The study sums up a situation that is still evolving rapidly: a single mechanism allowing copyright holders to forbid generative AI from training on their content is yet to be found; but the standardisation of such a mechanism is key if the market for legitimate content licensing is to function properly and fulfil its high potential; thanks to our Carlo Lamantea and Federico Caruso for contributing to this article.
The development of generative artificial intelligence from a copyright perspective, published in May 2025 by the EUIPO (European Union Intellectual Property Office), explores the developments of Generative Artificial Intelligence (GenAI in the following) from the point of view of European Union copyright law.
The paper’s purpose is to understand, from a technical point of view, the relation between GenAI and copyright; its approach is to start from technical, legal and economic analyses of how GenAI works and of what its development entails, and then examine in depth the use and generation of content in the development of GenAI services from the point of view of copyright.
A rapidly evolving picture
The result is a rapidly changing picture with research concluding, inter alia, that no single opt-out reservation mechanism (exception or limitation to rights) has yet been defined to enable copyright holders to exercise the right to prevent GenAI systems from using their online content, as provided by Article 4 of the EU Directive on copyright and related rights in the Digital Single Market (CDSM Directive) .
On the one hand the number of legal disputes concerning copyrights and GenAI is growing, but on the other the study shows that so are high-value agreements between GenAI developers and copyright holders on the use of copyright-protected content for AI training. The study concludes and that this kind of agreement is to play an increasingly important role in generating revenue streams from copyright.
The potential of direct licensing agreements
It is interesting to note that according to the EUIPO’s study the market for licensing agreements for use of copyright-protected content by IA developers actually has a high potential for further development.
It is a market that has been created precisely by the opt-out mechanism put in place by Article 4 of the CDSM Directive mentioned above; this provision indeed makes the use by AI developers of works that have been opted out an infringement of copyrights.
The paper also points out other growth factors for this market, the first being the perception of an imminent shortage of data for machine learning, followed closely by the role of data quality and the importance of metadata and data annotation (easier to read for AI).
Other sources also forecast that licensing agreements will be fundamental on a market heavily impacted by AI. The International Confederation of Societies of Authors and Composers has estimated that by 2028, 24% of music creators’ revenues will be at risk owing to GenAI; by that same date, the musical output of GenAI systems, including both complete and partial works, will be worth 40 billion euro, with an estimated yearly value of 16 billion euro for 2028 alone. Meanwhile, leading music industry players are already engaging in discussions with AI developer startups about music generative models and the possibility of reaching agreements.
Working towards a standard opt-out mechanism
Clearly a well-functioning, standard opt-out system is key to a well-functioning market for legitimate content licensing. Several entities are already working in this direction; the World Wide Web Consortium (W3C) has created a protocol to make declarations reserving exclusive reproduction rights by an automatic mechanism, Chat GPT has an online form that can be filled to reserve one’s rights and other entities (including Getty Images, Tik Tok, Tinder and Giphy) have included a clause in their terms of use and service forbidding the use of their data for machine learning; although it must be noted here that under the CDSM Directive, the reservation must be made available in machine-readable form to be valid.
The Content Authenticity Initiative, founded by Adobe, has improved the C2PA “Content Provenance and Authenticity” protocol, which was originally intended to carry information on the provenance and history of digital media, so that it now also allows creators to attach “do not train” credentials to their content expressing the intention to forbid its use for machine learning.
