation set out in the Code of Practice.

2.1.2.3. Understanding and compliance requirements

70Article 53(1)(b)(ii) provides that the relevant information must include, ‘at a minimum,’ the elements set out in Annex XII. In addition, Article 53(1)(b)(i), in turn, points to other information that should be communicated, irrespective of Annex XII’s minimal requirements. A first element is that the provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. of the downstream AI system Article 3(1) AI Act: ‘AI system’ means a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. must be able to understand the capabilities and limitations of the GPAI model. This is a sensible requirement that strongly underpins the first paragraph of Annex XII. Where the peculiarities of the GPAI model, or advances in the state-of-the-art in GPAI techniques, necessitate relevant information beyond that listed in Annex XII for proper understanding, the GPAI model provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. should supply it.¹³⁷

71A second element under Article 53(1)(b)(i) requires that the information provided enable downstream system providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. ‘to comply with their obligations pursuant to this Regulation.’ This requirement is notable for several reasons.

72First, it is remarkable that such a broad obligation is not further clarified in Article 53(1)(b)(i) itself, nor is it more explicitly supported, e.g., by the transparency requirements in Annex XII. Nevertheless, this information is particularly significant for providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. of downstream high-risk AI systems Article 3(1) AI Act: ‘AI system’ means a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. , who are subject to extensive obligations under Articles 8–15 of the AI Act.¹³⁸ The more specific requirements made explicit in Annex XII appear insufficient to ensure compliance with some of these obligations.

73This is more apparent for certain requirements than others. For instance, Article 9’s requirement to develop a risk Article 3(2) AI Act: ‘risk’ means the combination of the probability of an occurrence of harm and the severity of that harm. management system does find some support by the obligation to provide ‘ risk Article 3(2) AI Act: ‘risk’ means the combination of the probability of an occurrence of harm and the severity of that harm. profile’-appropriate information under Annex XII, although this raises the question of what the express requirement in Article 53(b)(i) to provide this information truly adds, given that much of this information would already be necessary for high-risk system compliance, as discussed earlier.¹³⁹

74Article 10 may be more demanding, setting out extensive data governance requirements for high-risk system training. Relatedly, Annex XII(2)(c)’s requirement that the GPAI model provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. should supply ‘information on the data used for training, testing and validation, where applicable, including the type and provenance of data and curation methodologies’ should be extended in such contexts, in line with the downstream compliance information requirement established by Article 53(1)(b)(i). Moreover, the likely deployment of a GPAI model in downstream high-risk systems imposes specific training data Article 3(29) AI Act: ‘training data’ means data used for training an AI system through fitting its learnable parameters. requirements¹⁴⁰ that are not made explicit in Chapter V of the AI Act¹⁴¹ but may nonetheless prove crucial.

75Articles 11 and 13 arguably sit at the core of the particular relationship between downstream GPAI model information requirements and the implicit assumption that the market would fail to secure these discussed earlier.¹⁴² These provisions impose extensive technical documentation and transparency obligations on high-risk system providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. . While similar in nature to the information required by Annex XII, compliance with Articles 11 and 13 requires system providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. to obtain far more extensive information from the GPAI model provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. whose model they implement.

76Notably, several of Article 13’s requirements, including certain explainability requirements, extend well beyond what is required under Article 53 and Annex XII. The same is true for some of the more substantive high-risk requirements set out in Articles 14 (enabling human oversight) and 15 (concerning accuracy, robustness and cybersecurity). For such provisions, the AI Act appears to implicitly rely on market functioning to ensure that GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. adapt their models in ways that enable compliance by downstream high-risk system providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. , without explicitly obliging them to do so. At the same time, the Act is far more reluctant to assume that providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. would share the relevant information, even though – by assumption – they would already have adapted their models due to market incentives. Consequently, the Act’s approach to market functioning in this respect appears inconsistent rather than systematic.

77Notably, these requirements for GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. whose model is implemented in high-risk systems may have an interesting spillover effect. GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. could market their models broadly – not limited to either high- or low-risk applications¹⁴³ – and thus may need to ensure that their documentation and information, or even the model itself, meets the standards set for high-risk systems. For reasons of convenience, they may choose to share this information more broadly with prospective downstream system providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. ,¹⁴⁴ even if those providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. intend to use the model in a low-risk system. As a result, GPAI model (self-¹⁴⁵) regulation may indirectly impose certain high-risk system requirements on low-risk systems through the adoption of a shared, high-risk-compatible GPAI model.

2.1.2.4. Limitations & modalities

78These information and documentation requirements are subject to several important limitations. First, there is the open-source exception in Article 53(2), discussed below. Secondly, Article 53(1)(b) limits the information and documentation obligations outlined above by stating that their scope is ‘[w]ithout prejudice to the need to observe and protect intellectual property rights and confidential business information or trade secrets in accordance with Union and national law’.

79The ‘confidential business information’¹⁴⁶ and ‘trade secrets’¹⁴⁷ limitation is particularly pertinent here, as leading GPAI models often rely on proprietary techniques that providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. would be reluctant to see disclosed to competitors or the wider public. Both concepts allow the GPAI model provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. a degree of discretion, provided the information is not generally known.¹⁴⁸ By limiting transparency obligations for confidential material – as the AI Act and its GPAI provisions do elsewhere¹⁴⁹ – the Act seeks to balance the need for information to enable downstream compliance with the commercial interest in retaining certain information. The ‘need to observe and protect’ such information likely encompasses the use of non-disclosure agreements for its sharing. Notably, however, the Act addresses this tension only in relation to the explicit information obligations in Article 53(1)(b), without addressing its wider implications, for example regarding the broader compliance requirements for high-risk system providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. .

80The parallel ‘intellectual property rights’ reservation is likely intended to address European copyright provisions. This primarily concerns copyright,¹⁵⁰ which could protect aspects of the model’s computer code or documentation, as well as database rights¹⁵¹ that may cover databases used for training. This is particularly relevant where third parties hold the relevant rights, potentially limiting the GPAI model provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. ’s ability to share such information. Awarded patents are likely to play a more limited role in this context,¹⁵² as they do not typically entail confidentiality requirements, though the situation may differ in the case of unpublished patent applications. Much like with copyright, Article 53(1)(b) does not require GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. to breach their own confidentiality obligations,¹⁵³ for example when they did not develop the GPAI model alone and/or are otherwise bound by non-disclosure agreements, or jeopardise the confidentiality of their know-how.

