A European Study Reveals that AI Training Infringes Copyright

The artificial intelligence feeds on all the data it finds on the internet, without any filter. This is why a European study, conducted by the Initiative, a European copyright initiative that includes 45 member organizations to represent the interests of about 140,000 authors from different fields, has concluded that AI training infringes copyright. 'Copyright and Training in Generative AI – Technological and Legal Foundations' is the study in which a computer scientist and a jurist shed light on the black box of processing steps in AI training, for the first time on this scale.

In spring, the Copyright Initiative commissioned Prof. Dr. Tim W. Dornis (University of Hannover) in collaboration with Prof. Dr. Sebastian Stober (University of Magdeburg) for a joint expert opinion on the technological and legal aspects of training generative AI models. Their interdisciplinary research provides urgently needed new insights into the technically necessary intermediate steps in training generative artificial intelligence. For the first time on this scale, a computer scientist and a jurist jointly create evidence regarding the processing steps in AI training. During the event, many open questions about protected materials were answered in a well-founded, reliable manner and according to the current state of technology.

The work of Professors Dornis and Stober focuses on the copyright evaluation of protected material processing in AI training: "As a closer examination of generative AI model technology reveals, training such models is not a case of text and data mining. It constitutes copyright infringement, without any exception in German and European copyright law," states Professor Stober Dornis. Professor Stober explains that "parts of the training data can be memorized wholly or partially by current generative models – LLM and (latent) diffusion models – and thus can be regenerated with appropriate prompts by end users and thereby reproduced."

Axel Voss, MEP and host of today's event at the European Parliament, expressly thanks scientists Dornis and Stober and welcomes that "the study not only demonstrates that training generative AI models is not covered by text and data mining but also provides other important indications and suggestions to achieve a better balance between protecting human creativity and promoting innovation in AI."

"This study is explosive because it shows that we are facing large-scale intellectual property theft. The ball is now in the politicians' court to draw the necessary conclusions and put an end to this theft at the expense of journalists and other authors," commented Hanna Möllers, legal advisor at DJV and representative of the European Federation of Journalists (EFJ).

For her part, Katharina Uppenbrink, General Director of the Copyright Initiative, emphasizes that the conclusions are "a pioneering result if we now have proof that reproducing works by an AI model constitutes reproduction relevant for copyright purposes and furthermore that its availability on the EU market may infringe public availability rights."

Composer and spokesperson for the Copyright Initiative, Matthias Hornschuh, assures that "there would be a new and profitable licensing market on the horizon, but no remuneration flows while generative AI prepares to replace those whose content it lives off in its own market. This endangers professional knowledge work and cannot be in society's, culture's or economy's interest. All the better that our study authors provide both technological and copyright bases so that legal consideration of generative AI finally becomes a reality."