The Higher Regional Court of Hamburg has issued a ruling that contains important guidelines for the admissibility of AI training and data mining.

In its ruling dated 10 December 2025, the court dismissed the appeal brought by the photographer Robert Kneschke against the first-instance judgment of the Regional Court of Hamburg dated 27 September 2024 (Ref.: 310 O 227/23). The claimant had already been unsuccessful at first instance with his action for injunctive relief against the use of one of his photographs for AI training purposes by the non-profit association LAION e.V. The judgment of the Regional Court had already attracted considerable attention as the first German decision to address the scope of the statutory opt-out for text and data mining in the context of AI training.

The appellate decision generally confirms the first-instance judgment. However, the judges take a different view on two material aspects: First, court expressly stated that the reproduction of the photograph in question is covered by the exemption for text and data mining, whereas the Regional Court left this question open. And second, contrary to the lower court, the Higher Regional Court of Hamburg found that the opt-out, formulated in natural language did not satisfy the statutory requirement of machine-readability pursuant to Section 44b para. 3 of the German Copyright Act and Article 4 of the DSM Directive (EU) 2019/790.

Facts of the case

The subject matter of the proceedings was a photograph to which a hyperlink was included in a dataset comprising 5.85 billion image-text pairs, compiled and made publicly available free of charge by the defendant, a non-profit association. The dataset included hyperlinks to images freely accessible on the internet as well as descriptions of these images. In the course of creating the dataset, the defendant extracted image URLs, downloaded the corresponding images, and analyzed the correspondence between text and image. The photograph in question was downloaded by the defendant from the website of a stock photo agency to which the photographer had licensed the image.

Since at least 13 January 2021, the following usage restriction was posted in English on the website of the stock photo agency:

RESTRICTIONS

YOU MAY NOT:

(…) 18. Use automated programs, applets, bots or the like to access the website or any content thereon for any purpose, including, by way of example only, downloading Content, indexing, scraping or caching any content on the website.“

Key Points of the First-Instance Judgment

At first instance, the Hamburg Regional Court dismissed the photographer’s claim, holding that the reproduction, namely, the downloading and processing of the disputed image, was covered by the scientific research exception set out in Section 60d of the German Copyright Act that is based on Article 3 of Directive (EU) 2019/790. In the court’s view, the mere creation of a dataset serving as the basis for training AI systems may be qualified as scientific research. The court considered the research purpose to be fulfilled, as the defendant made the dataset expressly available free of charge to scientists. It was deemed irrelevant by the court that the datasets could also be utilised by commercial entities for AI training purposes, and that LAION e.V., as the data collector, did not itself pursue independent research objectives.

Furthermore, the court, in an obiter dictum, commented on the possible applicability of the text and data mining exception of Section 44b German Copyright Act based on Article 4 of the DSM Directive (EU) 2019/790 to reproductions in connection with AI training, but expressly left this question open. With regard to the text and data mining exception, the court clarified that a usage reservation formulated in “natural language” is generally sufficient for effectiveness. What is decisive is the mere “machine interpretability” of the opt-out, which the court considered adequate.

Key Points of the Appellate Decision

In the view of the Higher Regional Court of Hamburg, the reproduction of the photograph (i.e., the downloading of the image) is covered by the statutory exception set out in Section 44b para 2 of the German Copyright Act. Unlike the first-instance judgment, which left this question open, the Higher Regional Court of Hamburg held that the download in question constitutes a reproduction for the purposes of text and data mining. The download was carried out in order to compare the image content, using software, with the image description already stored. This comparison constitutes an analysis for the purpose of obtaining information within the meaning of the text and data mining exception. According to the court, this follows from the wording of the provision, which presupposes the automated analysis of one or more digital or digitized works in order to obtain information about patterns, trends, and correlations.

The court expressly limited its reasoning to the applicability of the text and data mining exception to the defendant’s preparatory measures prior to the actual AI training, namely the comparison between image and image description. The court considered it unnecessary, for the purposes of the present case, to address the controversial question as to whether subsequent training of generative AI models meets the requirements of Section 44b para. 1 of the German Copyright Act.

The observations of the Higher Regional Court of Hamburg regarding the form of the opt-out merit particular attention. In the court’s view, the opt-out declared in natural language on the website of a stock photo agency did not meet the form required by the text and data mining exception. The provision requires that, for works accessible online, an opt-out is only effective if it is declared in machine-readable form. Whether a reservation pre-formulated in natural language (without the use of a protocol such as a robots.txt file) meets the requirements of Section 44b para 3 sentence 2 of the German Copyright Act is highly controversial. According to the Higher Regional Court of Hamburg, what matters is that the reservation “can be interpreted by machines”; mere machine detectability is not sufficient. The Higher Regional Court of Hamburg relied primarily on the German legislative rationale, according to which the provision is intended to ensure that, for online content, “automated processes, which are a typical feature of text and data mining, can actually be carried out automatically” (BT-Drs. 19/27426, p. 89). In the court’s view, the claimant did not demonstrate that the usage reservation was machine-readable in the second half of 2021.

Finally, the Higher Regional Court of Hamburg, in line with the first-instance decision, found that the use of the photograph was also justified by scientific research exception, as it was for the purposes of scientific research. According to the court, it is sufficient for scientific research that the dataset in question is aimed at future knowledge gain. The court considered the non-profit association LAION to be an “other institution” within the meaning of Section 60d para. 2 of the German Copyright Act, which conducts scientific research and is thus within the group of beneficiaries under the provision. Furthermore, the defendant did not pursue commercial purposes within the meaning of Section 60d para 2 sentence 2 no. 1 of the German Copyright Act, as it is a non-profit organization according to its statutes and makes its results available free of charge.

Outlook

The judgment makes clear that the legal questions surrounding text and data mining are far from settled. In particular, the divergent views on the machine-readability of an opt-out show that this issue remains highly controversial in both legal literature and case law. As long as no uniform standards have been developed regarding machine-readability, there will continue to be significant legal uncertainties regarding the analysis of copyright-protected content for AI training purposes.

It should also be noted that the Higher Regional Court of Hamburg confirmed the broad interpretation of the term “scientific research” under Section 60d German Copyright Act and Article 3 of Directive (EU) 2019/790 adopted by the Hamburg Regional Court, thereby further strengthening the position of research institutions.

The question of the extent to which the text and data mining exception applies to all stages of generative AI training remains open, as the Higher Regional Court of Hamburg limited its comments exclusively to the comparison between image and image description prior to the actual training of the AI.

Although the decision further clarifies the legal guidelines for AI training using pre-existing works, the legal dispute of Robert Kneschke v. LAION is not yet concluded. The Higher Regional Court of Hamburg allowed a further appeal to the Federal Court of Justice, so the claimant still has the opportunity to challenge the judgment. The decision therefore represents only a further step towards greater legal clarity regarding text and data mining for AI training purposes.

Resources

You can view our latest campaign – Algorithm to Advantage which gives you actionable insights you can turn into your competitive edge.