New Code of Practice for GPAI Models Raises Compliance Questions

On July 10, 2025, a new code of practice for providers of General Purpose AI (GPAI) models was published, offering essential guidance on compliance with the European Union’s AI Act. However, the code has also elicited concerns regarding its clarity, particularly concerning copyright obligations. Experts in the field have highlighted the need for more detailed information to ensure that businesses can effectively navigate this regulatory framework.

The GPAI models code, released by the European Commission, is intended to aid providers in understanding and adhering to the AI Act, which takes full effect on August 2, 2025. The code is organized into three main sections addressing security and safety, copyright, and transparency. According to Dr. Nils Rauer, a partner at Pinsent Masons based in Frankfurt, and Wouter Seinen, head of the Amsterdam office, the guidance is a positive step but lacks the specificity necessary for businesses to make informed decisions. Their comments were made following the code's release, which was delayed by more than two months.

"While early guidance is welcome, the code does not provide the level of detail businesses require to ensure compliance with the new law," said Dr. Rauer. He emphasized that the copyright chapter is particularly vague, failing to clarify what specific information providers must share about the copyright material used in training their models. This is critical, as Article 53 of the AI Act mandates that providers implement an EU-compliant copyright policy, publish a sufficiently detailed summary of the training content, and allow rights holders to reserve their rights regarding their works.

The copyright chapter's ambiguity poses a significant hurdle for GPAI model providers, as the anticipated template that would outline the necessary details is yet to be made available by the AI Office. According to Wouter Seinen, the guidance on copyright is less helpful than desired, noting that the recommendations to reproduce only lawfully accessible content and to comply with rights reservations are overly simplistic. "Such guidance lacks the nuance required for businesses to navigate the complexities of copyright law," he stated.

Moreover, the code does not address how the extraterritorial application of Article 53(1)(c) should be interpreted, which could affect providers whose models are trained outside the EU. Dr. Rauer pointed out, "The lack of clarity regarding the application of copyright laws across jurisdictions is concerning, especially for models trained in the United States."

As the industry prepares for the implementation of the AI Act, the code's voluntary nature means that adherence is not mandatory for compliance. Providers may prove compliance through alternative methods, which raises questions about the effectiveness of the code in standardizing practices across the industry.

The implications of this code extend beyond mere compliance; they highlight the ongoing challenges in aligning technological advancements with existing legal frameworks. As businesses grapple with these new regulations, the need for clearer guidelines and supportive tools becomes increasingly apparent. The AI sector is at a critical juncture, and how it addresses these compliance issues will significantly impact its future trajectory.

In conclusion, while the GPAI models code of practice marks a significant step towards regulatory clarity, its limitations underscore the complexities inherent in aligning AI development with copyright and compliance obligations. Stakeholders are urged to seek further clarification from the European Commission to ensure that the forthcoming guidelines meet the needs of the industry and adequately address the concerns raised by experts. The future of GPAI model compliance will depend on a collaborative approach between the regulatory bodies and industry providers to foster an environment of innovation while protecting intellectual property rights.