81Both the reservation regarding confidential business information/trade secrets and the reservation regarding intellectual property rights arguably restrict the level of detail and types of information that GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. must disclose. This is reinforced by the less detailed requirements in Annex XII compared to Annex XI.¹⁵⁴ These limitations also suggest that GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. may be incentivised to interpret Article 53(1)(b) narrowly to maximise the protection of their trade secrets, confidential business information, and intellectual property interests. There is a direct risk Article 3(2) AI Act: ‘risk’ means the combination of the probability of an occurrence of harm and the severity of that harm. , therefore, that the balance between downstream providers Article 3(68) AI Act: ‘downstream provider’ means a provider of an AI system, including a general-purpose AI system, which integrates an AI model, regardless of whether the AI model is provided by themselves and vertically integrated or provided by another entity based on contractual relations. ’ need for transparency and the confidentiality interests of GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. could be determined solely at the discretion of the latter. In this context, it is worth highlighting two important safeguards for downstream providers Article 3(68) AI Act: ‘downstream provider’ means a provider of an AI system, including a general-purpose AI system, which integrates an AI model, regardless of whether the AI model is provided by themselves and vertically integrated or provided by another entity based on contractual relations. . First, Article 53(1)(b)(ii) introduces a minimum safeguard by requiring communication of certain key elements.¹⁵⁵ Second, GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. must ensure that the information they do share ‘enable[s] providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. of AI systems Article 3(1) AI Act: ‘AI system’ means a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. to have a good understanding of the capabilities and limitations of the general-purpose AI model Article 3(63) AI Act: ‘general-purpose AI model’ means an AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications, except AI models that are used for research, development or prototyping activities before they are placed on the market. and to comply with their obligations pursuant to this Regulation’.¹⁵⁶ Both of these ‘safeguard provisions’ can be enforced effectively,¹⁵⁷ as similar information must be provided in more detail to the AI Office Article 3(47) AI Act: ‘AI Office’ means the Commission’s function of contributing to the implementation, monitoring and supervision of AI systems and general-purpose AI models, and AI governance, provided for in Commission Decision of 24 January 2024; references in this Regulation to the AI Office shall be construed as references to the Commission. and national competent authorities Article 3(48) AI Act: ‘national competent authority’ means a notifying authority or a market surveillance authority; as regards AI systems put into service or used by Union institutions, agencies, offices and bodies, references to national competent authorities or market surveillance authorities in this Regulation shall be construed as references to the European Data Protection Supervisor. upon request (Article 53(1)(a)). In other words, the discretion granted to GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. under Article 53(1)(b)’s confidentiality and IP provisions is, arguably, sufficiently constrained, with adequate incentives to implement their obligations in a manner that safeguards the interests of downstream system providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. .

82As the model developer is required to inform system providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. who ‘intend’ to integrate the model into their systems, an implication is that these providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. should generally have access to the relevant information before incorporating the model.¹⁵⁸ The text is less clear, however, as to whether this means they should receive the information prior to making a decision or before, for example, becoming contractually bound to implement the model. Both the level of detail in Annex XII and the wording of Article 53(1)(b)(i) suggest that the primary aim of the information is to ensure compliance, rather than to inform prospective downstream providers Article 3(68) AI Act: ‘downstream provider’ means a provider of an AI system, including a general-purpose AI system, which integrates an AI model, regardless of whether the AI model is provided by themselves and vertically integrated or provided by another entity based on contractual relations. before they select a model – though it should also ensure that they have a clear understanding of the model’s limitations and capabilities.

83In any case, this provision does not preclude prior non-disclosure agreements.¹⁵⁹ While the Code of Practice states that GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. should provide this information ‘proactively’,¹⁶⁰ the need to preserve confidentiality may in practice require prospective system implementers to contact the provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. before receiving the relevant documentation. Moreover, given the express reference to ‘confidential business information’ and ‘trade secrets’, the GPAI model provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. should be assumed to be able to refuse abusive requests, for example by system providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. that clearly would not adhere to the model’s use policy or by competitors or their proxies without an actual intent to integrate the model.¹⁶¹

84Interestingly, unlike Annex XI, neither Annex XII nor Article 53(1)(b) more generally refer to the risk Article 3(2) AI Act: ‘risk’ means the combination of the probability of an occurrence of harm and the severity of that harm. profile of the model.¹⁶² This is noteworthy, as Recital 101 emphasises the importance of proportionality in this context.¹⁶³ Such a requirement would also be sensible: depending on whether the downstream system qualifies as a high-risk system or one liable to be used in a prohibited AI practice, its provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. is subject to more stringent obligations.¹⁶⁴ While GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. may be inclined to adopt a ‘one size fits all’ approach by supplying the same information to different downstream providers Article 3(68) AI Act: ‘downstream provider’ means a provider of an AI system, including a general-purpose AI system, which integrates an AI model, regardless of whether the AI model is provided by themselves and vertically integrated or provided by another entity based on contractual relations. ,¹⁶⁵ this is arguably addressed by the requirement that the information provided must enable downstream providers Article 3(68) AI Act: ‘downstream provider’ means a provider of an AI system, including a general-purpose AI system, which integrates an AI model, regardless of whether the AI model is provided by themselves and vertically integrated or provided by another entity based on contractual relations. ‘to comply with their obligations pursuant to this Regulation’, which implies that less information should be required for low-risk systems. Some authors nevertheless argue that Recital 101’s principle of proportionality should be read more fully into Article 53(1)(b) and Annex XII,¹⁶⁶ limiting the extent of information to be provided according to the size and risk Article 3(2) AI Act: ‘risk’ means the combination of the probability of an occurrence of harm and the severity of that harm. profile of the GPAI model. At the same time, Recital 109 introduces a countervailing consideration,¹⁶⁷ imposing more stringent requirements on larger model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. , which would appear to restrict this interpretation to relatively small GPAI providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. offering smaller models with more nuanced risk Article 3(2) AI Act: ‘risk’ means the combination of the probability of an occurrence of harm and the severity of that harm. profiles. In any event, the impact of a proportional approach is limited, given that Article 53(1)(b) and Annex XII are already less demanding than Article 53(1)(a) and Annex XI – meaning that any obligation to share more information with downstream providers Article 3(68) AI Act: ‘downstream provider’ means a provider of an AI system, including a general-purpose AI system, which integrates an AI model, regardless of whether the AI model is provided by themselves and vertically integrated or provided by another entity based on contractual relations. remains comparatively lighter than under the latter provisions.

2.1.3. Article 53(1)(c): Copyright compliance policy

2.1.3.1. Positioning and scope

85Article 53(1)(c) requires GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. to establish a policy to ensure compliance with Union law and copyright and related rights regulation.¹⁶⁸ In particular, this policy should ensure – including through state-of-the-art technologies – that the model provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. processes the text and data mining (TDM) opt-out specified in Article 4(3) of Directive (EU) 2019/790. TDM refers to the automated analysis – including for AI model training¹⁶⁹ – of text and data to identify patterns and extract useful information, often using web crawlers.¹⁷⁰ Rightsholders must permit TDM by research organisations and cultural heritage institutions for scientific research purposes.¹⁷¹ For commercial purposes, TDM is permitted by default unless the rightsholder opts out.¹⁷² If an opt-out is exercised, it must be in machine-readable form.¹⁷³ In the absence of an opt-out, TDM is permissible.¹⁷⁴ This obligation is reiterated by various measures in the Code of Practice.¹⁷⁵

86Where there is no opt-out, the party wishing to conduct TDM must have lawful access to the data and text,¹⁷⁶ meaning, for example, that they must not circumvent subscription models or paywalls.¹⁷⁷ Additionally, the Code of Practice states that GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. should ‘exclude from their web-crawling websites that make available to the public content and which are, at the time of web-crawling, recognised as persistently and repeatedly infringing copyright and related rights on a commercial scale by courts or public authorities in the European Union and the European Economic Area’.¹⁷⁸ The Code of Practice further emphasises the efforts – including deploying state-of-the-art technologies¹⁷⁹ – that GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. should undertake to identify opt-outs.¹⁸⁰ It also provides that GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. who also operate a search engine should not interpret a TDM opt-out as also constituting an opt-out from search indexing.¹⁸¹ If a TDM opt-out were interpreted that broadly, exercising it for copyright reasons could prove punitive for the rightsholder, as it would cut off an important stream of traffic to their content.

