Skip to main content
Praxikon

Article 53 AI Act: the obligations of every provider of a general-purpose AI model

··17 min read

A municipality uses a supplier's chatbot built on a large language model. Its innovation team wants to go further: download an open language model, fine-tune it on the municipality's own policy texts and share the result with other municipalities. Does the municipality then become a provider of a general-purpose AI model, with the obligations of Article 53? And what can it expect from its supplier if it does not?

The examples are fictional.

In short

Every provider of a general-purpose AI model has four obligations (paragraph 1): technical documentation for the authorities (point (a)), information for those who integrate the model (point (b)), a copyright policy (point (c)) and a public summary of the training content (point (d)). For a model under a free and open-source licence, with its parameters made public and without monetisation, only points (a) and (b) fall away, and not where the model has systemic risk (paragraph 2). Article 53 has applied since 2 August 2025; the Commission has been able to impose fines since 2 August 2026.

According to the Commission guidelines, someone who fine-tunes a model only becomes the provider of the modified model if the modification significantly changes the model's generality, capabilities or systemic risk; as an indicative criterion, they point to more than a third of the compute of the original training (paragraphs 59 and 60). In our reading, most municipalities and banks do not become the provider of the model, but often are the deployer or provider of the AI system built on it.

The additional obligations for models with systemic risk are in our piece on Article 55, the authorised representative for providers outside the EU in the piece on Articles 22 and 54 and supervision in the piece on Articles 90 to 94.

What is a general-purpose AI model?

Article 3(63) defines it as an AI model that "displays significant generality and is capable of competently performing a wide range of distinct tasks" and can be integrated into a variety of downstream systems or applications, except AI models used for research, development or prototyping activities before they are placed on the market. The Commission guidelines on general-purpose AI models (C(2025) 7719 final, 19 November 2025) set as an indicative criterion training compute greater than 10^23 FLOP and the ability to generate language (text or audio), text-to-image or text-to-video (paragraph 17). The criterion is indicative: a model that meets it but can only competently perform a narrow set of tasks, such as transcribing speech to text, is exceptionally out of scope, and a model that does not meet it but displays significant generality is exceptionally in scope (paragraph 20 and the examples that follow).

Article 53 paragraph by paragraph

Paragraph 1, point (a): technical documentation. The provider draws up and keeps up-to-date the technical documentation of the model, "including its training and testing process and the results of its evaluation", containing at a minimum the information in Annex XI, for the purpose of providing it, upon request, to the AI Office and the national competent authorities. Section 1 of that Annex applies to all providers, "as appropriate to the size and risk profile of the model", from architecture and licence to data provenance and energy consumption. Section 2 applies only where there is systemic risk. Point (a) does not require publication.

Paragraph 1, point (b): information for those who integrate the model. The provider draws up, keeps up-to-date and makes available information and documentation "to providers of AI systems who intend to integrate the general-purpose AI model into their AI systems". That information enables them to have a good understanding of the capabilities and limitations of the model and to comply with their own obligations, and contains at a minimum the elements in Annex XII, such as the acceptable use policies and information on the training data. All of this is without prejudice to the protection of intellectual property rights and trade secrets.

Paragraph 1, point (c): copyright policy. The provider puts in place "a policy to comply with Union law on copyright and related rights", in particular to identify and comply with, including through state-of-the-art technologies, a reservation of rights expressed pursuant to Article 4(3) of Directive (EU) 2019/790. That provision makes the general text and data mining exception in paragraph 1 of that Article conditional on the use not having been "expressly reserved by their rightholders in an appropriate manner, such as machine-readable means in the case of content made publicly available online". The AI Office monitors whether the provider has put a policy in place and published the summary, but does not carry out a work-by-work assessment of the training data for copyright compliance; enforcement of copyright itself is unaffected (recital 108).

Paragraph 1, point (d): summary of the training content. The provider draws up and makes publicly available "a sufficiently detailed summary about the content used for training", according to a template provided by the AI Office. We discuss the template in our piece on the public summary. According to the explanatory notice, it is made public at the latest when the model is placed on the market (point 32) and on further training updated at six-month intervals, or sooner for a materially significant update (point 29).

Paragraph 2: free and open-source licence. Points (a) and (b) do not apply to a model released under a free and open-source licence that allows for the access, usage, modification and distribution of the model, and whose parameters, including the weights, the information on the model architecture and the information on model usage, are made publicly available. The exception does not apply where there is systemic risk, and never to points (c) and (d): an open-source release does not necessarily reveal substantial information on the training data or on copyright compliance (recital 104). According to the guidelines, a limitation to non-commercial use (paragraph 80) and monetisation, such as a requirement to purchase support (paragraph 83), among other things disqualify the exception. Access, use, modification or distribution that requires the collection or other processing of personal data is treated as monetisation, unless that processing is exclusively and strictly limited to the security of the model, without any commercial or financial gain (paragraph 84).

Paragraph 3: cooperation. The provider cooperates "as necessary" with the Commission and the national competent authorities.

