Kneschke v LAION: AI and copyright

The Federal Court of Justice of Germany has scheduled the delivery of its decision in the dispute between photographer Robert Kneschke and LAION for 17 December 2026 at 08:45. The hearing took place on 3 September. The court may resolve the appeal or, should it deem an interpretation of Union law necessary, refer preliminary questions to the Court of Justice.

The case will not decide in a general sense whether training an artificial intelligence with protected works is always lawful or unlawful. Nor does it directly concern the generation of images by a model. It examines a specific reproduction carried out in 2021: LAION downloaded a preview of a photograph to verify via software whether its content matched the associated textual description during the creation of a dataset.

The distinction is significant. Framing the litigation as a general trial on 'AI training' artificially expands its scope and obscures the specific legal questions that must be resolved.

The actions of LAION

LAION is a registered non-profit association that provides the public with free access to a dataset comprising 5.85 billion image-text pairs. The collection contains links to images accessible online and associated data, including textual descriptions. It does not contain the images themselves.

To construct the dataset, LAION utilised an existing American dataset containing image URLs and descriptions. It extracted the addresses, downloaded the images, and employed a programme to verify whether each image sufficiently matched its description. It discarded those that failed the verification and incorporated the metadata of the remainder into the new dataset.

During this process, it downloaded a watermarked preview of a photograph by Robert Kneschke hosted on a stock image agency's website. The website's terms of use prohibited the use of automated programmes, applets, bots, or similar means to access the site or its content, including for the purposes of downloading, indexing, scraping, or caching. Kneschke sought to prohibit LAION from reproducing the photograph for the creation of datasets intended for AI training.

The two exceptions for text and data mining

Directive (EU) 2019/790 distinguishes between two regimes for text and data mining.

Article 3 establishes an exception for research organisations and cultural heritage institutions when they perform mining for the purposes of scientific research and meet the relevant requirements. This exception does not depend on the rightholder having formulated a reservation against mining.

Article 4 contains a general exception for reproductions and extractions of lawfully accessible works. In this instance, the rightholder may expressly reserve such use. For content made available to the public online, the directive indicates that the reservation should only be considered appropriate if it is effected by machine-readable means.

Germany transposed these regimes, respectively, into Sections 60d and 44b of its Copyright Act. Section 44b(3) requires a machine-readable reservation for works available online.

The controversy does not merely concern whether a modern computer can process a standard sentence. The question is normative: whether the prohibition drafted in natural language and published in the agency's terms of use met the form of reservation required by law at the time of the download.

The ruling of the court of first instance

The Hamburg Regional Court dismissed the claim on 27 September 2024.

It considered that the downloading and analysis constituted text and data mining. It applied the exception for scientific research under Section 60d: it held that LAION could be classified as a research organisation, that its activity pursued a non-commercial scientific purpose and that the subsequent use of the dataset by companies did not in itself demonstrate a decisive influence by them over the association.

The court also examined the general exception under Section 44b and appeared open to the possibility that a natural language reservation could be machine-readable. However, that consideration was not the necessary reason for dismissing the claim, as the ruling was sustained by the scientific research exception. It is therefore advisable not to present that passage as if it had determined the outcome of the case.

Developments on appeal

The Hamburg Higher Regional Court confirmed the dismissal on 10 December 2025 and authorised the appeal to the Federal Court of Justice.

The appeal considered that the automated comparison between image and description served to obtain information regarding their correlation and, therefore, fell within the legal definition of text and data mining. It deemed the general exception under Section 44b applicable because the rightholder had not proven that, at the time the download occurred, a machine-readable reservation existed. According to the court, the claimant had not proved that the natural-language reservation was machine-readable when the download took place in 2021.

It also considered Section 60d on scientific research to be applicable. The fact that third parties could commercially use the dataset was insufficient to exclude LAION from the exception; the relevant question was whether a company exercised decisive influence over the entity or had preferential access to the results.

The appeal judgment is not final. The appeal is pending before the German Federal Court of Justice under the reference I ZR 281/25.

The decision facing the Federal Court of Justice

The appeal involves at least four related issues:

Whether the automated verification of the correspondence between an image and a description constitutes text and data mining within the meaning of the Directive and German law.
What conditions a reservation must meet to be 'machine-readable' when it appears in a website's terms of use.
Who bears the burden of proving that such a reservation existed and was machine-readable at the time the reproduction was made.
Under what conditions a non-profit entity may rely on the scientific research exception when the resulting dataset may subsequently be utilised by commercial operators.

Not all of them have the same scope. Even if the Federal Court of Justice were to consider the reservation ineffective against the general exception, it would still have to assess the independent path of scientific research. And if it were to confirm the application of Section 60d, the question of the reservation might not decide the dispute on its own.

The limitations of the case

The decision will not automatically determine the legality of all training datasets or all generative models. The facts pertain to a download intended to verify image-text pairs and to a dataset incorporating links and metadata, rather than the downloaded images themselves.

Nor will it decide whether generated output infringes upon a work, whether a model reproduces protected fragments, whether a platform complied with its contracts, or whether an AI-assisted creation possesses a human author. These are distinct issues. The guide on authorship and AI-generated works under Spanish law examines the latter.

The judgment will have direct authority in Germany. Its interpretation may influence other countries, but it will not by itself bind their courts. If the Federal Court of Justice refers a preliminary question and the Court of Justice interprets Articles 3 or 4 of the directive, that interpretation must guide the uniform application of Union law.

The stakes for creators

The practical challenge is genuine. A clause comprehensible to a person may fail to produce the intended legal reservation if the regulation also demands a technically processable format. This is particularly detrimental to those who publish on services that do not permit the modification of headers, metadata, or control files.

However, the contrary conclusion also carries costs. If any sentence included in any terms of use could function as a machine-readable reservation, those conducting large-scale mining would face difficulties in detecting and applying it uniformly. The litigation necessitates precision regarding where the burden of expressing, detecting, and proving the reservation lies.

In my view, this burden should not fall entirely upon individual authors who publish their work online. This assessment does not permit the disregard of the applicable text: for online content, the Directive and German law attribute specific relevance to machine readability.

Pending the decision

The most prudent approach to reserving rights against the general mining exception is to combine a clear declaration for human readers with a machine-readable reservation and to retain proof of when it was published. No technical signal alone prevents downloading or eliminates copies obtained previously.

The guide on how to reserve rights against data mining and AI explains the various available layers and their limitations. A reservation against the general exception does not necessarily block the separate exception for scientific research, which has its own specific requirements.

The article must be updated after 17 December. Until then, what exists is a pending appeal, not a doctrine established by the Federal Court of Justice.

About the open conversation

This text distinguishes between facts, legal interpretation and opinion. It does not constitute legal advice. If you work in intellectual property, develop datasets or publish work online and wish to clarify any details or discuss this reading, this notebook remains open: write to me.

The author's relationship to the cases

There is no financial, professional or collection relationship with the individuals, companies and institutions mentioned.

Kneschke v LAION: The German Federal Court of Justice to rule on AI and copyright