87The inclusion of copyright in the AI Act has attracted some criticism,¹⁸² as it conflates the private rights nature of copyright protection with the public interest objectives of the AI Act.¹⁸³ It is also notable that the Act imposes a specific obligation on GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. to comply with copyright law, without imposing a similar requirement – despite a generally higher level of depth and detail in the relevant provisions – on high-risk system providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. or, indeed, on all AI model and system providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. .¹⁸⁴

88Interestingly, a dominant reading gives this obligation a strikingly broad territorial scope.¹⁸⁵ Recital 106 states that GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. should comply with copyright law ‘regardless of the jurisdiction in which the copyright-relevant acts underpinning the training of those general-purpose AI models Article 3(63) AI Act: ‘general-purpose AI model’ means an AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications, except AI models that are used for research, development or prototyping activities before they are placed on the market. take place.’¹⁸⁶ This extension – sometimes characterised as a ‘maximalist’ interpretation¹⁸⁷ – is not confirmed by the text of the regulation itself. Notably, it would also extend the traditional territorial scope of EU copyright law.¹⁸⁸ A more ‘minimalist’ reading would suggest that the territorial reach of EU copyright law should set an upper limit to the AI Act’s effect in this area:¹⁸⁹ as EU copyright law and the TDM opt-out do not apply to a foreign developer training their model outside the EU, such developers could argue that they comply with EU copyright law irrespective of their training practices (i.e. even if their actions would violate EU copyright law if conducted within the EU).¹⁹⁰ This would mean that the EU TDM exception could become highly relevant in foreign jurisdictions, even where it is not part of the local copyright regime.¹⁹¹ A less convincing intermediate approach would require compliance with the TDM opt-out only if the data is hosted on a European server.¹⁹²

89The maximalist interpretation is most consistent with the objectives of the AI Act, as set out in Recital 106. This ‘product regulation’ approach¹⁹³ is further reflected in other provisions of the Act, such as the requirement that training data Article 3(29) AI Act: ‘training data’ means data used for training an AI system through fitting its learnable parameters. for high-risk systems meet specified anti-bias standards.¹⁹⁴ Resultingly, even if parties develop their model or system outside the Union, their conduct during development becomes subject to the AI Act – and, in this case, to EU copyright law – if they subsequently put their system into service within the EU.¹⁹⁵ This interpretation best reflects the intention to create a level playing field, ensuring that developers outside the EU cannot exploit more lenient copyright regimes abroad to gain an advantage with their GPAI models within the EU.¹⁹⁶

90Some authors argue that this obligation does not mean that all AI development is covered by this provision, for example when the provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. of the GPAI model (within the meaning of Article 3(3)) did not themselves develop the model but merely places on the EU market a model that was developed outside the EU by a party who does not qualify as a provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. under Article 3(3) AI Act.¹⁹⁷ In that case, the language of Article 53(1)(c), which targets the ‘ provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. ’ of the model, does not appear to apply to the model training.¹⁹⁸ This, however, fundamentally depends on how Article 53(1)(c) AI Act is to be understood. If it is regarded as a form of product regulation,¹⁹⁹ imposing requirements on how the model was developed, it could be taken to bind the provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. who places the model on the market. If, instead, it is viewed as a form of entity regulation,²⁰⁰ its scope would be confined to the actions of the provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. of the model. While the nature of the obligation and the language of Article 53(1)(c) suggest an interpretation as entity regulation, Recital 106 could also be read as supporting a characterisation as product regulation, which the provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. must ensure compliance with regardless of who trained the model beforehand.²⁰¹

91The obligation to comply with EU copyright law and TDM opt-outs, at least under the AI Act, is sometimes argued to be one of best efforts rather than strict liability for any (minor) breach²⁰² – a view supported by Recital 108²⁰³ – because it is currently technologically impossible to perfectly filter out copyright-protected material.²⁰⁴ This grants the AI Office Article 3(47) AI Act: ‘AI Office’ means the Commission’s function of contributing to the implementation, monitoring and supervision of AI systems and general-purpose AI models, and AI governance, provided for in Commission Decision of 24 January 2024; references in this Regulation to the AI Office shall be construed as references to the Commission. significant authority in the copyright sphere, as it may set the relevant standard,²⁰⁵ though Recital 108 makes clear that this standard should not require perfection. While the establishment of a policy is strictly required, this does thus not equate to an obligation of absolute compliance with EU copyright law. In this context, it is interesting that some German commentators have suggested that Article 53(1)(c) might constitute a so-called Schutznorm²⁰⁶ – a notion in German law that refers to a provision intended to protect specific parties or interests from harm²⁰⁷ – which would potentially allow affected individuals to derive rights from it, such as to bring liability claims. However, it is important to emphasise that a breach of copyright itself does not necessarily entail a breach of this provision – for instance, a GPAI developer could have a robust policy but still be unable, due to external factors or technical limitations, to prevent every possible copyright violation.²⁰⁸ This does not, of course, preclude rightsholders from bringing claims under copyright law itself.

92As a result, we would question the assertion some authors make that violations of this provision are obvious,²⁰⁹ as the text does not make clear how comprehensive the required policy must be or which specific technical measures GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. are expected to implement to achieve compliance. Even if this provision were to confer an individual right of action under certain national legal systems, claimants would likely find it challenging to demonstrate that a specific copyright violation resulted from the absence of an adequate policy, particularly given the inherent difficulty in eliminating all copyright infringements, even with robust policies, and the causality issues that typically characterize cases of AI harm explored elsewhere in this work.²¹⁰

93Lastly, it is worth emphasising that the policy requirement in Article 53(1)(c) applies not only to the model’s training data Article 3(29) AI Act: ‘training data’ means data used for training an AI system through fitting its learnable parameters. (input) but also to its output. While some maintain that GPAI providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. are not obliged to verify the output of another provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. ’s system incorporating the GPAI model²¹¹ – nor can they be held liable for infringements arising at the level of the implementing system²¹² – it is nevertheless clear that the (potential) output of the GPAI model should fall within the scope of the copyright policy.²¹³ The Code of Practice further states that GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. should implement safeguards to help prevent their models from infringing copyright and that they should require downstream system providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. to accept terms and conditions designed to prevent copyright violations.²¹⁴

2.1.3.2. Policy requirements

94In general, a copyright policy requires that the GPAI model provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. first assess the extent to which the training and use of the model may give rise to copyright infringements.²¹⁵ In other words, the provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. must determine for which aspects of training and use the rightsholder’s permission is necessary.²¹⁶ The Code of Practice addresses this most explicitly in the context of TDM opt-out detection, as discussed above, but the compliance policy requirement extends further, also covering other potential copyright infringements when collecting or processing copyrighted material.

