The GPAI Code of Practice is an industry-over-friendly fig leave, and an invitation to a self-service buffet, serving (non-EU) AI companies.

It could have been a beacon of clarity well balanced between the interests of an industry and protection for Europe, for European authors and translators as well as for the whole book value chain and creative sectors. Instead, the Code of Practice missed out on protecting and supporting the future of European markets and the European creative economies. This Code is an industry-over-friendly fig leave, not even a compromise rather than an invitation to a self-service buffet, serving (non-EU) AI companies. It also normalises exploitation and products manufactured from stealing and imitating.

The European Writers’ Council (EWC), representing 250,000 writers and translators from 53 organisations in 35 countries, urges Member States for improvements or to object the GPAI Code of Practice and the Template for the Public Summary of Training Content for general-purpose AI models. 

Brussels, 25^th July 2025

Find here:

• EWC’s General Position
• GPAI Code of Practice – procedures & insights
• Evaluation: most critical points
• Final remarks and call upon the Member States

EWC Position on the GPAI Code of Practice as pdf

EWC’s General Position

The GPAI Code of Practice is an industry-over-friendly fig leave, and an invitation to a self-service buffet, serving (non-EU) AI companies.

It could have been a beacon of clarity well balanced between the interests of an industry and protection for Europe, for European authors and translators as well as for the whole book value chain and creative sectors. Instead, the Code of Practice missed out on protecting and supporting the future of European markets and the European creative economies. This Code is an industry-over-friendly fig leave, not even a compromise rather than an invitation to a self-service buffet, serving (non-EU) AI companies. It also normalises exploitation and products manufactured from stealing and imitating.

Europe’s culture and creatives are essential to democracy, economics, science, and education. AI impacts everything and is capable of causing tremendous harm here – today and in the future. A clear and strict Code could have prevented this. Examples for efficient regulations exist (i.e. Digital Operational Resilience Act/DORA[1]). DORA sets clear binding regulations and certain measures to be taken up by the Finance Sector in Germany, as a crucial branch for society and economy. The finance industry has to fulfil the rules and improve their standards to prevent themselves and society.

In contrast to DORA, the Code is in favour of only one side: The tech-industry has hardly to take measures, e.g. on transparency, liability and a functioning copyright policy, or invest on measures like unlearning or confirming rights reservations. The Code distributes the risks and long-term costs for damages on Europe and its cultural actors and is harming the interests of authors.

---

GPAI Code of Practice and Template

for the Public Summary of Training Content GPAI models: flawed procedures, influenced by high tech industry, useless for enforcing the rights of authors

Maïa Bensimon, Vice President, Nina George, Political Commissioner, and Nicole Pfister Fetz, Secretary General, took part in the negotiation process for the Code on behalf of the EWC in all four working groups (copyright and transparency, risk assessment, risk mitigation, governance).

Over 1,400 organisations and individuals– whose affiliation has not been made completely transparent – took part in the nine-month process, which covered four drafts to set a Code helping the “industry comply with the AI Act’s rules on general-purpose AI, which will enter into application on 2 August 2025”. The rules become enforceable by the AI Office of the Commission one year later as regards new models and two years later as regards existing models[2].

“The procedure was a pseudo-democratic illusion”, Nina George states and continues:

AI providers, who are to be regulated by the AI Act, have been given inappropriate special treatment in the negotiation of the code – from regular, informal preliminary discussions of their own, to the submission of the final draft even before the member states had a look on it.”

“This is a convenient way to write your own interpretations of the law”, says EWC’s Vice President Maïa Bensimon:  

It is more than urgent that Europe focuses much more on its citizens, among them, authors, and less on foreign companies benefiting from their activities outside of Europe.”

The EWC criticised already since November 2024 frequently the lack of transparency of the procedure and the opaque nature of the individual participants, whose affiliation was not disclosed. Together with the federations of journalists’ associations EFJ and translators’ organisations CEATL, the EWC regularly contacted EU Commissioner Virkkunen, the AI Board and informed the public, to draw attention to the procedure, which was jeopardised by mostly non-European oligopolies.

”Furthermore, the Template presented by the AI Office on 24 July, one week before it came into force, actively prevents the enforcement of rights for individual works, as it does not mandate to any title-specific documentation”, underlines Nicole Pfister Fetz, Secretary General:

Together with the Code, it sends a clear message to European book authors: your copyright, your authors’ rights and your contribution to diversity and democracy are worthless to us. The EWC hopes that the Member States, as the co-legislators, will express a clear signal to the Commission and the responsible EU Commissioner and send both the Code and the Template back to the drawing board, putting an end to the sell-out of European culture to non-European companies. The goal must be to enforce legitimate rights – not to bend the law in favour of a few companies.”

