The defendant in this case was a non-profit organisation called LAION eV. LAION provides a publicly accessible database with nearly six billion image-text pairs that can be used to train AI systems. One of the images in this database was taken by Robert Kneschke, a photographer who sought a court order to prohibit its use.

LAION had previously obtained the image and metadata itself from another provider, Bigstock, where Robert Kneschke had uploaded his works, and used software to check it for matches with the associated image description for its database. Bigstock had, however, included a restriction in its terms of use, stating that it was not allowed to “download” or “scrape” the images using “automated programs”.

The decision was eagerly awaited, since it involves the highly debated legal question of whether the scientific text and data mining exceptions in Sections 60d and 44b of the German Copyright Act justify using copyrighted works for AI training. The provisions go back to Art. 3 and 4 of the Directive on Copyright in the Digital Single Market (the DSM Directive), and allow the collection and automated analysis of digital or digitised content in order to extract information, in particular about patterns, trends and correlations.

Copyright law privileges text and data mining for scientific purposes

The court agreed with this position and dismissed the lawsuit. It made clear that LAION’s act of reproduction, the downloading of such an image, was covered by the exception of Section 60d of the German Copyright Act.

In the opinion of the Hamburg District Court the comparison of image and description is a privileged analysis for the purpose of obtaining information about correlations (in this case between image content and image description) and, therefore, a privileged scientific purpose. The fact that the data set could later be used to train AI applications did not lead to a different assessment, since the original purpose of data collection was for scientific research.

Obiter dictum on machine-readability of rights reservations

Interestingly, the court then devoted a great deal of attention to a question that was no longer relevant to the case, namely, the exception for text and data mining in Section 44b (3) of the German Copyright Act, according to which reproductions of “lawfully accessible” works are permitted for text and data mining. A work is “lawfully accessible” if it is freely available on the internet or if users are authorised to access it. A restriction of use must be declared in “machine-readable” format in order to be effective.

However, the exact meaning is still disputed. The court took the position that the restriction of use on Bigstock’s website was written in “natural language”. The court hinted that such a disclaimer in natural language in the website’s terms might be effective, and considered them “machine-readable”, since modern technology must also be able to recognise and process licence conditions by machines, thereby making reference to the AI Act (Art. 53 (1) (c) of EU Regulation 2024/1689 for providers of AI models who need to ensure that any asserted restriction of rights is observed).

Takeaways and outlook

The judgment is a significant first step in terms of the copyright permissibility of training AI training, though it does not provide clarity on the question whether an image can generally be used for AI training. LAION was allowed to download the image in order to compare it with the image description for its database due to the text and data mining exception, that is, the use of images for scientific purposes is permissible under certain circumstances.

But the judgment also creates further discussion; the court applied a wide interpretation of the term “scientific purposes”, extending the privilege of “scientific” text and data mining to the data collection, in particular, irrespective of the fact that the AI training data set could later be used to train potentially commercial AI applications. Further, the obiter dictum raises a lot of questions in terms of the machine-readability of rights restrictions and their significance in the context of the development and use of AI.

It remains to be seen whether the Hamburg District Court’s interpretation will prevail in a possible appeal.

Useful links

• The DSM Directive, the European Parliament, 17 April 2019: dycip.com/dsm-directive
• EU Regulation 2024/1689 (AI), the European Parliament, 13 June 2024: dycip.com/eu-regulation-2024-1689