The GPAI Code of Practice: A Complete Guide for AI Model Providers

The General-Purpose AI Code of Practice is a voluntary compliance instrument published by the European Commission on 10 July 2025. It sets out how providers of general-purpose AI models can demonstrate compliance with their obligations under Articles 53 and 55 of the EU AI Act.

The Code is voluntary. The underlying obligations are not. Article 53 and Article 55 have applied since 2 August 2025 regardless of whether a provider signs.

What changes this month is enforcement. From 2 August 2026, the Commission’s enforcement powers under the AI Act entered into application. The AI Office can now issue requests for information, demand model access, order recalls, and impose fines of up to €15 million or 3% of global annual turnover under Article 101.

Roughly 190 organisations have signed. Signatories were publicly listed in July 2026, ahead of the enforcement date.

This guide covers what the Code requires across its three chapters, what signing does and does not do, how the enforcement posture differs between signatories and non-signatories, and what providers should do now.

Key Definitions of GPAI Code of Practice

Term	Definition
GPAI model	A general-purpose AI model under Article 3(63) of the EU AI Act. Trained at scale, displays significant generality, capable of a wide range of distinct tasks
GPAI Code of Practice	The voluntary compliance instrument published 10 July 2025 covering Articles 53 and 55 obligations
Systemic risk model	A GPAI model with high-impact capabilities, presumed where training compute exceeds 10^25 FLOPs
AI Office	The European AI Office, the supervisory body for GPAI model providers
Signatory	A provider that has signed the Code and committed to its measures
Model Documentation Form	The standardised documentation template under the Transparency chapter
Safety and Security Framework	The systemic risk management document required under the Safety and Security chapter
TDM	Text and data mining. Article 4(3) of Directive (EU) 2019/790 permits rights holders to reserve rights against it
Signatory Taskforce	The AI Office-chaired body established to facilitate coherent application of the Code

What Is the GPAI Code of Practice

Article 56 of the EU AI Act directed the AI Office to facilitate the drawing up of codes of practice at Union level to contribute to the proper application of the regulation. The GPAI Code of Practice is the result.

GPAI Code of Practice was developed through a multi-stakeholder process involving nearly 1,000 participants, drafted by independent experts, and endorsed by the Commission and the AI Board comprising member state representatives. The final version was published on 10 July 2025, three weeks before the GPAI obligations took effect.

The Code is organised into three chapters. Transparency and Copyright apply to all GPAI model providers. Safety and Security applies only to providers of models with systemic risk, currently a small group of between five and fifteen companies worldwide.

There are twelve commitments in total: one in the Transparency chapter, one in the Copyright chapter, and ten in the Safety and Security chapter.

Chapter	Applies to	Commitments
Transparency	All GPAI model providers	1
Copyright	All GPAI model providers	1
Safety and Security	Systemic risk model providers only	10

GPAI Code of Practice Chapter 1: Transparency

The Transparency chapter operationalises Article 53(1)(a) and (b) of the AI Act, covering technical documentation and information provision to downstream providers.

The central mechanism is the Model Documentation Form, a standardised template signatories complete for each GPAI model placed on the EU market. The Form consolidates the information the AI Act requires into a single structured document rather than leaving each provider to design its own approach.

What the Model Documentation Form covers

Section	Content
Model identification	Name, version, release date, model family
Model properties	Architecture, modality, input and output types, model size
Training process	Training methodology, compute used, energy consumption
Training data	Data types, sources, provenance, curation and filtering methodology
Distribution channels	How the model is made available, licensing terms
Intended and prohibited uses	Acceptable use policy and restrictions
Technical means for integration	Information downstream providers need to integrate the model
Evaluation results	Benchmarks, performance metrics, known limitations

Retention and access. Documentation must remain current, reflecting material changes, and be preserved for at least ten years after the model’s initial release. Signatories must publish contact details enabling the AI Office and downstream providers to request documentation access.

The downstream dimension. The Transparency chapter addresses a structural problem in the AI value chain. A downstream provider building a high-risk AI system on a foundation model cannot complete its own Annex IV technical documentation without information only the model provider holds. The Model Documentation Form is designed to make that information available in a usable, standardised form.

The training data summary. Separately from the Model Documentation Form, Article 53(1)(d) requires providers to publish a sufficiently detailed public summary of the content used for training, using a template the Commission published on 24 July 2025. This is a public-facing document, distinct from the documentation held for the AI Office.

GPAI Code of Practice Chapter 2: Copyright

The Copyright chapter operationalises Article 53(1)(c), which requires GPAI model providers to put in place a policy to comply with Union copyright law, and in particular to identify and comply with rights reservations expressed under Article 4(3) of Directive (EU) 2019/790.

Signatories undertake to establish, maintain, and execute a copyright policy applying to all GPAI models distributed within the EU.

