A study by German authors’ group, Initiative Urheberrecht, argues that AI training falls outside the scope of the EU’s exception for uses of copyrighted material for text and data mining and that distributing such an AI model implicates the separately protected right to “make available” copyrighted works to the public. In other words, they argue that AI producers — which could include researchers — are acting akin to platforms that stream works for public consumption. This study was presented in the European Parliament.
We spoke to Felix Reda, a former Member of the European Parliament (MEP) from Germany during the negotiation of the EU’s Copyright in the Digital Single Market directive on her views on the paper. He provided the following views:
“The authors, who are not known for legal analysis of copyright, contradict the interpretation of the European legislator, who has clarified in the recently adopted AI Act that the exception for commercial text & data mining, and its corresponding opt-out regime, apply to the training of AI models. While the authors are not the first to argue that the European legislator could not have foreseen the rapid advancements of generative AI since adoption of the copyright directive in 2019, supporting AI innovation was an express objective of the European legislator when introducing the provision. It is the very nature of innovation-friendly legal changes to facilitate developments that have not happened yet. That said, generative AI applications did exist and were discussed at the time of the negotiations on the copyright directive, though primarily in the context of automated text generation and voice assistants. The authors’ claim that sharing an AI model is making its training data available to the public flies in the face of the widely voiced complaint of copyright holders that they do not know whether their works were used for AI training. Even Initiative Urheberrecht itself, the lobby group that commissioned the study, advocates for a right for copyright holders to know whether their works were used for training. How can sharing an AI model be considered to be making works available to the public if it does not put others in the position of accessing those works, let alone find out what those works are? It appears that this claim was shoehorned in to find a justification for applying German copyright law to AI models that were trained outside of Germany, or even outside the European Union. The claim that extracting syntactic information would somehow render the TDM exception inapplicable is indefensible given that syntactic information is not protected by copyright to begin with.”
Full article link here.
