Overview of the EU General-Purpose AI Code of Practice
The General-Purpose AI Code of Practice (GPAI CoP) is a compliance framework released by the European Commission on 10 July 2025. It serves as a voluntary tool for providers of general-purpose AI models to align with the obligations outlined in the European Union Artificial Intelligence Act (AI Act), particularly Articles 53 and 55. While adherence to the Code is not mandatory, providers that follow it benefit from reduced administrative burdens and increased legal certainty, as regulators will presume compliance with the AI Act’s requirements.
The Code is structured into three chapters: Transparency, Copyright, and Safety and Security. The Safety and Security chapter is the most comprehensive, serving as a template for frontier safety frameworks. It includes provisions for risk taxonomy, evaluations, red teaming, incident reporting, and weight security. This chapter applies specifically to models classified as systemic risks, which are defined as those trained with more than 10^25 floating-point operations (FLOPs) or those deemed equivalent in capability or impact under Annex XIII of the AI Act.
The GPAI CoP was developed by 13 independent experts through four thematic working groups: Transparency & Copyright, Risk Assessment for Systemic Risk, Technical Risk Mitigation for Systemic Risk, and Governance Risk Mitigation for Systemic Risk. The process involved contributions from nearly 1,000 stakeholders, including AI developers, academics, civil society organizations, national authorities, and international observers. The Code underwent three iterations before its final publication, reflecting iterative feedback and refinement.
Legal and Regulatory Context
The AI Act, adopted in 2024, established a risk-based regulatory framework for AI systems in the EU. The GPAI CoP was introduced under Article 56 of the AI Act, which authorizes the EU AI Office to develop voluntary guidelines for AI model providers. Articles 53 and 55 of the AI Act impose specific obligations on providers of general-purpose AI models, particularly those with high computational training (above 10^23 FLOPs) or systemic risk potential.
Models trained with more than 10^25 FLOPs are classified as systemic risks and are subject to enhanced safety requirements. The Commission may also designate a model as systemic if it meets equivalent criteria outlined in Annex XIII, even if its compute is below the threshold. The GPAI CoP provides operational clarity on how providers can meet these obligations, addressing gaps in the AI Act’s specificity regarding implementation.
Compliance with the AI Act remains mandatory, while adherence to the GPAI CoP is voluntary. Providers that do not sign the Code must still demonstrate compliance through alternative means. The European Commission has indicated that it may impose stricter actions against non-signatories, emphasizing the importance of alignment with the AI Act’s requirements.
Transparency and Copyright Chapters
The Transparency chapter requires providers to document their models’ capabilities, limitations, and points of contact. Signatories must make key documentation available to downstream providers and publish summaries of training data content. This aligns with Article 53 of the AI Act, which mandates transparency obligations for general-purpose AI models.
The Copyright chapter ensures that providers adopt policies aligning with EU copyright law. This includes mitigating risks of copyright-infringing outputs, such as unauthorized use of copyrighted material in training data or model responses. These chapters apply to all GPAI providers, regardless of model size or risk classification.
Safety and Security Chapter
The Safety and Security chapter is the most detailed and applies exclusively to providers of systemic-risk models. It outlines obligations under Article 55(1) of the AI Act, including conducting model evaluations, assessing and mitigating systemic risks, tracking and reporting serious incidents, and ensuring cyber and physical security of models.
This process includes identifying, analyzing, and mitigating risks until they reach acceptable levels. The chapter specifies four mandatory categories of systemic risks to evaluate: CBRN (chemical, biological, radiological, and nuclear), loss of control, cyber offense, and harmful manipulation. Providers must also identify additional risks to public health, safety, and fundamental rights.
Risk Identification and Evaluation
Providers are required to develop risk scenarios illustrating how identified risks could materialize in real-world conditions. The risk analysis process involves drawing on scientific literature, training data analysis, incident databases, and expert consultations. State-of-the-art model evaluations, such as benchmarking, red teaming, and human uplift studies, must be conducted to target each identified risk.
The Code mandates that providers use predefined, measurable acceptance criteria to determine whether risks are acceptable. These criteria must be established preemptively and cannot be adjusted flexibly ahead of deployment decisions. Only models with risks below acceptable thresholds may be deployed.
Independent External Evaluations and Continuous Risk Management
Appendix 3.5 of the Safety and Security chapter requires providers to engage independent external evaluators for model assessments. Exemptions are possible only if a provider demonstrates that their model is ‘similarly safe’ to another compliant model or if qualified evaluators are unavailable. The determination of ‘similarly safe’ is based on comparable benchmark performance and architectural similarity, typically applicable to models from the same provider or open-source models.
The Code emphasizes continuous risk management throughout a model’s lifecycle, including light-touch evaluations, ongoing mitigation, post-market monitoring, and incident tracking and reporting. Providers must establish organizational governance structures to oversee risk management, promote a ‘healthy risk culture,’ and ensure whistleblower protections for employees raising concerns about systemic risks.
Documentation and Transparency Requirements
Signatories are required to create two types of documentation: a Safety and Security Framework outlining their risk management implementation, and a Model Card summarizing key technical and risk-related information. The framework should detail the processes for risk identification, evaluation, mitigation, and continuous monitoring, while the Model Card provides a concise overview for stakeholders.
Signatories and Global Participation
Among these, xAI, founded by Elon Musk, has signed only the Safety and Security chapter. European AI companies such as Aleph Alpha and Mistral AI have also committed to the Code.
Notable absences include Meta, which has declined to sign, and major Chinese AI companies such as Alibaba, Baidu, and Deepseek. The European Commission maintains an updated list of signatories, reflecting the voluntary nature of the Code while underscoring the expectation that all providers comply with the binding requirements of the AI Act.
The GPAI CoP is intended to evolve alongside other regulatory tools, such as the training data summary template. The EU AI Office plans to update the Code continuously to address emerging risks and technological advancements, ensuring its relevance in the rapidly evolving AI landscape.
Key dates
- 10 Jul 2025Code published by the Commission
- 1 Aug 2025Commission and AI Board assessment of adequacy
- 2 Aug 2025Articles 53 to 55 obligations apply to GPAI providers
- 2 Aug 2026AI Office enforcement powers and fines exercisable