KI-Flash: Hamburg Regional Court: AI and Copyright - Who Will Win The Battle For Image Rights?

After reporting on the relation between AI systems and the planned Cyber Resilience Act in our last AI Flash, we would like to continue to provide you with legal impulses at regular intervals.

Today's topic: Hamburg Regional Court allows image usage in connection with AI training

On September 27, 2024, the Regional Court of Hamburg issued its first and very interesting judgement by a German court on the use of images for AI training (case number: 310 O 227/23, not yet legally binding). The dispute centered around the download and thus the reproduction of an image by the defendant, an association founded in 2021. The core issue was not directly whether the disputed image data could be used for AI training. 

The defendant offers its database online, free of charge, as a training dataset for AI systems and had included the disputed image in its dataset, which comprises over 5.85 billion image-text pairs. The photographer of the image claimed his rights under Section 16 of the German Copyright Act (GCA) were infringed by the reproduction and subsequent analysis process (comparison with the image description).

The Hamburg Regional Court has now ruled that the defendant association can invoke the exception und Section 60d GCA, which permits reproductions for text and data mining for the purpose of scientific research. 

Firstly, the Regional Court determined that the reproduction was not covered by the exception of Section 44a GCA. According this provision, reproduction is lawful if it is only temporary or incidental to a technical process without an independent purpose. However, the court opined that the reproduction in this case served as a preliminary procurement process for subsequent analysis and was therefore not merely an ancillary act of a technical process.

In contrast, the court found that the disputed use constituted an act of scientific research, defined as the methodical and systematic pursuit of knowledge. Thus, the creation of a dataset is already a fundamental and necessary step towards gaining knowledge at a later stage.

As it was not relevant to the decision, the Regional Court left open whether the act of reproduction would also have been comprised by the exception of Section 44b I GCA, which pertains to the use of text and data mining to obtain information about patterns, trends, and correlations. The court explicitly did not decide whether AI training in general falls under this exception. However, it considered Section 44b I GCA to be applicable on the merits, as the purpose of the image-text analysis was to compare the (non-)correspondence of the images with their descriptions, thus investigating correlations. In a detailed obiter dictum, the Chamber expressed its legal views on the interpretation of the exception:

The Regional Court rejected a teleological reduction of Section 44b I GCA, which would exclude its application to AI training. According to the court, the distinction discussed in the literature between the use of the content of the intellectual creation and the development of information hidden in the database does not sufficiently differentiate traditional text and data mining from use for AI training to justify a teleological reduction of the exception. Moreover, the European legislator has confirmed with the AI Regulation that the training of artificial neural networks is covered by the provision underlying Section 44b I GCA in Article 4 III of the DSM Directive.

However, the court noted that the image in dispute was subject to a possibly validly declared reservation of use, which argued against the application of the exception. Subsequently, the court examined the prevailing opinions on the requirements for machine-readability of the reservation of use within the meaning of Section 40b III 2 GCA. The court rejected the widely held view in the literature that a reservation of use written solely in natural language already meets the requirements of Section 40b III 2 GCA for a machine-readable reservation of use. According to the Chamber, machine comprehensibility must always be assessed based on the current state of technical development and on a case-by-case basis. The Chamber dismissed common arguments against the machine-readability of natural language. For example, the AI-supported review of the source websites does not entail circular reasoning, as a preliminary analysis of the websites does not itself constitute a text and data mining process; rather, crawling only constitutes a fleeting and incidental act of reproduction within the meaning of Section 44a GCA. Furthermore, the concept of machine readability from recital 35 of the PSI Directive, which presupposes simple recognizability, should not automatically be applied to Article 4 III of the DSM Directive.

As a result, the Chamber saw a certain contradiction in values in allowing providers of AI models to develop increasingly powerful text-understanding and text-creating AI models via the barrier in Section 44b II GCA, while not requiring them to use existing AI models within the framework of the counter-exception in Section 44b III 2 GCA. In this respect, the Hamburg Regional Court postponed a landmark decision in this case. However, other courts will likely gratefully accept the almost extravagant obiter dictum of the court on the machine-readability of a reservation of use as a template for future proceedings.