A German photographer, Robert Kneschke, brought a claim against the organisation LAION regarding the downloading of one of his photographs during the creation of a dataset of image-text pairs usable for training generative artificial intelligence systems. In December 2025, the Hanseatic Higher Regional Court of Hamburg dismissed his appeal. Among the grounds for the judgment is one that shows a significant shift in the protection of copyright: the terms of the agency that marketed the photograph prohibited bots, scraping, and automated downloading, but the claimant did not prove that this reservation could be automatically interpreted in 2021. It was drafted in natural language and was understandable to a person; it was not demonstrated that it could then be located and interpreted automatically as an opt-out instruction.
The German appellate court applied German legislation that transposed a European directive; the ultimate interpretation of that directive would, if necessary, fall to the Court of Justice of the European Union. The decision is not final: Kneschke filed an appeal on points of law before the Bundesgerichtshof, the German Federal Court of Justice, under case number I ZR 281/25. The hearing was held on 3 September 2026, and the ruling is scheduled for 17 December. What exists for now is a relevant appellate judgment, but one limited to a specific reproduction and set of circumstances, not a decision of the Bundesgerichtshof.
The dataset published by LAION does not contain copies of the images, but rather links and descriptions. The reproduction in question was the internal downloading of the photograph to verify that it corresponded to its description.
The European Copyright Directive mandated that Member States introduce exceptions for text and data mining, the technical term for the automated analysis of large volumes of material to extract information. Germany incorporated these into two sections of its law, and the court applied both as independent grounds for its decision. Section 60d covers mining for scientific research purposes: the court deemed LAION to be a research organisation not pursuing commercial objectives and considered the construction of the dataset to constitute applied research. The reservation of rights provided for in Section 44b does not apply to this exception. The latter establishes the general exception and allows the rights holder to exclude such uses via an effective reservation. Kneschke, therefore, did not lose solely due to the form of the reservation: even had he excluded the application of Section 44b, the independent basis of Section 60d would have remained. However, it is the reasoning regarding Section 44b that introduces a novel issue.
For content available online, the reservation under Section 44b is only effective if it is machine-readable. The court clarified the meaning of this requirement: it is insufficient for a machine to merely process the characters of a text; it must be able to locate and interpret it automatically as a reservation applicable to the content, such that the system excludes it from analysis. A notice in natural language does not necessarily satisfy this requirement simply because its characters can be read. In this instance, it was not proven that this was technically feasible in 2021.
The agency's reservation expressed a prohibition on bots, scraping, and automated downloads in natural language. It did not mention data mining; the court interpreted it to encompass LAION's conduct and to be attributable to the photographer. Even so, the court did not consider it effective against Section 44b, as the claimant did not prove that it could be automatically interpreted in the second half of 2021. The examples provided using subsequent tools did not demonstrate what was possible at that time.
A reservation of rights is traditionally directed at another person: it is written, published, and made intelligible to those who must respect it. What this judgment reveals, in a specific context, is that this condition is no longer sufficient on its own. When the reservation must be detected and applied by a program that crawls millions of pages, the declaration no longer counts only for what it says: its effectiveness also depends on whether it can be incorporated into the automated flow that must respect it.
The court was more cautious than any rights holder. It did not state that natural language can never suffice, nor did it establish a mandatory format. It adopted an open criterion: what matters is whether the reservation was machine-readable at the time of use, given the technical possibilities of that moment, and that may change. It also left another question unresolved: whether the party performing the mining is obliged to develop its own tools to detect a reservation when available programs cannot interpret it.
Even with such prudence, the criterion unevenly distributes the possibilities of making a reservation effective. The effectiveness of the reservation may depend on capabilities that many creators do not control: knowing the available mechanisms, being able to incorporate them into the site where the work appears, and preserving proof that they were interpretable at the time of crawling. The problem grows when the author publishes through an agency, a platform, or a social network whose code and terms they do not manage, as was the case here. An organisation with technical resources may be better positioned than an independent creator to implement and document these mechanisms; yet not even the intervention of an agency guarantees that the reservation meets the requirements, as this case demonstrates. The entitlement exists. The position from which to exercise it effectively is not the same for everyone.
