Das Landgericht Hamburg hat im sog. Laion-Fall als erstes deutsches Gericht zu der Frage entschieden, ob durch das Tranieren von künstlicher Intelligenz Urheberrechte verletzt werden. Es hat die Klage eines Fotografen abgewiesen. Die Nutzung seines Fotos sei durch die sog. Text- und Data-Mining-Schranke aus § 60d UrhG gedeckt.
The Decision:
Yesterday, the Hamburg Regional Court ruled on the Laion case as the first German court to address whether AI training infringes copyright. The case attracted much attention because it involves the unresolved legal question of whether the text and data mining exceptions in § 44b UrhG and § 60d UrhG justify using copyrighted works for AI training. The court, at least partially, agrees with this position and dismissed the lawsuit filed by a photographer over the use of his image.
The photographer had sued the non-profit research network Laion. Laion provides a publicly accessible database with nearly 6 billion image-text pairs that can be used to train AI systems. One of the images in this database belonged to the plaintiff, who sought a court order to prohibit its use. In the first instance, the photographer has now lost the case before the Hamburg Regional Court (Ruling from September 27, 2024 – 310 O 227/23). The written reasons for the ruling are not yet available.
It is important to note that the legal dispute is not about whether the image can generally be used for AI training, but whether Laion was allowed to download it to compare it with the image description for its database purposes. Downloading such an image constitutes a copyright-relevant reproduction of a protected work, which requires the permission of the copyright holder. However, the Hamburg court considers this use to be justified by the text and data mining exception in § 60d UrhG. This provision permits the use of copyrighted works for scientific purposes, particularly for text and data mining, without infringing on the copyright holder’s rights. Text and data mining refer to converting unstructured data into structured formats to identify meaningful patterns and generate new insights, a process that relies on vast data collections.
The Hamburg court believes that Laion’s comparison of the image and its description falls under this exception. It views this process as an analysis to identify correlations between image content and its description, which is considered a privileged scientific purpose. The fact that the data was later used for AI training does not change this assessment, as the original purpose of data collection was for scientific research.
Interestingly, the court also touched on the pressing question, in an obiter dictum (a statement not crucial to the ruling), of whether using such data for commercial purposes would be permissible under § 44b UrhG if the copyright holder includes a usage restriction in machine-readable language alongside their work. In this case, the photo agency from which Laion obtained the image had posted such a restriction in “natural language” on its website. The court hinted that such restrictions in natural language might be considered machine-readable if modern AI technologies can comprehend them.
Comparison with Dornis and Stober’s Study:
In September 2024, Tim W. Dornis and Sebastian Stober published a study titled “Copyright and the Training of Generative AI Models – Technological and Legal Foundations.” This study analyzes the copyright framework for training generative AI models and arrives, at least at first glance, at a different conclusion than the Hamburg Regional Court regarding the applicability of the text and data mining exception under § 60d UrhG.
According to the study, § 60d UrhG cannot be applied to the training of generative AI models because such training does not only involve the analysis of purely semantic information but also processes the syntactic information of the data used. Such an extensive use of data exceeds the narrow scope of the legal exception, which is primarily focused on analyzing content for scientific purposes. Additionally, when introducing the text and data mining exception, the legislator was not prepared for the technological development of AI models and their disruptive nature, making its application to AI training improper.
This discussion is certainly far from over, even following the Laion case before the Hamburg Regional Court.
