I. Background: Dataset Compilation for AI Training
The Hanseatic Higher Regional Court (Hamburg Court of Appeal) has issued a decision on the copyright implications of creating AI training datasets. In the proceeding brought by a photographer against LAION e.V., the Court of Appeal dismissed the plaintiffs‘ appeal and confirmed the first-instance judgment of the Regional Court of Hamburg (310 O 227/23). The dispute centered on the creation of a dataset, which provides links to image-text pairs. These pairs can be used to train generative AI systems. For creating the dataset, LAION downloaded a copyrighted photograph taken by the plaintiff.
The plaintiff claimed that LAION’s downloading and use of the photograph infringed his copyright. LAION, by contrast, argued that its activities fell within the statutory text-and-data-mining (TDM) exception.
II. TDM Exception Applies to Dataset Creation
The Court of Appeal held that LAION’s use of the plaintiff’s photograph was covered by the TDM exception (Article 4 of the DSM Directive and § 44b of the German Copyright Act (UrhG).
A key aspect of the ruling concerns rightsholders‘ ability to opt out of TDM. Such an opt out was declared by the agency commissioned to exploit the plaintiff’s photograph. The court held that, although this opt-out in principle also applies in favor of the plaintiff, the opt-out in the present case was invalid. Contrary to the first instance ruling, which leaned to accept the opt out as valid, the appeal court held that the opt out declared in natural language did not qualify as an opt out in machine-readable form as required by § 44b (3) UrhG.
III. Exception for Scientific Research Purposes Applies
In addition, the Court of Appeal confirmed that LAION can also invoke § 60d UrhG. The creation of the dataset by LAION serves scientific research purposes. The applicability of § 60d UrhG is also not precluded by the fact that the dataset may be used by commercial providers for AI training. This is because these commercial providers do not have a decisive influence on LAION.
IV. Significance and Broader Context
The decision specifies the conditions under which the creation of training data sets consisting of copyright-protected works is permissible.
It should be noted that the ruling of the Higher Regional Court of Hamburg does only concern the admissibility of creating datasets but not the admissibility of training generative AI using these datasets. In contrast, the Munich Regional Court dealt with the latter aspect in its judgment in GEMA v. OpenAI (42 O 14139/24) (see https://grunecker.de/insights/ai-and-copyright-law-munich-regional-court-rules-against-openai/). The Munich Regional Court found that using copyrighted works for training AI models and reproducing these works in outputs can constitute copyright infringement.
V. Conclusion and Outlook
The ruling strengthens the position of organizations compiling datasets for AI training, provided they respect rightsholders‘ opt-out rights. However, the ruling does not address the liability for subsequent model training – a topic that continues to evolve in parallel litigation. In addition, the plaintiff has the option of appealing. In this case, the Federal Court of Justice (BGH) would decide on the case in the final instance.