Paragraph 4: codes of practice and standards. Until a harmonised standard is published, the provider may rely on a code of practice within the meaning of Article 56 to demonstrate compliance with paragraph 1; such a code does not give rise to a presumption of conformity (recital 41 of Regulation (EU) 2026/1744). Compliance with a European harmonised standard grants a presumption of conformity to the extent that the standard covers those obligations. A provider who does not adhere to an approved code of practice or does not comply with a European harmonised standard demonstrates alternative adequate means of compliance for assessment by the Commission. The General-Purpose AI Code of Practice of 10 July 2025 has two chapters for Article 53: Transparency (points (a) and (b)) and Copyright (point (c)); the Commission and the AI Board confirmed the code as adequate on 1 August 2025. More in our piece on the code. For downstream providers, Measure 1.2 of the Transparency Chapter matters most: signatories provide them with additional information upon request insofar as it is necessary to understand the capabilities and limitations of the model and to comply with their own obligations, within a reasonable timeframe and no later than 14 days, save for exceptional circumstances, without prejudice to intellectual property rights and trade secrets. This binds signatories only and does not apply to open models covered by paragraph 2.

Paragraphs 5 and 6: delegated acts. The Commission may detail measurement and calculation methodologies for Annex XI and amend Annexes XI and XII. On EUR-Lex we found no delegated act under either power on 22 November 2026; that does not mean that none is being prepared.

Paragraph 7: confidentiality. Information or documentation obtained pursuant to Article 53, including trade secrets, is treated in accordance with the confidentiality obligations in Article 78.

Since when?

Chapter V, which contains Article 53, has applied since 2 August 2025 (Article 113, third paragraph, point (b)). Article 101 is excepted from that date: the Commission has been able to impose fines of up to 3 % of annual worldwide turnover or EUR 15 000 000, whichever is higher, since 2 August 2026. Providers of models placed on the market before 2 August 2025 take the necessary steps to comply by 2 August 2027 (Article 111(3)).

Regulation (EU) 2026/1744, the Digital Omnibus on AI, did not amend Article 53. It did replace Article 56(6): the Commission can no longer approve a code and give it general validity by implementing act, but assesses whether a code covers the obligations of Articles 53 and 55 and publishes its assessment of adequacy.

When does a municipality or bank become a provider itself?

A provider is anyone that "develops" a general-purpose AI model or "has" one "developed" and places it on the market "under its own name or trademark, whether for payment or free of charge" (Article 3(3)). Placing on the market means the first making available on the Union market (point (9)).

Buying in or using via an API. The actor that develops the model and makes it available for the first time to a downstream actor on the Union market is the provider of the model (guidelines, paragraph 55). If the municipality buys a chatbot, it is the deployer of that AI system (Article 3(4)). If it builds the chatbot itself on a model accessed via an API and puts it into service under its own name, it is both provider and deployer of that system, but not the provider of the model.

Fine-tuning. According to the Commission, not every modification makes someone a provider. A downstream modifier only becomes the provider of the modified model if the modification "leads to a significant change in the model's generality, capabilities, or systemic risk" (paragraph 59). The indicative criterion: the training compute used for the modification is greater than a third of the training compute of the original model (paragraph 60). If that value is unknown and cannot be estimated, the threshold is a third of 10^23 FLOP, or a third of 10^25 FLOP if the original model has systemic risk (paragraph 61). The Commission writes that currently few modifications may meet the criterion (paragraph 64). Whether the original provider or the downstream actor modifies the model is assessed case by case; with fine-tuning via an API, who has control over the model's weights may be an important factor (footnote 13).

If the modification meets the threshold, the obligations are limited to the modification and the data used in it (paragraph 65, recital 109). Whoever modifies a model with systemic risk in this way has a model with systemic risk itself and notifies the Commission (paragraphs 67 and 68).

An own model. When a provider integrates its own model into its own AI system that is made available on the market or put into service, the model is considered placed on the market (recital 97, paragraph 54). Use for purely internal processes that are not essential for providing a product or a service to third parties, where the rights of natural persons are not affected, falls outside the obligations for models under recital 97; use that is essential for a service to third parties or that affects those rights is given by the guidelines as an example of placing on the market, to be assessed case by case (paragraphs 51 and 52).

Back to the municipality. The team fine-tunes an open model on a few thousand policy texts. In our reading, that takes a fraction of a third of the original training compute, also measured against the fallback threshold of paragraph 61. The municipality then does not become the provider of the modified model; because it does not take on that role, sharing the result with other municipalities does not, in our reading, change this, and the original developer remains the provider of the model. If it puts a chatbot for residents into service with it under its own name, it is provider and deployer of that AI system, with among other things the provider's obligation to design the chatbot so that residents are informed that they are interacting with an AI system, unless this is obvious to a reasonably well-informed resident (Article 50(1), applicable since 2 August 2026; see making a chatbot recognisable as AI).

