Facts
The claimant, photographer Robert Kneschke, sued LAION, a non-profit association whose stated purpose is to promote research and education. LAION makes freely available a dataset of 5.85 billion image-text pairs that can be used to train generative AI. To build it, LAION took image URLs from an existing US dataset, downloaded the images, used software to check whether image and caption matched, and kept only the URLs and captions. In the process it downloaded a preview of the claimant's photograph from a stock agency website whose terms prohibited the use of automated programs or bots. The Hamburg Regional Court dismissed the claim on 27 September 2024.
Question
On appeal the question was whether reproducing the photograph while building the dataset fell within § 44b of the German Copyright Act (UrhG), the general text and data mining exception, and § 60d UrhG, which covers mining for scientific research. Section 44b applies only where the rightholder has not reserved use in machine-readable form, so the court also had to decide whether a prohibition written in natural language in the agency's terms of use met that requirement.
Decision
The court dismissed the appeal. It held that the copying was an automated analysis of digital works to obtain information about patterns, trends and correlations and therefore fell within § 44b. The text-based prohibition in the agency's terms did not satisfy the requirement of a machine-readable reservation and so did not prevent the use. The court also treated LAION as a research organisation pursuing non-commercial purposes and the creation of the dataset as methodical, verifiable scientific research, so § 60d applied as well; the possibility that commercial firms might later use the dataset did not change this. It found the three-step test in Article 5(5) of Directive 2001/29 satisfied and granted leave to appeal on points of law.
Why it matters
This is the first German appellate ruling on the scope of the text and data mining exceptions when datasets are built for AI training. The dispute concerned the preparation of a training dataset rather than the training of a model. It bears directly on the technical form that a rightholder's reservation must take. The Federal Court of Justice's ruling will matter for the interpretation of Article 4 of the EU Digital Single Market Directive and for the EU AI Act's requirement that general-purpose model providers respect such reservations.
Related stages
First instance: Hamburg Regional Court, 27 September 2024, 310 O 227/23 (claim dismissed). Further appeal: Federal Court of Justice (BGH), I ZR 281/25; heard on 3 September 2026. The BGH will announce its decision on 17 December 2026; a reference to the CJEU for a preliminary ruling was raised at the hearing.