What the copyright policy must address

Element	Requirement
Lawful access	Reproduce and extract only lawfully accessible copyright-protected content
Circumvention	Do not circumvent technological protection measures
Paywall respect	Do not crawl content behind paywalls without authorisation
Rights reservation identification	Identify and comply with TDM opt-outs expressed in machine-readable form
Robots.txt	Respect robots.txt as a rights reservation mechanism
Piracy sources	Exclude websites persistently and repeatedly infringing copyright from crawling
Output safeguards	Implement measures to mitigate the risk of models producing infringing outputs
Complaints mechanism	Designate a point of contact for rights holders and handle complaints

The rights reservation problem. Article 4(3) of the Copyright Directive permits rights holders to reserve rights against text and data mining, but the Directive did not specify a technical standard for expressing that reservation. Robots.txt is the most established mechanism and the GPAI Code of Practice treats it as a valid expression of reservation. Other machine-readable methods exist but are inconsistently implemented.

The Commission launched a stakeholder consultation on protocols for reserving rights from text and data mining specifically to address this gap. The technical standard remains unsettled, which creates genuine uncertainty for providers attempting to comply in good faith.

No safe harbour. GPAI Code of Practice is explicit that adherence to the chapter does not substitute for the underlying obligation to comply with Union and national copyright law. Signatories retain full responsibility for legal compliance independently of the Code.

GPAI Code of Practice Chapter 3: Safety and Security

The Safety and Security chapter applies only to providers of GPAI models with systemic risk, meaning models with high-impact capabilities. Article 51(2) presumes systemic risk where cumulative training compute exceeds 10^25 floating-point operations. The AI Office may also designate models below that threshold.

Ten commitments make up this chapter. Together they describe a systemic risk management lifecycle running from model development through to post-market monitoring.

Commitment	Substance
1. Safety and Security Framework	Adopt, implement, and maintain a documented Framework setting out risk management processes
2. Systemic risk identification	Identify systemic risks arising from the model, using a structured taxonomy
3. Systemic risk analysis	Analyse identified risks, including through state-of-the-art model evaluations
4. Systemic risk acceptance determination	Determine whether risks are acceptable, against defined criteria, before proceeding
5. Safety mitigations	Implement mitigations proportionate to identified risks
6. Security mitigations	Protect model weights and infrastructure against unauthorised access and exfiltration
7. Safety and Security Model Reports	Produce a report for each systemic risk model, provided to the AI Office
8. Systemic risk responsibility allocation	Allocate clear internal responsibility for systemic risk management across governance levels
9. Serious incident reporting	Establish processes to detect, document, and report serious incidents to the AI Office
10. Non-retaliation protections	Protect workers who report concerns about systemic risk from the model

The Safety and Security Framework. This is the anchoring document. It describes how the provider identifies, analyses, evaluates, and mitigates systemic risk across the model lifecycle. It is a governance artefact rather than a technical one, comparable in function to a quality management system under Article 17 for high-risk AI systems.

Model evaluation. Commitment 3 requires state-of-the-art evaluations, including adversarial testing where appropriate to the risk profile. The Code does not prescribe specific benchmarks, which reflects the pace at which evaluation methodology is developing.

Serious incident reporting. Article 55(1)(c) requires reporting of serious incidents to the AI Office without undue delay. The Code operationalises this but leaves the criteria for what constitutes a serious incident partly undefined, which is one of the acknowledged gaps in the instrument.

Non-retaliation. Commitment 10 is unusual in a compliance code and reflects concerns raised during drafting about internal suppression of safety findings. It commits signatories to protecting workers who raise systemic risk concerns.

What Signing Actually Does

The Code is voluntary. Understanding what signing changes, and what it does not, is the practical question for providers.

What signing does

Signing creates a presumption in the enforcement posture. For signatories, the Commission focuses enforcement on monitoring adherence to the Code rather than assessing compliance with Articles 53 and 55 from first principles. This offers greater predictability and legal certainty across the EU, and reduced administrative burden.

Signatories benefit from streamlined compliance regardless of their place of establishment or which market surveillance authorities are competent. For a provider operating across multiple member states, this uniformity has real value.

The AI Office also treats commitments made under the Code as mitigating factors when determining fine levels. Adherence does not exclude a fine but can reduce it.

What signing GPAI Code of Practice does not do

Signing does not create a legal safe harbour. Compliance with the Code does not exclude the imposition of fines under Article 101.

Signing does not substitute for compliance with the underlying obligations. Articles 53 and 55 apply regardless. The Code is a route to demonstrating compliance, not a replacement for it.

Signing does not resolve the copyright position. The Code is explicit that adherence does not discharge obligations under Union and national copyright law.

What non-signatories face

Providers that do not sign must demonstrate compliance through other adequate means. In practice this means more requests for information from the AI Office and a requirement to provide more detailed information than signatories would.

There is no penalty for not signing. There is an administrative cost, borne through heavier scrutiny and the absence of the predictability that adherence provides.