95At the next stage, the GPAI model provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. should devise a plan to mitigate the risks Article 3(2) AI Act: ‘risk’ means the combination of the probability of an occurrence of harm and the severity of that harm. of copyright infringement.²¹⁷ This could involve deploying state-of-the-art techniques – which Article 53(1)(c) refers to specifically in relation to the TDM opt-out – to prevent copyright violations more generally. For model outputs, the Code of Practice similarly requires GPAI developers to implement ‘appropriate and proportionate’ technical safeguards to prevent the model from reproducing copyright-protected material used during training, and to prohibit copyright-infringing uses through contractual terms with downstream providers Article 3(68) AI Act: ‘downstream provider’ means a provider of an AI system, including a general-purpose AI system, which integrates an AI model, regardless of whether the AI model is provided by themselves and vertically integrated or provided by another entity based on contractual relations. , or, in the case of open-source models, at least to alert users to the prohibition of copyright-infringing use.²¹⁸ On the training side, GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. could impose similar obligations on parties assisting in the acquisition of training material.²¹⁹ It is also important to stress that the policy must not be merely theoretical but must be effectively implemented.²²⁰

96According to the Code of Practice, GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. should, as part of their policy, designate a point of contact and enable stakeholders to submit complaints.²²¹ Such complaints should be handled diligently and in accordance with due process, without prejudicing any potential copyright-based claims by rightsholders.²²² Several other provisions in the Code of Practice also emphasise the need to make relevant information available to rightsholders who believe their copyright may have been infringed by the model.²²³

2.1.4. Article 53(1)(d): Summary of training content

2.1.4.1. Requirement and rationale

97According to Article 53(1)(d), the GPAI model provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. is also required to prepare a summary of the data used to train the model. To this end, the AI Act mandates the AI Office Article 3(47) AI Act: ‘AI Office’ means the Commission’s function of contributing to the implementation, monitoring and supervision of AI systems and general-purpose AI models, and AI governance, provided for in Commission Decision of 24 January 2024; references in this Regulation to the AI Office shall be construed as references to the Commission. to provide a template for providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. to complete,²²⁴ which it has.²²⁵ While Recital 107 emphasises the need to respect trade secrets and confidential business information, both Article 53(1)(d) and Recital 107 indicate that the summary should be reasonably detailed, though not necessarily highly technical. Recital 107 further clarifies that the purpose of the summary is to ‘facilitate parties with legitimate interests, including copyright holders, to exercise and enforce their rights under Union law, for example by listing the main data collections or sets that went into training the model, such as large private or public databases or data archives, and by providing a narrative explanation about other data sources used.’²²⁶ The level of detail – while not necessarily technical²²⁷ – should enable those parties to assess their legal position and identify any potential concerns regarding the data used. The non-binding Explanatory Notice to the Template clarifies that this legal position extends beyond copyright concerns²²⁸ to include all rights protected under Union law,²²⁹ such as data protection²³⁰ and the freedom to receive information and conduct scientific research.²³¹

98Additionally, this information can be valuable for downstream system providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. considering integration of the model, allowing them to assess the diversity of the data used.²³² It is also argued that such transparency may lead to more competitive markets, as it enables downstream actors to better evaluate how their data and models have been used, thus reducing lock-in effects.²³³ At the same time, this transparency requirement has been criticised for imposing a significant burden on providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. .²³⁴

99The wording of Article 53(1)(d) strongly suggests that the information must be shared using the template provided by the AI Office Article 3(47) AI Act: ‘AI Office’ means the Commission’s function of contributing to the implementation, monitoring and supervision of AI systems and general-purpose AI models, and AI governance, provided for in Commission Decision of 24 January 2024; references in this Regulation to the AI Office shall be construed as references to the Commission. . Recital 107 takes a softer approach, stating that ‘[it] is appropriate for the AI Office Article 3(47) AI Act: ‘AI Office’ means the Commission’s function of contributing to the implementation, monitoring and supervision of AI systems and general-purpose AI models, and AI governance, provided for in Commission Decision of 24 January 2024; references in this Regulation to the AI Office shall be construed as references to the Commission. to provide a template for the summary, which should be simple, effective, and allow the provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. to provide the required summary in narrative form’ without expressly indicating that this form should be used by providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. . Nevertheless, the more literal reading of Article 53(1)(d) – that the AI Office Article 3(47) AI Act: ‘AI Office’ means the Commission’s function of contributing to the implementation, monitoring and supervision of AI systems and general-purpose AI models, and AI governance, provided for in Commission Decision of 24 January 2024; references in this Regulation to the AI Office shall be construed as references to the Commission. template must be used – is reaffirmed by the Explanatory Notice to the Template²³⁵ and, arguably, aligns more closely with the core objective of this obligation: to enable parties with legitimate interests to assess their legal position, with uniformity supporting both this goal and the document’s accessibility. Even if Article 53(1)(d) were interpreted as not mandating use of the template, employing the template would remain the most straightforward way to comply with the duty to disclose the summary of training content, compared to communicating that information by other means. However, if a GPAI model provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. can communicate the same information as effectively through another format, one might question whether anyone would be disadvantaged and whether the aims of the provision are not still met.²³⁶

100It is also important to note that the GPAI model provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. must make this document ‘publicly available’. This availability should go beyond mere theoretical access.²³⁷ Given the wide range of potentially interested parties, it is reasonable to require that the document be accessible online in a digital format. This is also what the Explanatory Notice suggests: ‘[the Summary] should be published on the provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. ’s official website in a clearly visible and accessible manner, clearly indicating which model(s) (and possibly model version(s)) the Summary covers […]. The Summary should also be made publicly available together with the model across all its public distribution channels (e.g. online platforms).’²³⁸ While the wording of Article 53(1)(d) does not entirely preclude a provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. from meeting its obligations by making the document ‘publicly available’ on-site, such in-person availability would run counter to the transparency that the provision and Recital 107 seek to establish. In any case, the document must be made available no later than the date on which the model is placed on the Union market.²³⁹

2.1.4.2. Detail and modalities

101The Template that was published by the European Commission provides more insight into the information that should be shared. It clarifies that the information should be provided with sufficient detail,²⁴⁰ covering all types of data used throughout the model’s lifecycle – from pre-training to post-training – including model alignment and fine-tuning.²⁴¹ Data used during model operation, such as in retrieval-augmented generation, falls outside the scope unless it contributes to model training.²⁴² The information should be comprehensive,²⁴³ presented in a narrative, simple, and effective format to ensure it is understandable to the relevant parties.²⁴⁴ It must be accurate, comprehensive, and provided in good faith.²⁴⁵ The AI Office Article 3(47) AI Act: ‘AI Office’ means the Commission’s function of contributing to the implementation, monitoring and supervision of AI systems and general-purpose AI models, and AI governance, provided for in Commission Decision of 24 January 2024; references in this Regulation to the AI Office shall be construed as references to the Commission. will assess compliance and, if necessary, request corrective measures or seek enforcement.²⁴⁶

102At the same time, it is important to emphasise that the template aims to balance transparency requirements with the need to protect confidential business information and trade secrets.²⁴⁷ This has resulted in significant limitations on the data that must be disclosed. The disclosure obligation for licensed data is limited, which makes sense as the relevant rightsholders are already parties to those licence agreements.²⁴⁸ Private datasets that are not commercially licensed only need to be listed if they are publicly known or if the provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. chooses to make them public.²⁴⁹ Commercially sensitive details concerning data sources, model curation, or training methods do not need to be disclosed.²⁵⁰ Only minimal information is required regarding user data from interactions, explicitly excluding data licensed via commercial agreements or customer fine-tuning data.²⁵¹ For synthetic data, disclosure is limited to the names of the models used if they are placed on the market Article 3(9) AI Act: ‘placing on the market’ means the first making available of an AI system or a general-purpose AI model on the Union market. , or a general description of model training data Article 3(29) AI Act: ‘training data’ means data used for training an AI system through fitting its learnable parameters. where necessary.²⁵²

103Furthermore, only a high-level aggregated overview of training data Article 3(29) AI Act: ‘training data’ means data used for training an AI system through fitting its learnable parameters. size per modality, presented in broad ranges, is required. For publicly available datasets, more detail must be provided.²⁵³ For scraped data, only a summarised narrative list of the most relevant domain names must be provided – not a full list of URLs.²⁵⁴ In accordance with Recital 107, the Explanatory Notice also stresses that disclosures should be non-technical and presented in a summarised narrative form, so as to avoid revealing sensitive information.²⁵⁵ Only high-level aggregates regarding the mix and composition of data sources must be disclosed, without specifying the exact mix.²⁵⁶ With respect to crawlers, providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. must disclose their purpose and collection periods, but not their precise technical implementation.²⁵⁷

