Skip to main content
Praxikon

Article 56 AI Act: what the code of practice for general-purpose AI models is and is not

··14 min read

A municipality buys a writing assistant for its staff. According to the tender, the language model behind it comes from a provider that "has signed the EU code of practice". Is the model approved, and what can the municipality do with that in procurement?

The answer starts with Article 56. The examples are fictional.

In short

Article 56 has the AI Office encourage codes of practice that cover at least the obligations in Articles 53 and 55. Since 27 July 2026 the Commission, taking utmost account of the opinion of the Board, assesses whether a code is adequate and publishes that assessment (paragraph 6, replaced by Regulation (EU) 2026/1744). Approval by implementing act has been removed.

The code of practice of 10 July 2025 was found adequate on 1 August 2025. A provider that adheres to it may rely on it to demonstrate compliance (Article 53(4), Article 55(2)), but gets no presumption of conformity. Non-signatories demonstrate alternative adequate means of compliance. In our reading, a signature is a useful procurement signal for a municipality or bank, not proof of compliance.

The obligations are covered in our pieces on Article 53 and Article 55, the code's content in the code of practice piece and contract clauses in the procurement piece.

Article 56 paragraph by paragraph

Paragraph 1: the AI Office encourages. The AI Office "shall encourage and facilitate" the drawing up of codes of practice at Union level. It is not a separate body: references to it are construed as references to the Commission (Article 3, point (47)).

Paragraph 2: the minimum content. The AI Office and the Board shall aim to ensure that the codes cover at least the obligations in Articles 53 and 55, including keeping the information under Article 53(1), points (a) and (b), up to date, the level of detail of the training content summary, and the identification, assessment and management of systemic risks at Union level. "Shall aim to ensure" is an effort, not a guarantee.

Paragraph 3: who takes part. The AI Office may invite all model providers and relevant national competent authorities. Other stakeholders, "such as downstream providers and independent experts", may help; recital 116 also mentions the scientific panel.

Paragraphs 4 and 5: objectives and reporting. The AI Office and the Board shall aim to ensure that a code sets out its objectives and contains commitments or measures, including key performance indicators as appropriate. The AI Office shall aim to ensure that participants report regularly; that too is an effort, not a statutory reporting duty for the participant.

Paragraph 6: assessment, no longer approval. In the original text, the AI Office and the Board assessed whether a code covers the obligations in Articles 53 and 55. A second subparagraph allowed the Commission, "by way of an implementing act", to approve a code and "give it a general validity within the Union". Article 1, point (21), of Regulation (EU) 2026/1744, the Digital Omnibus on AI, replaces paragraph 6. Monitoring and publication were already there. What is new is that the Board no longer co-assesses but advises: the Commission, "taking utmost account of the opinion of the Board", assesses whether the codes cover those obligations. And the approval subparagraph is gone.

Recital 41 of Regulation (EU) 2026/1744 explains why: these codes of practice "have limited legal effect, and in particular do not grant a presumption of conformity", so approval is not strictly necessary; providers should be able to rely on codes assessed as adequate under paragraph 6. We found no earlier approval decision on EUR-Lex. The new text applies since 27 July 2026, the day of entry into force (Article 4 of that Regulation).

Three references were left in the Regulation: Article 53(4) and Article 55(2) still refer to "an approved code of practice" and Article 89(1) to "approved codes of practice". In our reading, given that recital of the omnibus, this now means a code assessed as adequate under paragraph 6. Points 91 and 96 of the GPAI guidelines also still describe the old regime (assessment by the AI Office and the Board, approval by implementing act); on those points they have been outdated since 27 July 2026.

Paragraphs 7 and 8: adherence and review. The AI Office may invite all providers to adhere to the codes. A provider of a model without systemic risk may limit its adherence to the obligations in Article 53, unless it declares explicitly its interest to join the full code. The AI Office also encourages their review and adaptation, in particular in light of emerging standards.

Paragraph 9: deadline and fallback. Codes had to be ready by 2 May 2025. If, by 2 August 2025, a code could not be finalised, or if the AI Office deems it not adequate following its assessment under paragraph 6, the Commission may provide common rules for the obligations in Articles 53 and 55 by means of implementing acts, under the examination procedure in Article 98(2). Paragraph 9 still names the AI Office; as that is a reference to the Commission, in our reading it fits the new paragraph 6. We found no such rules on EUR-Lex on 22 November 2026, which does not rule out a draft.

The code of practice of 10 July 2025

The Commission received the final General-Purpose AI Code of Practice on 10 July 2025, after the deadline of 2 May but before 2 August 2025. It has three chapters: Transparency and Copyright for all providers (Article 53), and Safety and Security only for providers of models with systemic risk (Article 55).

On 1 August 2025 the Commission concluded in its opinion C(2025) 5361 final that the code adequately covers the obligations in Articles 53 and 55 and meets the aims of Article 56 (point 57). The Board concluded the same separately. Both assessments fell under the old paragraph 6, and the omnibus contains no transitional rule. We found no new assessment. In our reading the 2025 assessment remains usable: the Commission assessed the code itself, and its page still presents it as adequate. The opinion does not rule out that the Commission finds the code inadequate later (point 3).

On 22 November 2026 the Commission's page listed 22 signatories, including Amazon, Anthropic, Google, IBM, Microsoft, Mistral AI and OpenAI. xAI signed only the Safety and Security Chapter and, according to the Commission, has to demonstrate transparency and copyright compliance via alternative adequate means; signatures can be withdrawn at any time.