Feature	Signatory	Non-signatory
Legal obligation to comply with Articles 53 and 55	Yes	Yes
Route to demonstrating compliance	The Code	Other adequate means, to be justified
Enforcement focus	Adherence monitoring	Compliance assessment from first principles
Information requests	Fewer, more targeted	More frequent, more detailed
Fine mitigation	Commitments taken into account	No equivalent mitigation
Public listing	Yes, listed since July 2026	Not listed

GPAI Code of Practice Timeline

Date	Development
1 August 2024	EU AI Act enters into force
10 July 2025	Final GPAI Code of Practice published
24 July 2025	Training data summary template published
2 August 2025	GPAI obligations under Articles 53 and 55 apply to new models
July 2026	Signatories publicly listed
2 August 2026	Commission enforcement powers enter into application. Article 101 fines available
2 August 2027	Providers of models placed on the market before 2 August 2025 must comply

The period between August 2025 and August 2026 functioned as a good-faith window. The AI Office did not treat signatories that had not fully implemented every commitment as being in breach, working with them instead toward full compliance.

That window has closed. From 2 August 2026, the Commission enforces all obligations for GPAI providers and may impose fines for non-compliance.

Enforcement Powers From August 2026

The AI Office’s enforcement toolkit under Chapter IX of the AI Act includes several instruments that entered into application this month.

Power	Legal basis	Effect
Requests for information	Article 91	Provider must supply documentation and information on request
Model evaluations	Article 92	AI Office may evaluate models, including requesting API or other access
Requests for corrective measures	Article 93	AI Office may require the provider to remedy identified non-compliance
Restriction, withdrawal, recall	Article 93	AI Office may require a model be withdrawn from the EU market
Fines	Article 101	Up to €15 million or 3% of global annual turnover, whichever is higher
Periodic penalty payments	Article 101(2)	Up to 1.5% of average daily worldwide turnover for continued non-compliance

Article 101(1) fines apply where a provider intentionally or negligently fails to comply with a requested corrective measure, fails to provide model access under Article 91, or fails to notify the AI Office of a serious incident under Article 55.Providers retain procedural rights. Article 94 gives providers the right to be heard before the AI Office takes any individual measure against them, with a minimum ten business day response period.

The Signatory Taskforce

Signatories have established a Signatory Taskforce to facilitate coherent application of the Code. The taskforce is chaired by the AI Office.

The function is practical: the Code leaves several questions open, including the criteria for reporting serious incidents and the technical standards for identifying rights reservations. The Taskforce provides a forum for developing consistent interpretation rather than leaving each signatory to reach its own conclusions in isolation.

For providers considering signing, the Taskforce is a meaningful part of the value proposition. It converts an individual compliance exercise into a coordinated one with direct AI Office involvement.

Open Questions and Criticisms around GPAI Code of Practice

GPAI Code of Practice has been criticised on several grounds and the criticisms are worth understanding before signing.

Undefined thresholds. The criteria for what constitutes a serious incident requiring notification are not fully specified. Providers have leeway in how they implement, which cuts both ways: flexibility now, uncertainty about whether an interpretation will be accepted later.

Audit and governance conflicts. The Code contains specific guidance for designing compliance and audit structures. For providers with established governance frameworks, particularly US companies with existing enterprise risk management structures, these can conflict with existing processes and responsibility allocations.

The copyright technical gap. Article 4(3) rights reservation lacks a settled technical standard. Robots.txt is treated as valid but is not comprehensive. Providers attempting good-faith compliance face genuine ambiguity about what a rights holder has effectively reserved.

Extraterritorial reach. GPAI Code of Practice applies to any provider placing a GPAI model on the EU market regardless of establishment. Non-EU providers face obligations designed substantially around EU regulatory concepts, with limited ability to influence their development.

What Providers Should Do Now

If you provide a GPAI model and have not assessed your position

• Determine whether your model meets the Article 3(63) definition.
• Determine whether it exceeds the 10^25 FLOP threshold or has been designated as systemic risk. If it exceeds the threshold, Article 52(1) requires notification to the AI Office before market placement.

If you are not established in the EU

Article 54 requires non-EU GPAI model providers to appoint an EU-established authorised representative by written mandate before placing the model on the EU market. This obligation applied from 2 August 2025. A UK-established entity does not qualify.

If you are considering signing

• Assess the Model Documentation Form against your existing documentation.
• Assess your copyright policy against the Chapter 2 requirements, particularly rights reservation identification and crawling practices.
• If you are a systemic risk provider, assess your Safety and Security Framework against the ten commitments.

Signing commits you at the level of a senior executive with authority to bind the provider. The signatory form is submitted to the AI Office directly.

If you have decided not to sign

• Document the alternative means by which you demonstrate compliance with Articles 53 and 55.
• Expect more frequent and more detailed information requests from the AI Office.
• Ensure your documentation would withstand assessment from first principles rather than against the Code’s structure.

If your model was placed on the market before 2 August 2025

You have until 2 August 2027 under the transitional provisions in Article 111(3). This is not an open-ended deferral. The AI Office retains authority to engage with you during the transitional period, particularly for systemic risk models.

FAQ