Article 53: Obligations for providers of general-purpose AI models
Article 53 requires providers of General Purpose AI (GPAI) models to draw up technical documentation, provide information to downstream providers, implement a copyright policy, and publish a summary of training data.
Praxikon tracks Article 53 (Obligations for providers of general-purpose AI models) under the EU AI Act, citing the source for every statement.
Official text
Source: EUR-Lex, Regulation (EU) 2024/1689. Text reproduced verbatim.
Download AI Act (PDF) →Decision tree
Are you the provider of a general-purpose AI model, and which obligations apply?
These obligations concern the model itself, not the application built on it. Most organisations that use or integrate a large language model are not model providers.
Question 1 · Art. 3(63)
Is it an AI model that displays significant generality, is capable of competently performing a wide range of distinct tasks and can be integrated into a variety of downstream systems or applications?
According to the guidelines an indicative criterion is that the model was trained with more than 10^23 FLOP and can generate language, images or video. A model for a single narrow task generally falls outside.
Full diagram
- 1
Is it an AI model that displays significant generality, is capable of competently performing a wide range of distinct tasks and can be integrated into a variety of downstream systems or applications?
Art. 3(63)
- Yes → Question 2
- No → Outcome A
- 2
Is the model only used for research, development or prototyping activities before it is placed on the market?
Art. 3(63)
- Yes → Outcome B
- No → Question 3
- 3
What do you do with the model?
Art. 3(3) · Art. 3(68)
- We develop it (or have it developed) and place it on the EU market under our own name → Question 5
- We modify another party’s model, for example by fine-tuning, and place that on the market → Question 4
- We integrate another party’s model into our own AI system → Outcome C
- We use an AI system that runs on such a model → Outcome D
- 4
Does your modification use more than one third of the training compute of the original model?
Art. 3(3) · richtsnoeren GPAI van de Commissie
- Yes → Question 5
- No → Outcome C
- 5
Is the cumulative training compute greater than 10^25 FLOP, does the model otherwise have high-impact capabilities, or has the Commission designated it as having systemic risk?
Art. 51(1) en 2
- Yes → Outcome H
- No → Question 6
- 6
Do you release the model under a free and open-source licence that allows access, usage, modification and distribution, and do you make the parameters, including the weights, the model architecture and the usage information publicly available?
Art. 53(2)
- Yes → Outcome E
- No → Question 7
- 7
Is your organisation established outside the EU?
Art. 54(1)
- Yes → Outcome G
- No → Outcome F
Outcome A · Does not apply
Not a general-purpose AI model
The obligations of Article 53 do not apply. If the model is used in an AI system, the rules for that system apply, depending on your role and the risk.
Art. 3(63)
Outcome B · Does not apply
Not yet a general-purpose AI model within the meaning of the law
Models used only for research, development or prototyping before being placed on the market fall outside the definition. As soon as you place the model on the market, the obligations apply.
Art. 3(63) · Art. 2(8)
Outcome C · Partly applies
You are a downstream provider, not a model provider
You need not comply with the model obligations of Article 53 yourself. You are entitled to the information and documentation from the model provider (Article 53(1)(b)), and the ordinary rules apply to the AI system you build, such as Article 50 and possibly the high-risk requirements.
Art. 3(68) · Art. 53(1)(b) · art. 50
Outcome D · Does not apply
No model obligations: you are a deployer
Whoever uses an AI system running on a general-purpose AI model has no obligations under Article 53. Article 4 and, depending on the use, Article 50(3) and (4) and Article 26 apply to you.
Art. 3(4) · art. 4 · Art. 50(3) · Art. 50(4) · art. 26
Outcome E · Partly applies
Open-source model: lighter obligations
The technical documentation for the AI Office and the information for downstream providers (paragraph 1(a) and (b)) and the authorised representative obligation (Article 54(6)) fall away. The copyright policy and the public summary of training content (paragraph 1(c) and (d)) still apply, as does the duty to cooperate with the Commission and the competent authorities (paragraph 3). If the model is provided against a price or otherwise monetised, the exception does not apply according to recital 103. The obligations have applied since 2 August 2025; since 2 August 2026 the Commission can enforce them and impose fines. For models placed on the market before 2 August 2025, the date is 2 August 2027.
Art. 53(2) · Art. 53(1)(c) en d · Art. 53(3) · Art. 54(6) · Art. 111(3) · overweging 103
Outcome F · Applies
You are the provider: the four obligations of Article 53 apply
Draw up and keep up to date technical documentation (Annex XI), provide information and documentation to downstream providers (Annex XII), put in place a copyright policy, and publish a sufficiently detailed summary of training content. Cooperate with the Commission and the competent authorities. The obligations have applied since 2 August 2025; since 2 August 2026 the Commission can enforce them and impose fines. For models placed on the market before 2 August 2025, the date is 2 August 2027. If you became the provider through a modification, the obligations are limited to your modification according to the Commission guidelines.