The information under paragraph 1, point (b), is intended for providers of AI systems that integrate the model. If the municipality buys in, the supplier receives that information; in our reading the municipality does well to require in the contract that the supplier passes on the relevant parts, such as capabilities and limitations, acceptable use policies and information on the training data. If it builds itself, it is a downstream provider entitled to that information and may lodge a duly reasoned complaint alleging an infringement (Article 89(2)).

Bank or insurer. A bank that fine-tunes an open model on internal procedures for a staff assistant does not, in our reading, become a model provider either. If a bank or insurer develops a general-purpose AI model itself and uses it for services to customers, the model should in principle be considered placed on the market under recital 97 and paragraph 51, and the obligations of Article 53 apply. A system built on an external model that evaluates the creditworthiness of natural persons, or that is intended to be used for risk assessment and pricing in relation to natural persons in the case of life and health insurance, is in principle high-risk (Annex III, point 5(b) and (c)), with obligations from 2 December 2027; for a system already placed on the market or put into service before that date, only if its design is subject to significant changes (Article 111(2)). If it builds such a system itself under its own name, it needs the information under paragraph 1, point (b), for its own obligations as provider, such as the technical documentation of Article 11.

Where the GDPR comes in

Training data often contain personal data; Article 53 does not deal with that, the GDPR does. If the municipality fine-tunes a model on documents containing personal data, it needs, as controller, a legal basis (Article 6 GDPR) and assesses whether a DPIA is required beforehand (Article 35 GDPR; see when a DPIA is required for AI).

According to EDPB Opinion 28/2024 on AI models (17 December 2024), AI models trained on personal data cannot, in all cases, be considered anonymous; this is assessed case by case (paragraph 34). According to the EDPB, supervisory authorities should take into account whether a controller deploying a model developed by another has conducted, as part of its accountability obligations, an appropriate assessment to ascertain that the model was not developed by unlawfully processing personal data, considering for example the source of the data and whether a supervisory authority or court has found an infringement; the degree of the assessment depends on the risks (paragraphs 129 and 130). The information under Annex XII and the summary under point (d) help with this. See also the Dutch DPA's investigation into Reddit data for AI training.

Which AI Act and GDPR obligations apply to your use case? See the case check.

Frequently asked questions about Article 53

Sources

EUR-Lex: Regulation (EU) 2024/1689 (AI Act), original text, recitals 97, 104, 108 and 109 (13 June 2024, published 12 July 2024, accessed 22 November 2026)
EUR-Lex: Regulation (EU) 2026/1744 (Digital Omnibus on AI), recital 41 and Article 1, point 21; Article 53 not amended (8 July 2026, published 24 July 2026, accessed 22 November 2026)
European Commission, AI Office: Explanatory Notice and Template for the Public Summary of Training Content for general-purpose AI models, C(2025) 8311 final, points 29 and 32 (published 24 July 2025, version of 5 December 2025, accessed 22 November 2026)
European Commission: Commission Opinion on the assessment of the General-Purpose AI Code of Practice (1 August 2025, accessed 22 November 2026)
EUR-Lex: Regulation (EU) 2016/679 (GDPR), Articles 6 and 35 (27 April 2016, accessed 22 November 2026)

Newsletter

Every Tuesday, the AI Act week ahead in 5 minutes

A practical briefing on deadlines, new guidance and enforcement, so you know what matters this week. No spam and you can unsubscribe in one click.

Practical and short · No spam · One-click unsubscribe

Written by

Zahed Ashkara, freelance AI and privacy consultant and legal adviser

Zahed Ashkara

Freelance AI and privacy consultant and legal adviser · AI governance & safety

Zahed Ashkara is a freelance AI and privacy consultant and legal adviser for the GDPR and the EU AI Act. Organisations hire him as AI governance consultant, privacy counsel, external DPO or AI governance lead, have him carry out a DPIA, FRIA or AI Act classification, or book an in-company training. He focuses on AI governance and safety, including AI agents. He works exclusively as a freelancer, through Embed AI (trade name Praxikon). He is a Certified AI Compliance Officer (CAICO) and has been a member of the NEN standards committee Artificial Intelligence & Big Data since July 2026.

Follow on LinkedInView expert profile

Use these sources in your AI assistant too: Praxikon in Claude and other assistants

Real-World Stories

Discover how organizations apply this in practice

Related articles

Article 55 AI Act: general-purpose AI models with systemic risk and the requirements for external evaluators

Article 55 of the AI Act gives providers of the most capable AI models four additional obligations: model evaluation, risk assessment, serious incident reporting and cybersecurity. What each paragraph says, whether the law itself requires external evaluators, what the AI Office wrote on 21 October 2026 about their qualification and independence, and what a municipality or bank buying such a model can do with it.

EU AI ActArticle 55General-purpose AI models

Article 101 AI Act: Commission fines for providers of general-purpose AI models

Article 101 of the AI Act lets the European Commission fine providers of general-purpose AI models, up to 3% of worldwide annual turnover or 15 million euro. Which four kinds of conduct can be fined, which safeguards apply, what Implementing Regulation (EU) 2026/1755 provides, how it relates to Articles 99 and 100, and whether a municipality or bank can receive such a fine.

EU AI ActArticle 101Fines