---

Evaluation of the Code and the Template’s effectiveness

In an unnecessary anticipatory obedience, the Code gives away the chance to be a clear, useable and legally binding guide.

Most critical points:

• Useability: The Code is closer to a soft mandate and far from offering clarity.
• Two tiers’ grace period: The Commission’s Q&A confirms that developers who do not comply after signing will still be considered to be acting “in good faith”. In this way, signatories are shielded from penalties for a year, even if non-compliant. Non-signatories are not. This grace period for signatories only runs until 2 August 2026.
• No clarity about text and data mining (TDM): Although the Code does not interpret or change Union’s Copyright law, the relation to Art. 4(1) and 4(3) TDM exception of the CDSM Directive (EU) 2019/790 gives the impression, the Code drafters set text and data mining processes equally to generative AI training. As this is most doubtful, like the several studies and the own-initiative report by MEP Axel Voss show, the CoP falls short in an adequate guidance for AI providers to deal with this ambivalence. Additionally, considerations on Art. 3 of the CDSM Directive (EU) 2019/790 and private partnerships for commercial use were ignored.
• Piracy – using books and text from piracy websites is not banned in full: Developers must exclude infringing websites from training data sources and apply safeguards to stop models from generating outputs that reproduce protected works. But the devil is still in the details: piracy websites shall only be excluded if they operate on a “commercial scale”. Which means in practice, that all those “non commercially” operating sites seem to be “permitted”, which is devastating.
• Gap in Data provenance and documentation over the whole AI development chain: The Code removed one key provision that required developers to assess the legality of third-party datasets, e.g. when they purchase or obtain works from a third party instead of self-scraping. Likewise, the former provision to set up a copyright policy for the whole cycle, from collection, storing, copying, memorising, making publicly available, is gone. This leaves authors with many gaps and hinders to exercise their rights.
• Gap in confirmation of rights reservation: There is no mandatory measure to have opt outs confirmed by AI providers; at the same time, they are “encouraged” to set up voluntary exchange with stakeholders onto a standardisation of rights reservation methods.
• The risks by generative AI on authors’ rights, economy, society or educational purposes are NOT addressed: The sections Copyright and Transparency, towards the Safety and Security Chapter, do not interact and leave gaps of accountability, traceability, and practical and concrete guidelines to assess, mitigate and govern risks related to authors rights, fundamental rights and data privacy. Copyright plays not a single role in risks, mitigation, governance, evaluation.
• No binding definition in the chapters of Safety and Security. The Code leaves it to the industry of AI providers that should be regulated to define what´s a risk and on which scale; what´s appropriate and the adequacy to handle it; what’s the measures for Safety and Security to be.
• Authors’ protection is dependent only on the good will of AI providers: There is no confirmation of rights reservations applied on their works by the genuine rightsholders, the authors. As it says to respect machine-readable rights like robots.txt – BUT this is the only standard mentioned. This is not enough as human readable or otherwise appropriate opt outs, incl. contractual opt outs, are not clearly considered.

On a positive note:

Wording: The Code is in the wording stricter than earlier drafts. Language once referring to “reasonable” or “best efforts” has been replaced with binding commitments.

Prevent Infringing output: AI providers must take effective measures to prevent that AI products infringe the copyright of existing work.

What will – or will not – follow

Vice President Monika Pfundmeier points out:

To have to take these effective measures is a clear attack on the AI industry’s success concept: the imitation and reproduction of lifelong works. Even if it would be the right, moral, honourable and long-term (also economically) beneficial measure to implement by AI providers, we doubt that this provision will be met by the big five companies.”

---

Final remarks

This Code is an industry-over-friendly fig leave, not even a compromise but rather an invitation to a self-service buffet, serving (non-EU) AI-companies the lifelong work and intellectual capital of European authors and cultural professionals. It is useless for most of the interests of authors, performers and other rightsholders, thus harming European cultural and creative economy more instead of protecting us. Together with the useless template, which hinders authors, as the genuine Intellectual Property rightsholder, from enforcing their rights instead of being an instrument for compliance with EU copyright law, the Code and the Template are a shameless sell-out of European cultural and creative works, for non-European companies building their profitable plagiarism and repetition machines.  

We call upon the Member States for improvements or to object the Code and the Template. 

 

 

Nicole Pfister Fetz                    Maïa Bensimon                     Monika Pfundmeier                  Nina George   

Secretary General                       Vice President, France             Vice President, Germany           Commissioner for Political Affairs  

 

EWC Position on the GPAI CoP as pdf

---

Links & Footnotes     

To: European Commission | The General-Purpose AI Code of Practice

[1] https://www.bafin.de/DE/Aufsicht/DORA/DORA_node.html

[2] https://ec.europa.eu/commission/presscorner/detail/en/ip_25_1787