104In any case, providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. may go beyond these minimum requirements and disclose more information than is required by the template.²⁵⁸ Providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. are also encouraged to respond to requests from relevant stakeholders with legitimate interests who wish to better assess their own legal position regarding the data used, provided this does not breach the provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. ’s own obligations. This is particularly encouraged for information scraped or crawled from the internet.²⁵⁹

2.1.4.3. Template content

105Without seeking to reiterate the full content of the template itself,²⁶⁰ it consists of three main sections.²⁶¹ First, it collects general information to identify the provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. and the model, as well as the characteristics of the training data Article 3(29) AI Act: ‘training data’ means data used for training an AI system through fitting its learnable parameters. , broken down by modality (text, image, audio, video, etc.), estimated data size, and general content types. Second, it requires a detailed list and categorisation of all data sources used for training, including public datasets, private or licensed datasets, data scraped from online sources, user data, synthetic data, and any other sources, together with descriptions and, where relevant, lists of domain names. Third, it addresses data processing aspects, including measures to respect copyright opt-outs, the removal of illegal content, and other relevant steps taken to ensure compliance with Union law.²⁶²

2.1.4.4. Adjusted content for pre-existing GPAI models/updates

106The requirements outlined above are affected if a downstream entity sufficiently²⁶³ modifies a model already placed on the Union market so that they themselves become the provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. of the modified model.²⁶⁴ In such cases, the modifying entity should report only the training data Article 3(29) AI Act: ‘training data’ means data used for training an AI system through fitting its learnable parameters. used for the modification in the template, as well as the name of the original model.²⁶⁵ If a provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. continues to train their own model, they should also update the information provided at six-month intervals, or sooner if the training results in a materially significant update to the content of the summary.²⁶⁶ The updated summary should then be made available alongside the modified model.²⁶⁷

107The Explanatory Notice does not specify the criteria for determining whether changes to the training data Article 3(29) AI Act: ‘training data’ means data used for training an AI system through fitting its learnable parameters. constitute a ‘significant update’. However, minor changes – such as correcting labelling errors, adding relatively small amounts of data, or retraining on the same domains without significant alteration – are unlikely to qualify. By contrast, a significant update is more likely where entirely new data sources or domains are added (for example, the inclusion of medical or financial data), where there is a substantial expansion of the dataset, or where the dataset comes to include non-copyrighted material in addition to copyrighted material, or vice versa. The same applies in cases of substantial reweighting of the various data types used.

108In any case, this assessment involves a degree of judgement on the part of the provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. . Crucially, the purpose of the information being shared must be taken into account. Accordingly, an update is required sooner than the six-month interval if the changes made could have potentially important consequences for stakeholders such as rights managers and copyright holders. Bearing this purpose in mind, it is unlikely that changes to the training data Article 3(29) AI Act: ‘training data’ means data used for training an AI system through fitting its learnable parameters. which do not meet the described thresholds but which result in significantly altered model behaviour or performance are relevant here. While such updates may arguably be considered ‘significant’, their significance does not necessarily extend to stakeholders concerned with the data used.

109A single summary may be used for multiple versions of the same model if their summaries are identical, provided that those models and versions are clearly identified. If one of these versions has already been placed on the Union market, requiring an earlier summary, the summaries for subsequent versions need only cover the training data Article 3(29) AI Act: ‘training data’ means data used for training an AI system through fitting its learnable parameters. specifically used to modify the original version, along with a clear reference to the original summary.²⁶⁸

2.2. Article 53(2): Open-source exception

110Article 53(2) provides a partial exception from the preceding obligations for open-source GPAI models.²⁶⁹ Where a GPAI model provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. meets certain requirements, they are not obliged to provide documentation to the AI Office Article 3(47) AI Act: ‘AI Office’ means the Commission’s function of contributing to the implementation, monitoring and supervision of AI systems and general-purpose AI models, and AI governance, provided for in Commission Decision of 24 January 2024; references in this Regulation to the AI Office shall be construed as references to the Commission. and national competent authorities Article 3(48) AI Act: ‘national competent authority’ means a notifying authority or a market surveillance authority; as regards AI systems put into service or used by Union institutions, agencies, offices and bodies, references to national competent authorities or market surveillance authorities in this Regulation shall be construed as references to the European Data Protection Supervisor. nor to supply information to downstream AI system Article 3(1) AI Act: ‘AI system’ means a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. , provided those models do not present a systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. . If the model does present systemic risks Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. , however, these information requirements do apply²⁷⁰ – and are, in part, extended by Section 2. of Annex XI. This open-source exception is motivated by the need to foster innovation and the growth opportunities that such models could offer for the European Union.²⁷¹

111To qualify for the exception, the model must be released under a free and open-source licence. The Commission Guidelines clarify that the term ‘licence’ should be interpreted broadly, referring to the granting of permissions such that the original provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. does not use their intellectual property rights to restrict the use of the model or charge for its use.²⁷² Furthermore, they indicate that the licence should provide for access, use, modification and distribution; modifying, using, or distributing the model should therefore be possible without restriction, although limited conditions are permissible.²⁷³ These ‘limited conditions often consist only in crediting the author(s) and retaining their copyright notice, i.e. attribution.’²⁷⁴

112The Commission Guidelines also identify certain restrictions that disqualify a licence from meeting these criteria.²⁷⁵ These include limitations to non-commercial or research use only, prohibitions on distributing the model or its components, usage restrictions relating to user scale thresholds (which require additional licensing), and requirements to obtain a specific licence for certain use cases.²⁷⁶

113Moreover, a model is not considered to be ‘free’ if it is monetised indirectly, including through additional services (for example, for necessary technical support or security),²⁷⁷ or if access to the model requires the purchase of support or training.²⁷⁸ This exclusion also covers instances where the model provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. collects data – other than for the purpose of improving security, compatibility, or interoperability – for monetisation purposes.²⁷⁹ Furthermore, it precludes models offered under licences that permit free academic use but require payment for commercial or scaled use, or licences under which use of the model necessarily requires purchasing access to a platform or server hosted by the provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. .²⁸⁰ Permissible practices, however, include situations where the model is offered alongside paid services that are purely optional or where such paid services or support are made available in the form of premium versions or extensions of the model.²⁸¹

114In addition, there is an information or transparency requirement. The provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. should make publicly available the model’s parameters (including weights) and the model architecture, as well as relevant information on model usage. This should arguably be interpreted broadly, given the purpose of the exception²⁸² – as wider access would better enable the model’s open-access status to foster innovation, facilitate further development, and allow downstream providers Article 3(68) AI Act: ‘downstream provider’ means a provider of an AI system, including a general-purpose AI system, which integrates an AI model, regardless of whether the AI model is provided by themselves and vertically integrated or provided by another entity based on contractual relations. to integrate the model. To this end, the Commission Guidelines similarly stress that this information should at least include ‘[i]nformation about the model’s input and output modalities, capabilities, and limitations [including] the technical means (e.g. instructions for use Article 3(15) AI Act: ‘instructions for use’ means the information provided by the provider to inform the deployer of, in particular, an AI system’s intended purpose and proper use. , infrastructure, tools) required for the model to be integrated into AI systems Article 3(1) AI Act: ‘AI system’ means a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. , which may include the appropriate configuration for the intended use cases, where applicable.’²⁸³