Art. 53(1) · Art. 53(3) · Art. 53(4) · art. 101 · Art. 111(3)
Outcome G · Applies
You are a provider outside the EU: also appoint an authorised representative
The four obligations of Article 53 apply. In addition, before placing on the market, you appoint in writing an authorised representative established in the EU, which among other things keeps the documentation for ten years and cooperates with the AI Office. The obligations have applied since 2 August 2025; since 2 August 2026 the Commission can enforce them and impose fines. For models placed on the market before 2 August 2025, the date is 2 August 2027.
Art. 53(1) · Art. 54(1) · Art. 54(3) · Art. 111(3)
Outcome H · Applies
Model with systemic risk: Article 53 and Article 55
Notify the model to the Commission without delay and within two weeks at the latest (Article 52(1)). You may argue that the model, despite meeting the threshold, does not present systemic risk; the Commission decides on that (Article 52(2) and (3)). In addition to the Article 53 obligations you perform model evaluations including adversarial testing, assess and mitigate systemic risks, report serious incidents to the AI Office and ensure adequate cybersecurity. The open-source exception does not apply. Established outside the EU: appoint an authorised representative. The obligations have applied since 2 August 2025; since 2 August 2026 the Commission can enforce them and impose fines. For models placed on the market before 2 August 2025, the date is 2 August 2027.
art. 51 · Art. 52(1) · Art. 52(2) · art. 53 · art. 54 · Art. 55(1) · Art. 111(3)
Official guidance on this article
10- Guidelines on the scope of obligations for providers of general-purpose AI models under the AI ActDetermines who qualifies as a provider of a general-purpose AI model at all and thus bears the documentation and copyright obligations, with the distinction between substantial and minor modifications.PublishedGuidelinesEuropean Commission18 Jul 2025
- The General-Purpose AI Code of PracticeTranslates the documentation, information and copyright obligations for general-purpose AI model providers into concrete practices, including a fillable documentation form and a copyright policyPublishedCode of practiceEuropean Commission / AI Office10 Jul 2025
- Explanatory Notice and Template for the Public Summary of Training Content for general-purpose AI modelsIs the mandatory template by which paragraph 1, point d is completed: without this format the public summary of training data does not meet the requirement.PublishedTemplateEuropean Commission / AI Office24 Jul 2025
- Questions and answers on the Code of Practice for General-Purpose AILinks the transparency and copyright chapters of the code to providers' documentation and copyright obligationsPublishedQ&AEuropean Commission, DG CONNECT20 Jul 2026
- Guidelines on obligations for General-Purpose AI providers - Questions & AnswersProvides the practical boundary of when an organisation that fine-tunes a model itself incurs the documentation and copyright obligations of a providerPublishedQ&AEuropean Commission, DG CONNECT11 Nov 2025
- General-Purpose AI Models in the AI Act - Questions & AnswersDescribes what technical documentation, information to users, copyright policy and summary of training data a provider must supplyPublishedQ&AEuropean Commission, DG CONNECT9 Sept 2025
- Commission Opinion on the assessment of the General-Purpose AI Code of PracticeMakes clear that signatories can rely on the code to demonstrably fulfil the obligations of this Article.PublishedOpinionEuropean Commission1 Aug 2025
- Opinion 28/2024 on certain data protection aspects related to the processing of personal data in the context of AI modelsDetermines when a general-purpose AI model trained on personal data itself still contains personal data, so that the provider may qualify not only under the general-purpose AI obligations from the AI Act but also as a processor under data protection lawPublishedOpinionEDPB (European Data Protection Board)17 Dec 2024
- AI Regulation timeline (Algoritmekader)Places the documentation and information obligations for providers of general-purpose AI models on 2 August 2025.PublishedNational guidanceMinistry of the Interior and Kingdom Relations (Algorithm Framework)17 Jul 2025
- Guidance on generative AI and the GDPRSets out alongside the general-purpose AI obligations in the AI Act the Dutch GDPR standard for developing and deploying generative models, so that organisations must establish both tracks in parallelPublishedGuidelinesDutch Data Protection Authority (AP)13 Jul 2026
Related recitals
The notion of general-purpose AI models should be clearly defined and set apart from the notion of AI systems to enable legal certainty. The definition should be based on the key functional characteri…
Whereas the generality of a model could, inter alia, also be determined by a number of parameters, models with at least a billion of parameters and trained with a large amount of data using self-super…
Large generative AI models are a typical example for a general-purpose AI model, given that they allow for flexible generation of content, such as in the form of text, audio, images or video, that can…