What the code is in law

A way to demonstrate compliance. A provider "may rely on" a code of practice to demonstrate compliance with Article 53(1), "until a harmonised standard is published" (Article 53(4)); the same applies to Article 55(1) (Article 55(2)). Recital 117 calls codes "a central tool". According to the GPAI guidelines (C(2025) 7719 final, 19 November 2025), such a code is "a straightforward way of demonstrating compliance", and for signatories the Commission focuses enforcement on adherence to the code. Any opt-out from a chapter loses that benefit for that part (point 91).

No presumption of conformity. That presumption is granted by a European harmonised standard, to the extent that it covers the obligations (Article 53(4), Article 55(2); guidelines point 97), and also by Commission common specifications, which may be adopted for the obligations in Chapter V (Article 41(1) and (3)). We found no standard or common specification for Articles 53 or 55 on 22 November 2026.

No new obligations. According to the Commission, the code does not impose obligations beyond the AI Act. A signatory does commit to its measures; when fixing a fine the Commission takes into account commitments made in relevant codes of practice (Article 101(1); guidelines point 93).

Providers that do not sign. A provider that adheres neither to the code nor to a harmonised standard "shall demonstrate alternative adequate means of compliance for assessment by the Commission" (Article 53(4), Article 55(2)). According to the guidelines, it reports its measures to the AI Office, explains how they ensure compliance, for instance with a gap analysis against the code, and may face more requests for information (point 92). Not signing is therefore not an infringement; not complying is.

Since when. Chapter V, including Article 56, has applied since 2 August 2025 (Article 113, third paragraph, point (b)); the Commission can impose fines since 2 August 2026. For models placed on the market before 2 August 2025, the deadline is 2 August 2027 (Article 111(3)).

Not to be confused with Article 95 or the GDPR

A code of conduct under Article 95 is a voluntary code for AI systems other than high-risk systems, without evidential effect (see our piece on Article 95). Nor is the code of practice a code of conduct within the meaning of Article 40 GDPR. Such a code is prepared by associations or bodies representing controllers or processors; the competent supervisory authority approves it (paragraphs 2 and 5) and a body referred to in Article 41 GDPR monitors compliance (paragraph 4). The code of practice for AI models concerns AI Act obligations; in our reading, a signature says nothing about whether the training of the model complies with the GDPR.

What does a municipality or bank gain?

The code binds only the signatory; the municipality is not a party. It buys an AI system and is its deployer (Article 3, point (4)). The supplier that integrates the language model is a downstream provider (Article 3, point (68)), and the model provider's information obligation runs to that supplier (Article 53(1), point (b), with Annex XII), except for a model under a free and open-source licence without systemic risk (Article 53(2)). Under Measure 1.2 of the Transparency Chapter, signatories also give such downstream providers additional information on request, insofar as it is necessary to understand the model and to comply with their own AI Act obligations and subject to trade secrets, within a reasonable timeframe and no later than 14 days, save for exceptional circumstances.

In our reading, the municipality asks its supplier four questions, not required by law:

  1. Which model and version is used, is the model provider on the Commission's list, and for which chapters?
  2. If the model provider did not sign, or only in part: how does it otherwise demonstrate compliance with Article 53, and where relevant 55?
  3. Does the supplier pass on the relevant information from Annex XII, such as tasks, acceptable use policies, input and output format, and licence?
  4. Will the supplier report if the signature is withdrawn or the model is replaced?

A signature does not prove compliance; whether the provider adheres to the code is for the Commission to assess. It does not change the municipality's own obligations and concerns the model, not the obligations in Article 50 for the AI system; for marking and labelling AI content there is a separate code of practice.

Bank or insurer. A bank that builds a staff assistant on a model accessed through an API and puts it into service under its own name is both provider and deployer of that AI system (Article 3, points (3) and (4)) and itself a downstream provider. It is then itself entitled to the information under Article 53(1), point (b), signature or not, unless the model is released under a free and open-source licence and has no systemic risk (Article 53(2)). In our reading, a signature mainly adds a public point of contact and the 14-day timeframe from Measure 1.2. If information is not forthcoming, the bank, like an insurer in the same position, may lodge a duly reasoned complaint (Article 89(2)).

Check your own use case with the case check.

Frequently asked questions on Article 56

Sources

EUR-Lex: Regulation (EU) 2024/1689 (AI Act), original text, Article 56(6) before the amendment and recitals 116 and 117 (13 June 2024, published 12 July 2024, accessed 22 November 2026)
EUR-Lex: Regulation (EU) 2026/1744 (Digital Omnibus on AI), recital 41, Article 1, point 21, and Article 4 (8 July 2026, published 24 July 2026, accessed 22 November 2026)
European Commission: The General-Purpose AI Code of Practice: chapters, assessment and list of signatories (code of 10 July 2025, list accessed 22 November 2026)
European Commission: Code of Practice for General-Purpose AI Models, Transparency Chapter, Measure 1.2 (10 July 2025, accessed 22 November 2026)
European Artificial Intelligence Board: Conclusion of the Artificial Intelligence Board on the Assessment of the General-Purpose AI Code of Practice (1 August 2025, accessed 22 November 2026)
European Commission: Questions and answers on the code of practice for General-Purpose AI (accessed 22 November 2026)
European Commission: Signing the General-Purpose AI code of practice (questions and answers) (accessed 22 November 2026)
EUR-Lex: Regulation (EU) 2016/679 (GDPR), Articles 40 and 41 (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 53 AI Act: the obligations of every provider of a general-purpose AI model

Article 53 of the AI Act gives every provider of a general-purpose AI model four obligations: technical documentation, information for those who integrate the model, a copyright policy and a public summary of the training content. What each paragraph says, what the open-source exception does and does not cover, and when a municipality or bank that fine-tunes a model becomes a provider itself.

EU AI ActArticle 53General-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