2.3. Article 53(3): Duty of cooperation

115Article 53(3) imposes a duty on GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. to cooperate with the Commission and national competent authorities Article 3(48) AI Act: ‘national competent authority’ means a notifying authority or a market surveillance authority; as regards AI systems put into service or used by Union institutions, agencies, offices and bodies, references to national competent authorities or market surveillance authorities in this Regulation shall be construed as references to the European Data Protection Supervisor. in the exercise of their AI Act competencies and powers. This obligation should also be understood as extending to the AI Office Article 3(47) AI Act: ‘AI Office’ means the Commission’s function of contributing to the implementation, monitoring and supervision of AI systems and general-purpose AI models, and AI governance, provided for in Commission Decision of 24 January 2024; references in this Regulation to the AI Office shall be construed as references to the Commission. .²⁸⁴ In some respects, Article 53(3) constitutes the GPAI provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. equivalent of Articles 21, 23(7), 24(6) and 26(12), which impose cooperation duties on various other actors – namely, high-risk system providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. , importers Article 3(6) AI Act: ‘importer’ means a natural or legal person located or established in the Union that places on the market an AI system that bears the name or trademark of a natural or legal person established in a third country. , distributors Article 3(7) AI Act: ‘distributor’ means a natural or legal person in the supply chain, other than the provider or the importer, that makes an AI system available on the Union market. , and deployers Article 3(4) AI Act: ‘deployer’ means a natural or legal person, public authority, agency or other body using an AI system under its authority except where the AI system is used in the course of a personal non-professional activity. , respectively.

116Article 53(3) states that the duty to cooperate applies ‘in the exercise of their competences and powers pursuant to this Regulation’. The material scope of this obligation thus extends beyond the other provisions of Article 53 and covers the entirety of the AI Act, with particular reference to Articles 88 to 94.²⁸⁵ It should also be read in conjunction with Article 101(1)(b), which imposes a fine on GPAI providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. who intentionally or negligently ‘[fail] to comply with a request for a document or for information pursuant to Article 91, or supplied incorrect, incomplete or misleading information’. The relationship between some of these provisions and Article 53 is, however, not always clear. For example, the Commission²⁸⁶ can request the documentation required by Article 53(1)(a) both under that provision itself and similarly under Article 91, whereas national competent authorities Article 3(48) AI Act: ‘national competent authority’ means a notifying authority or a market surveillance authority; as regards AI systems put into service or used by Union institutions, agencies, offices and bodies, references to national competent authorities or market surveillance authorities in this Regulation shall be construed as references to the European Data Protection Supervisor. can seemingly only do so on the basis of Article 53(1)(a), raising the question as to the additional function of Article 91. Its role appears clearer for other parts of Article 53(1), such as (1)(b) and (c), which do not grant a separate competence to request information. A different interpretation, suggested by the Article’s title (‘Obligations for providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. of general-purpose AI models Article 3(63) AI Act: ‘general-purpose AI model’ means an AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications, except AI models that are used for research, development or prototyping activities before they are placed on the market. ’) could see Article 53(1)(a) as only binding on the providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. without providing a power for the AI Office Article 3(47) AI Act: ‘AI Office’ means the Commission’s function of contributing to the implementation, monitoring and supervision of AI systems and general-purpose AI models, and AI governance, provided for in Commission Decision of 24 January 2024; references in this Regulation to the AI Office shall be construed as references to the Commission. the Commission or national competent authorities Article 3(48) AI Act: ‘national competent authority’ means a notifying authority or a market surveillance authority; as regards AI systems put into service or used by Union institutions, agencies, offices and bodies, references to national competent authorities or market surveillance authorities in this Regulation shall be construed as references to the European Data Protection Supervisor. to request the information, though that would leave the AI Office Article 3(47) AI Act: ‘AI Office’ means the Commission’s function of contributing to the implementation, monitoring and supervision of AI systems and general-purpose AI models, and AI governance, provided for in Commission Decision of 24 January 2024; references in this Regulation to the AI Office shall be construed as references to the Commission. and national competent authorities Article 3(48) AI Act: ‘national competent authority’ means a notifying authority or a market surveillance authority; as regards AI systems put into service or used by Union institutions, agencies, offices and bodies, references to national competent authorities or market surveillance authorities in this Regulation shall be construed as references to the European Data Protection Supervisor. without such a power under the AI Act.²⁸⁷

117The Article 53(3) duty to cooperate is not further clarified in the recitals, leaving its interpretation largely open. This is particularly relevant to the extent of the obligation. Given the phrase ‘as necessary’, one could interpret this obligation as requiring GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. only to respond to requests for information and documentation made by the AI Office Article 3(47) AI Act: ‘AI Office’ means the Commission’s function of contributing to the implementation, monitoring and supervision of AI systems and general-purpose AI models, and AI governance, provided for in Commission Decision of 24 January 2024; references in this Regulation to the AI Office shall be construed as references to the Commission. , the Commission, or national competent authorities Article 3(48) AI Act: ‘national competent authority’ means a notifying authority or a market surveillance authority; as regards AI systems put into service or used by Union institutions, agencies, offices and bodies, references to national competent authorities or market surveillance authorities in this Regulation shall be construed as references to the European Data Protection Supervisor. , provided such requests are made within the exercise of their AI Act powers and competences.

118While somewhat less convincing,²⁸⁸ one could also interpret this obligation more broadly.²⁸⁹ A broader reading might require GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. to proactively submit information that has not been expressly requested by the AI Office Article 3(47) AI Act: ‘AI Office’ means the Commission’s function of contributing to the implementation, monitoring and supervision of AI systems and general-purpose AI models, and AI governance, provided for in Commission Decision of 24 January 2024; references in this Regulation to the AI Office shall be construed as references to the Commission. , the Commission, or national competent authorities Article 3(48) AI Act: ‘national competent authority’ means a notifying authority or a market surveillance authority; as regards AI systems put into service or used by Union institutions, agencies, offices and bodies, references to national competent authorities or market surveillance authorities in this Regulation shall be construed as references to the European Data Protection Supervisor. , insofar as doing so could facilitate their enforcement of the AI Act. Such an interpretation appears less consistent with the literal wording of Article 53(3) and is also implicitly contradicted by the explicitly phrased duty to notify – a form of proactive cooperation – the AI Office Article 3(47) AI Act: ‘AI Office’ means the Commission’s function of contributing to the implementation, monitoring and supervision of AI systems and general-purpose AI models, and AI governance, provided for in Commission Decision of 24 January 2024; references in this Regulation to the AI Office shall be construed as references to the Commission. and, where relevant, national competent authorities Article 3(48) AI Act: ‘national competent authority’ means a notifying authority or a market surveillance authority; as regards AI systems put into service or used by Union institutions, agencies, offices and bodies, references to national competent authorities or market surveillance authorities in this Regulation shall be construed as references to the European Data Protection Supervisor. of serious incidents Article 3(49) AI Act: ‘serious incident’ means an incident or malfunctioning of an AI system that directly or indirectly leads to any of the following: (a) the death of a person, or serious harm to a person’s health; (b) a serious and irreversible disruption of the management or operation of critical infrastructure; (c) the infringement of obligations under Union law intended to protect fundamental rights; (d) serious harm to property or the environment. involving GPAI models with systemic risk Article 3(65) AI Act: ‘systemic risk’ means a risk that is specific to the high-impact capabilities of general-purpose AI models, having a significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or the society as a whole, that can be propagated at scale across the value chain. , as set out in Article 55(1)(c).²⁹⁰

119Regardless of these potential diverging interpretations, Article 53, taken together with the wording of Article 101(1)(b) – which requires that supplied information be correct, complete and not misleading – means that this obligation should be interpreted as requiring GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. to supply detailed, accurate and reliable information. These elements should be assessed in light of the enforcement objectives for which they are provided.