Just as the visibility of a work depends on its compatibility with the formats indexed by platforms, its protection against mining is beginning to depend on whether the reservation can be processed by machines. If it cannot, the reservation continues to exist, but it may fail to produce the specific effect of excluding mining covered by Section 44b.
What remains is what the judgment did not have to answer. The subject of the litigation was the downloading of the image to verify the image-text pair during the creation of the dataset. The court took into account that the dataset could be used later to train generative systems, but it did not rule separately on the legality of training a specific model, the potential memorisation of the photograph, or the images generated. In this case, it was not enough for the refusal to be understandable to a person: to exclude the general mining exception, it had to be proven that a machine could also interpret it at the time of crawling. The author's will does not disappear, but its effectiveness becomes dependent on an infrastructure that they do not always know or control.
On the open conversation
This text discusses the judgment of the Hanseatic Higher Regional Court of Hamburg of 10 December 2025 (case 5 U 104/24), which dismissed the appeal by photographer Robert Kneschke and upheld the dismissal of his claim against LAION regarding the downloading of one of his images during the creation of an AI dataset. I do not contend that the ruling legalises AI training in Europe: it is a German appellate court applying the national transposition of a European directive; the decision is not final and is subject to an appeal before the Bundesgerichtshof, the German Federal Court of Justice, which held the hearing for proceedings I ZR 281/25 on 3 September 2026 and has scheduled delivery of its decision for 17 December; furthermore, the litigation concerns the creation of the dataset, not the training of a model or the images it generates. The court further considered the mining exception for scientific research under Section 60d applicable, which is not subject to the reservation mechanism of Section 44b, meaning the ineffectiveness of the opt-out was not the sole basis for the dismissal. A reservation drafted in natural language and comprehensible to a person may not produce the intended effect if it is not proven that a machine could interpret it at the time of use. Anyone wishing to contribute from the perspective of copyright, engineering, or creative practice, will find this notebook open.
Sources
Hanseatisches Oberlandesgericht Hamburg, judgment of 10 December 2025, 5 U 104/24 (anonymised copy of the judgment). Case-law index on Dejure, which records the progression from the Landgericht Hamburg and the appeal before the Bundesgerichtshof (I ZR 281/25).
Official press release from the Hanseatisches Oberlandesgericht, 'KI und Urheberrecht: Hanseatisches Oberlandesgericht weist Berufung zurück', 10 December 2025, indicating that the decision was not final and that an appeal to the Bundesgerichtshof was admitted. Official PDF version.
Bundesgerichtshof, proceedings I ZR 281/25: delivery of the decision scheduled for 17 December 2026 (hearing held on 3 September 2026).
Directive (EU) 2019/790 on copyright in the Digital Single Market, Articles 3 and 4.
Regulation (EU) 2024/1689 on Artificial Intelligence, cited in the judgment as interpretative context.
German Copyright Act: Section 44b UrhG, text and data mining, and Section 60d UrhG, mining for scientific research purposes.
Notes on the ruling
Grounds: the court considered the reproduction justified by two independent exceptions under German copyright law, Section 44b (general mining, Art. 4 of the Directive) and Section 60d (mining for scientific research, Art. 3). It thus upheld the judgment of the Landgericht Hamburg of 27 September 2024 (case 310 O 227/23), which had relied on Section 60d and left the application of Section 44b open.
Reservation of rights: for online content, the reservation under Section 44b is only effective if it is machine-readable, in the sense that it can be automatically interpreted, not merely captured. The agency's natural-language reservation was not proven to be machine-readable in the second half of 2021. The court did not declare natural language invalid in principle and left open the question of whether those performing the mining must develop their own tools to detect reservations.
Scope: the litigation concerned the downloading of the photograph to verify the image-text pair during the creation of the dataset, which is published as links and metadata. The court did not rule separately on the legality of training a specific model, the possible memorisation of the photograph in its parameters or the lawfulness of the generated images.