GPAI models code provides guidance but also leaves questions unanswered

Frankfurt-based Dr. Nils Rauer and Amsterdam-based Wouter Seinen, both of Pinsent Masons, were commenting after the GPAI code was published on 10 July 2025 – more than two months after it was expected. Providers that adhere to the code can prove their compliance with the underlying AI Act provisions, but adherence to the code is entirely voluntary and compliance with the legislation can be proven by other means.

The code is split into three separate chapters, addressing security and safety; copyright; and transparency, respectively. The code is meant to help providers navigate the new piece of legislation given that no court has yet to hand down a decision on how the law is to be interpreted nor any authority yet applied the regulatory framework.

Rauer and Seinen said that, in principle, early guidance of this kind is very welcome. However, they said that what is needed to support those who want to apply the new law and to be compliant with it is an adequate level of detail and explanation allowing for business decisions to be based on the code. They said that, at least in regard to the copyright chapter, the code falls behind what businesses had hoped for. 

The copyright chapter addresses requirements arising under Article 53 of the AI Act. This provision requires providers of GPAI models, among other things, to put in place an EU law-compliant copyright policy, publish a “sufficiently detailed summary” about the content used for training of the model, and enable rights holders to reserve their rights not to have their works used for training. However, the code fails to clarify what information the providers will need to share about the copyright material they used in the training of their models, according to Rauer and Seinen.

The Article 53 rules – and other AI Act rules applicable to GPAI models – take effect on 2 August 2025. According to Article 53(1)(d), the summary that providers need to publish must accord to “a template provided by the AI Office”. The anticipated template was not published alongside the new code. Rayer said the said template would deliver the details of what is required to be published in the summary and so is a centrepiece of the copyright chapter. Pinsent Masons has therefore asked the European Commission, within which the AI Office sits, for clarity on when that template will be made available.

Rauer said: “Practically, it will be a challenge for businesses to compile and to make publicly available ‘sufficiently detailed summaries about the content used by the Signatories for the training of their general-purpose AI models’. It will be interesting to see the anticipated template the AI Office will publish in this respect.”

Rauer said the objectives of the copyright chapter of the new code are clear – and highlighted the need for GPAI model providers to put “policies and safeguards” in place to ensure potential copyright issues are addressed “by default and by design”.

Seinen said: “EU copyright laws themselves provide for an exception that allows for GPAI models to be trained using copyright works under certain conditions. Therefore, it is good to read that the code explicitly refers to this exception for text and data mining.”

According to Seinen, other code provisions designed to help GPAI model providers comply with their copyright obligations under the AI Act, are less helpful. He highlighted two measures as examples – first, the suggestion that providers “reproduce and extract only lawfully accessible copyright-protected content when crawling the World Wide Web” and second, that they “identify and comply with rights reservations when crawling the World Wide Web”.

Rauer added: “This guidance sounds obvious but the case Kneschke v LAION currently pending here in Germany before the courts in Hamburg has demonstrated how difficult it can be to decide whether material sitting on a website may be used or not, and if so, for commercial purposes or only for academic use. So, the code does not give definite answers, only food for thought and consideration. Providers should consider the code to be no more than guidance.”

The new code is also silent about how Article 53(1)(c) is to be interpreted. On the face of it, the provision states that any provider placing a GPAI model on the EU market should comply with the reservation of rights expressed by rightsholders in accordance with Article 4(3) of the EU’s 2019 digital single market copyright directive. A non-binding recital in the AI Act states that providers should comply with that obligation “regardless of the jurisdiction in which the copyright-relevant acts underpinning the training of those general-purpose AI models take place”.

Rauer said: “This language clearly suggests Article 53(1)(c) has extraterritorial effect. As a result, it would have been highly appreciated to learn what the AI Office thinks about GPAI models trained, for instance, in the US., since many models out in the market were trained across the Atlantic.”