2.4. Article 53(4): Compliance pathways

120Article 53(4)²⁹¹ sets out how GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. can demonstrate compliance with the requirements outlined in the previous sections. It offers three distinct pathways to this end. First, GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. may demonstrate compliance by adhering to harmonised standards Article 3(27) AI Act: ‘harmonised standard’ means a harmonised standard as defined in Article 2(1), point (c), of Regulation (EU) No 1025/2012. , which creates a presumption of conformity with the AI Act insofar as the relevant obligations are addressed by those standards.²⁹² In the absence of published harmonised standards Article 3(27) AI Act: ‘harmonised standard’ means a harmonised standard as defined in Article 2(1), point (c), of Regulation (EU) No 1025/2012. , model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. may rely on codes of practice.²⁹³ Finally, GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. may demonstrate compliance by any other ‘adequate means’.

2.4.1. Harmonised standards Article 3(27) AI Act: ‘harmonised standard’ means a harmonised standard as defined in Article 2(1), point (c), of Regulation (EU) No 1025/2012.

121Over time, compliance with harmonised standards Article 3(27) AI Act: ‘harmonised standard’ means a harmonised standard as defined in Article 2(1), point (c), of Regulation (EU) No 1025/2012. can be expected to become the principal mechanism for fulfilling GPAI model provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. obligations. This notion refers to the definition found in Article 2(1)(c) of Regulation (EU) No 1025/2012,²⁹⁴ which describes a harmonised standard Article 3(27) AI Act: ‘harmonised standard’ means a harmonised standard as defined in Article 2(1), point (c), of Regulation (EU) No 1025/2012. as ‘a European standard adopted on the basis of a request made by the Commission for the application of Union harmonisation legislation’. These standards, developed by the European Committee for Standardisation (CEN), the European Committee for Electrotechnical Standardisation (Cenelec) or, the European Telecommunications Standards Institute (ETSI),²⁹⁵ can be expected to reflect the state-of-the-art²⁹⁶ and will be formulated with a ‘balanced representation of interests involving all relevant stakeholders’,²⁹⁷ following a request by the Commission.²⁹⁸ Article 10(6) Regulation (EU) No 1025/2012 indicates that a reference of the standard will be published in the Official Journal of the European Union.

122While Recital 117 indicates that the AI Office Article 3(47) AI Act: ‘AI Office’ means the Commission’s function of contributing to the implementation, monitoring and supervision of AI systems and general-purpose AI models, and AI governance, provided for in Commission Decision of 24 January 2024; references in this Regulation to the AI Office shall be construed as references to the Commission. will assess whether such a standard constitutes a ‘suitable’ instrument to cover the relevant obligations,²⁹⁹ this idea is not directly reflected in the enacting terms. Instead, this likely refers to the Article 10(5) Regulation (EU) No 1025/2012 provision that the Commission ‘shall assess the compliance of the documents drafted by the European standardisation organisations with its initial request’.

123Once it is established that an obligation is covered by a relevant harmonised standard Article 3(27) AI Act: ‘harmonised standard’ means a harmonised standard as defined in Article 2(1), point (c), of Regulation (EU) No 1025/2012. , and a reference of that standard has been published in the Official Journal of the European Union, Article 53(4) introduces a presumption that compliance with the standard entails compliance with Article 53 vis-à-vis that obligation. It is key to note, however, that that presumption is likely rebuttable, as it would otherwise not be meaningful to call it a presumption.³⁰⁰

2.4.2. Codes of practice

124In the absence of a harmonised standard Article 3(27) AI Act: ‘harmonised standard’ means a harmonised standard as defined in Article 2(1), point (c), of Regulation (EU) No 1025/2012. , Article 53(4) indicates that GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. may comply with codes of practice within the meaning of Article 56.³⁰¹ Its last sentence indicates that such a code of practice must have been assessed adequate (‘approved’) within the meaning of Article 56(6) in order to serve its compliance function.³⁰² This approach differs from that adopted for high-risk systems, where the fall-back in the absence of harmonised standards Article 3(27) AI Act: ‘harmonised standard’ means a harmonised standard as defined in Article 2(1), point (c), of Regulation (EU) No 1025/2012. is the adoption of ‘ common specifications Article 3(28) AI Act: ‘common specification’ means a set of technical specifications as defined in Article 2, point (4) of Regulation (EU) No 1025/2012, providing means to comply with certain requirements established under this Regulation. ’ (Article 41 AI Act).

125While codes of practice are discussed more extensively in the chapter on Article 56,³⁰³ it is worth noting that Article 53(4) does not extend the presumption applicable to harmonised standards Article 3(27) AI Act: ‘harmonised standard’ means a harmonised standard as defined in Article 2(1), point (c), of Regulation (EU) No 1025/2012. to codes of practice. As a result, compliance with a code of practice does not amount to automatic compliance with the AI Act, nor does it generally give rise to a presumption of such compliance.³⁰⁴

126Even though codes of practice do thus not directly confer (a presumption of) compliance,³⁰⁵ they remain a valuable tool for interpreting the Act’s provisions. Given their assessment as adequate by the AI Office Article 3(47) AI Act: ‘AI Office’ means the Commission’s function of contributing to the implementation, monitoring and supervision of AI systems and general-purpose AI models, and AI governance, provided for in Commission Decision of 24 January 2024; references in this Regulation to the AI Office shall be construed as references to the Commission. , they also give rise to legitimate expectations.³⁰⁶ Nevertheless, the absence of an explicit presumption of conformity underscores the – at least theoretical³⁰⁷ – possibility that the Commission, the AI Office Article 3(47) AI Act: ‘AI Office’ means the Commission’s function of contributing to the implementation, monitoring and supervision of AI systems and general-purpose AI models, and AI governance, provided for in Commission Decision of 24 January 2024; references in this Regulation to the AI Office shall be construed as references to the Commission. , or national competent authorities Article 3(48) AI Act: ‘national competent authority’ means a notifying authority or a market surveillance authority; as regards AI systems put into service or used by Union institutions, agencies, offices and bodies, references to national competent authorities or market surveillance authorities in this Regulation shall be construed as references to the European Data Protection Supervisor. could still find a provider Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. in violation of the AI Act despite adherence to a code of practice deemed adequate.³⁰⁸ Nevertheless, the Commission Guidelines seem to equate compliance with an approved code of practice with compliance with the AI Act.³⁰⁹

2.4.3. Alternative adequate means

127There is no obligation for GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. to adhere to codes of practice or harmonised standards Article 3(27) AI Act: ‘harmonised standard’ means a harmonised standard as defined in Article 2(1), point (c), of Regulation (EU) No 1025/2012. , even when these are available. Irrespective of the availability of such measures, Article 53(4) indicates that providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. may also demonstrate compliance with the AI Act through ‘alternative adequate means’. This remains subject to assessment by the Commission.

