EU AI Act and General Purpose AI Models: What GPT-4, Llama and Others Must Do

When the EU AI Act was being finalized, one of the most contested additions was the chapter on General Purpose AI (GPAI) models. Foundation models like GPT-4, Llama, Gemini, Claude, and their successors are increasingly the infrastructure layer on which AI applications are built — and the EU legislature decided that this infrastructure layer needed its own governance regime.

The GPAI obligations came into effect on August 2, 2025 — one year after the Act entered into force. This guide explains what GPAI models are, who the obligations apply to, what they require, and what the systemic risk designation means for the most powerful models.

What Is a General Purpose AI Model?

Under the EU AI Act, a general purpose AI model (GPAI model) is an AI model that is trained on large amounts of data at scale, designed for general applicability, and that can competently perform a wide range of distinct tasks. The key characteristics are:

• Trained with large amounts of data using self-supervision, semi-supervision, or reinforcement learning at scale
• Designed for general applicability across a wide range of contexts
• Can be integrated into a variety of downstream systems or applications

This definition captures large language models (LLMs) like GPT-4, Claude, Gemini, and Llama; multimodal foundation models; and similar large-scale general-purpose AI systems. It does not capture AI models trained for specific narrow tasks (such as a model trained solely to classify images of defects in a manufacturing process).

Who Do the GPAI Obligations Apply To?

The GPAI obligations apply to providers of GPAI models — those who develop and make available a GPAI model to the public or integrate it into their own products and services. This means:

• OpenAI (provider of GPT-4, o1, and other models)
• Google DeepMind (provider of Gemini models)
• Anthropic (provider of Claude models)
• Meta (provider of Llama open-weight models)
• Mistral AI (provider of Mistral and Mixtral models)
• Any other company that trains and makes available a model meeting the GPAI definition

Companies that use GPAI models via API to build their own applications are not GPAI model providers — they are downstream deployers, with their own obligations under the high-risk AI system provisions if applicable, but not under the GPAI chapter directly.

Standard GPAI Model Obligations

All providers of GPAI models (not just those with systemic risk) must comply with a baseline set of obligations from August 2, 2025:

Technical Documentation

Providers must draw up and keep up to date technical documentation about the model, including:

• A general description of the GPAI model
• Description of the training process, including compute, data, and architecture
• Information about training data, including sources and filtering methods
• Evaluation results, including safety testing
• Known limitations and foreseeable risks

Copyright Compliance

Providers must implement policies to comply with EU copyright law, particularly the text and data mining exceptions. They must maintain a publicly available summary of the content used for training. This is significant for models trained on large web crawls, where copyright compliance is complex.

Information for Downstream Providers

When a GPAI model is made available for integration into third-party products (via API or open release), the GPAI provider must provide downstream providers with the technical documentation and other information they need to fulfill their own obligations under the Act.

Transparency to Users

When a GPAI model interacts directly with users (in a consumer-facing product), the provider must ensure that users are informed they are interacting with AI.

Systemic Risk GPAI Models: Additional Obligations

The EU AI Act introduces a special category for the most powerful GPAI models: those that pose systemic risk due to their scale and potential impact. A GPAI model is presumed to have systemic risk if:

• It was trained using a cumulative amount of compute greater than 10^25 FLOPs (floating point operations)

This threshold captures the most powerful frontier models — at the time of the Act's passage, this included GPT-4-level and above models. The AI Office can also designate models as having systemic risk based on their actual capabilities even if they fall below the compute threshold.

Providers of systemic risk GPAI models must fulfill significant additional obligations:

Model Evaluation and Adversarial Testing

Providers must perform model evaluations in accordance with standardized protocols, including adversarial testing ("red-teaming") to identify vulnerabilities, risks, and harmful capabilities. Results of these evaluations must be reported to the AI Office.

Incident Reporting

Providers must track, document, and report any serious incidents resulting from the use of their models to the AI Office. This creates an ongoing reporting obligation analogous to data breach notification under GDPR.

Cybersecurity Measures

Providers must implement appropriate cybersecurity protections for their models, including protections against unauthorized access to model weights and training infrastructure.

Energy Efficiency Reporting

Providers must report on the energy consumption of their models, reflecting the EU's concern about the environmental impact of large-scale AI training.

Information Sharing with the AI Office

The AI Office has authority to request additional information from systemic risk GPAI model providers, including access to models for evaluation purposes and access to documentation of internal testing.

The GPAI Code of Practice

The EU AI Act provides for a voluntary Code of Practice for GPAI model providers, developed through a multi-stakeholder process facilitated by the AI Office. The Code provides practical guidance on how to implement GPAI obligations and, critically, adherence to the Code creates a presumption of conformity with the Act's GPAI requirements.

The first Code of Practice was finalized through a multi-stakeholder process in 2025, with input from major AI providers, civil society, academics, and other stakeholders. Major AI providers including OpenAI, Google, Anthropic, and Meta participated in the drafting process.

For GPAI model providers, adherence to the Code of Practice is not mandatory — but it is the most efficient path to demonstrating compliance with the Act's GPAI requirements without facing regulatory scrutiny.

What About Open-Weight Models Like Llama?

The treatment of open-weight models (those where model weights are publicly released) was one of the most debated aspects of the GPAI provisions. The final Act provides a partial exception for providers of open-weight GPAI models: they must still comply with copyright compliance and information documentation requirements, but are exempt from some of the more burdensome technical documentation and downstream provider information requirements.

However, this exception does not apply if the open-weight model poses systemic risk. Meta's Llama models, which as of 2026 approach the systemic risk compute threshold, are subject to close scrutiny. Organizations hosting or fine-tuning open-weight models for others should also consider their own obligations as downstream providers.

Implications for Organizations Using GPAI Models

If your organization uses GPAI models via API to build products, you are a downstream deployer, not a GPAI provider. Your obligations depend on what you build with those models, not on the properties of the underlying model itself. However, several practical implications apply:

• You should verify that the GPAI models you use are compliant with their provider obligations, as this affects your ability to rely on those models in regulated contexts.
• When building high-risk AI systems on top of GPAI models, your technical documentation must describe the GPAI model's role and how you have addressed its known limitations.
• Changes to underlying GPAI models (model updates, version changes, deprecations) may trigger re-assessment of your AI system's risk profile and documentation.

AICerty's ISO 42001 certification process helps organizations navigate the governance implications of building on top of GPAI models, including supplier management policies, change management processes, and documentation requirements for downstream deployers.