The European Union’s AI Office set up a Working Group to produce a voluntary General Purpose AI (GPAI) Code of Practice (CoP) that AI model providers could sign, to demonstrate compliance with EU obligations. Over time, rightsholders and copyright advocacy groups have provided ongoing input to its drafts. Its third draft was released on March 11, 2025.
In response to this third draft iteration, a broad coalition of creators and rightholders’ organizations has spoken out strongly against it in a joint statement. The statement, released by IMPALA, the Independent Music Companies Association, said “The third draft of the GPAI Code of Practice undermines the objectives of the AI Act, contravenes EU law and ignores the intention of the EU legislator – we cannot support it. 38 organizations are recognized in the IMPALA press release, including AAPA, the Audiovisual Anti-Piracy Alliance.
The statement was direct and to the point: “Regrettably, the third draft does not meet the adequacy requirement provided by the EU AI Act, and, therefore, should not be approved without substantial improvements. No Code would be better than the fundamentally flawed third draft.”
Did the EU Working Group miss the point?
IMPALA’s statement contends that (the third draft of the EU CoP) creates legal uncertainty,…” and “… sets the bar so low as to provide no meaningful assistance for authors, performers and other rightsholders to exercise or enforce their rights. Even more fundamentally, it would also not ensure that GPAI providers comply with either EU copyright law or the AI Act.
(IMPALA) “participated constructively in the drafting process and provided substantive comments to the previous drafts of the Code of Practice, underlining severe deficiencies affecting all creative sectors. However, the drafters of the Code have largely ignored or dismissed those comments…”
Swedish piracy org weighs in
The Swedish Rights Alliance found that copyright measures were further diluted in the third draft, compared to previous drafts. The requirements for AI companies were lowered, and the ability to submit complaints is restricted.
Particularly concerning to them was Measure I.2.2, which allows AI services to scrape all pirate sites not recognized by courts or authorities as pirate services. Given the number of pirate services available and the length of court proceedings, AI companies could essentially gain access to all books, video games, music, movies, TV series, and sports broadcasts to train their data on without compensating copyright holders, they said.
Flaws identified
- The IMPALA statement identifies how the third draft (does not) ensure compliance with EU copyright rules and the AI Act itself:
- The draft’s statement that GPAI providers merely need to make “reasonable efforts” is ambiguous.
- The draft does not ensure meaningful due diligence obligations to comply with EU copyright law and the AI Act (and) would also risk guiding GPAI providers towards copyright infringements.
- It fails to provide meaningful guidance on the right of authors, performers and other right holders to choose how they reserve their rights, and toward what GPAI providers must do to comply with such reservations. It also “misinterprets EU copyright law, but also does not reflect the reality of how pirated content is accessed and distributed online.”
- The third draft removes the requirements to GPAI providers, to disclose whether or how they comply with the rights reservations expressed by authors, performers and other rightsholders.
The IMPALA statement notes that ROBOTS.TXT is considered to be the only method GPAI providers must recognize and respond to, while other reservation methods are ignored or treated as optional, in direct contradiction with EU law.
Research says AI modelers use pirated content extensively
A report released in March (2025) by the Rights Alliance identified the illegal content libraries being used to train AI models produced by commercial model suppliers that include Books3, Common Crawl, LibGen and Redpajama, and by generative AI platforms that are used directly by end-users, including from Apple, Meta Platforms, Microsoft, OpenAI and DeepSeek.
Special recognition might go to Meta Platforms, parent company of Facebook, Instagram and other social media platforms, which has suffered multiple accusations, from book authors (in the Kadrey case), by copyright advocates including the Danish Rights Alliance and CreativeFuture; and by others. In March 2025, French publishing and author organizations SNE, SNAC and SGDL sued Meta for its alleged use of their copyrighted content to train Meta’s AI models without permission.
And yet, Meta may actually be one of the least egregious offenders: A 2024 report by Stanford University compared and ranked ten AI models for transparency – willingness to reveal their sources – under the premise that copyrighted works are used regularly to train them and that the model providers might have something to hide. The model with the highest rank was Meta’s LLaMA2, at 54%; the lowest being Amazon’s Titan Text at just 12%.
Further reading
Third Draft of the General-Purpose AI Code of Practice published, written by independent experts. Press release with links to downloads. March 11, 2025. European Commission
Joint Statement by a coalition of authors, performers and other rights holders active across the EU’s cultural and creative sectors regarding the third draft of the EU AI Act’s GPAI Code of Practice. Press release. March 28, 2025. IMPALA (Independend Music Companies Association)
Code of practice for general-purpose AI, third draft. Press release. March 13, 2025. Rattighetsalliansen (Swedish Rights Alliance)
The Rights Alliance cannot support the proposed implementation of the EU’s AI regulation. Press release. March 28, 2025. Rettighedsalliancen (the Danish Rights Alliance)
Report: Classic pirate sources are widely used to train AI datasets, says Danish Rights Alliance. Article. March 20, 2025. Piracy Monitor
Why it matters
In its own statement, the Danish Rights Alliance noted that the “effective exercise and enforcement of copyright in generative artificial intelligence requires transparency about the works and creations on which AI models are trained. The EU’s AI Regulation contains provisions aimed at ensuring the effective exercise of rights, by obliging providers of AI models to implement policies for compliance with copyright and related rights; and to prepare and publish a sufficiently detailed description of the models’ training data.
“To this end, EU regulators drafted and iterated a General Purpose AI (GPAI) Code of Practice (CoP), which is intended to provide proper measures to facilitate and persuade the providers of General Purpose AI models to respect the two basic principles of copyright law: to seek prior authorization and abstain from unauthorised uses of copyrighted material, and to apply whenever a GPAI provider places their model into the EU market, regardless of where they are established or where the training of that model took place,” said the Rights Alliance.