128Even where providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. choose not to adhere to codes of practice or harmonised standards Article 3(27) AI Act: ‘harmonised standard’ means a harmonised standard as defined in Article 2(1), point (c), of Regulation (EU) No 1025/2012. , these instruments offer valuable guidance for alternative compliance routes – particularly regarding the types and extent of information to be documented, as well as the modalities (such as duration) of documentation. For instance, compliance might be achieved by documenting the same or similar information, for a comparable duration, but in a different format to that proposed by the codes of practice or harmonised standards Article 3(27) AI Act: ‘harmonised standard’ means a harmonised standard as defined in Article 2(1), point (c), of Regulation (EU) No 1025/2012. . The Commission Guidelines suggest conducting a ‘gap analysis’ to compare the adopted measures with those set out in the codes of practice. They also note that providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. choosing this path may face increased information requests, as it will generally be more challenging for the AI Office Article 3(47) AI Act: ‘AI Office’ means the Commission’s function of contributing to the implementation, monitoring and supervision of AI systems and general-purpose AI models, and AI governance, provided for in Commission Decision of 24 January 2024; references in this Regulation to the AI Office shall be construed as references to the Commission. to assess their compliance.³¹⁰

129A notable exception appears to be Article 53(1)(d), which seems to mandate use of the template provided by the AI Office Article 3(47) AI Act: ‘AI Office’ means the Commission’s function of contributing to the implementation, monitoring and supervision of AI systems and general-purpose AI models, and AI governance, provided for in Commission Decision of 24 January 2024; references in this Regulation to the AI Office shall be construed as references to the Commission. , as discussed earlier.³¹¹

2.5. Article 53(5): Delegated acts on Annex XI methodologies

130Articles 53(5) and 97 empower the Commission, pursuant to Article 290 TFEU, to adopt delegated acts providing further detail on measurement and calculation methodologies, ‘with a view to allowing for comparable and verifiable documentation’ to assess compliance with the provisions in Annex XI relating to the computational resources used to train the model, the model’s training time, and other relevant details of the training process (Annex XI(2)(d)), as well as the estimated or known energy consumption of the model (Annex XI(2)(e)), as discussed above.³¹² Such delegated acts are binding non-legislative acts and constitute secondary legislation under Article 290 TFEU, allowing the Commission to specify these technical elements. This mechanism is particularly pertinent given the absence of agreed technical standards for certain aspects.³¹³

131Interestingly, while Article 53(5) explicitly refers to delegated acts in relation to Annex XI Section 1(e) (and (d)), the Act – and Annex XI in particular – does not clarify the relationship with the phrase at the end of Annex XI Section 1: ‘With regard to point (e), where the energy consumption of the model is unknown, the energy consumption may be based on information about computational resources used.’ This could give rise to a contradiction if the Commission were to adopt a delegated act specifying an estimation method not based on computational resources used. Should this occur, it is relevant to note that the Commission is also empowered to adopt delegated acts to amend Annexes XI and XII in light of technological developments (Article 53(6)).³¹⁴ However, this estimation method is arguably not encompassed by that provision, leaving it as a valid alternative to the method identified in any delegated act adopted by the Commission.

132Procedurally, it is important to note that the European Parliament and the Council may object to a delegated act adopted by the Commission within three months of its notification.³¹⁵ This objection prevents the act from becoming binding, although it may become binding within that three-month period if both the European Parliament and the Council have informed the Commission that they do not intend to object (Article 97(6)).

2.6. Article 53(6): Delegated acts to amend Annexes XI and XII

133In contrast to Article 53(5), Article 53(6) confers a much broader mandate on the Commission to adopt delegated acts amending Annexes XI and XII. As discussed earlier, Annex XI sets out the technical documentation that GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. must supply, upon request, to the AI Office Article 3(47) AI Act: ‘AI Office’ means the Commission’s function of contributing to the implementation, monitoring and supervision of AI systems and general-purpose AI models, and AI governance, provided for in Commission Decision of 24 January 2024; references in this Regulation to the AI Office shall be construed as references to the Commission. and national competent authorities Article 3(48) AI Act: ‘national competent authority’ means a notifying authority or a market surveillance authority; as regards AI systems put into service or used by Union institutions, agencies, offices and bodies, references to national competent authorities or market surveillance authorities in this Regulation shall be construed as references to the European Data Protection Supervisor. , while Annex XII specifies the information that should be communicated to downstream AI system Article 3(1) AI Act: ‘AI system’ means a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. intending to integrate the GPAI model into their systems.

134Article 53(6) empowers the Commission to amend Annexes XI and XII ‘in light of evolving technological developments’. This should be seen as an attempt to ‘future-proof’ the AI Act,³¹⁶ and in particular, the regulatory oversight it establishes. In practical terms, this means the Commission may add elements to the documentation requirements for GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. (Annex XI) – for example, novel evaluation methods or new risk Article 3(2) AI Act: ‘risk’ means the combination of the probability of an occurrence of harm and the severity of that harm. assessment criteria – or require additional information to be shared with downstream providers Article 3(68) AI Act: ‘downstream provider’ means a provider of an AI system, including a general-purpose AI system, which integrates an AI model, regardless of whether the AI model is provided by themselves and vertically integrated or provided by another entity based on contractual relations. (Annex XII), such as emerging integration challenges or limitations. As a result, GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. cannot rely on a single documentation exercise but must remain vigilant and ensure their documentation remains up to date to avoid omitting any amendments.

135The procedure mirrors that described above (see also Article 97(2) AI Act). This power for the Commission to amend Annexes XII and, in particular, XI means that GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. cannot rely on a single documentation exercise; rather, they must remain vigilant and ensure their documentation remains up to date to avoid omitting any amendments.

2.7. Article 53(7): Confidentiality

136Article 53(7) provides that the AI Office Article 3(47) AI Act: ‘AI Office’ means the Commission’s function of contributing to the implementation, monitoring and supervision of AI systems and general-purpose AI models, and AI governance, provided for in Commission Decision of 24 January 2024; references in this Regulation to the AI Office shall be construed as references to the Commission. , the Commission, and national competent authorities Article 3(48) AI Act: ‘national competent authority’ means a notifying authority or a market surveillance authority; as regards AI systems put into service or used by Union institutions, agencies, offices and bodies, references to national competent authorities or market surveillance authorities in this Regulation shall be construed as references to the European Data Protection Supervisor. must treat any information received pursuant to Article 53 in accordance with the confidentiality obligations set out in Article 78. Notably, Article 78(2) requires that authorities generally exercise restraint in requesting sensitive data, ensuring that such requests are ‘strictly necessary’ for the exercise of their powers under the AI Act or their obligations under Regulation 2019/1020 on market surveillance and compliance of products. This requirement for restraint, however, is not expressly reiterated in Article 53 itself.

137Unlike Article 78, which applies only to ‘the Commission, market surveillance authorities Article 3(26) AI Act: ‘market surveillance authority’ means the national authority carrying out the activities and taking the measures pursuant to Regulation (EU) 2019/1020. and notified bodies Article 3(22) AI Act: ‘notified body’ means a conformity assessment body notified in accordance with this Regulation and other relevant Union harmonisation legislation. and any other natural or legal person involved in the application of this Regulation’, Article 53(7) adopts a broader scope, seemingly extending to downstream system providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. who have received information under Article 53(1)(b), insofar as the information is sensitive. However, as penalties under the AI Act are directed solely at GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. (Article 101), it appears that breaches by downstream system providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. would not be penalised under the AI Act. This further reinforces the earlier point that it is permissible, and arguably advisable, for GPAI model providers Article 3(3) AI Act: ‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge. to make access to relevant information conditional upon a non-disclosure agreement.³¹⁷

138As a direct consequence of this confidentiality requirement, the general public will not have access to any information documented pursuant to Article 53, with the notable exception of Article 53(1)(d) – the training data Article 3(29) AI Act: ‘training data’ means data used for training an AI system through fitting its learnable parameters. summary.³¹⁸