Kneschke v. LAION: Europe's First Appellate Ruling on AI Training Data and Copyright

• The Hamburg Higher Regional Court's December 2025 decision marks Europe's first appellate ruling on text and data mining (TDM) exceptions for AI training datasets, establishing crucial precedent for how copyright law applies to machine learning development across the EU.

• German courts confirmed that non-commercial research organizations can lawfully compile AI training datasets from publicly available content under TDM exceptions, validating the open-source AI development model.

• Natural language terms of service prohibiting web scraping do not constitute valid "machine-readable" opt-outs under EU copyright law, requiring rights holders to implement technical measures like robots.txt or metadata protocols.

• The ruling leaves critical questions unresolved—including whether actual AI model training and AI-generated outputs infringe copyright—setting the stage for potential Federal Court of Justice review and possible CJEU referral.

When AI Meets Copyright: The LAION Case Reshapes European Data Mining Law

In what has become the defining copyright case for generative AI in Europe, German stock photographer Robert Kneschke challenged LAION e.V., a non-profit organization that created one of the world's largest AI training datasets. The LAION-5B dataset contains approximately 5.85 billion image-text pairs and has served as foundational training material for prominent AI image generators including Stable Diffusion and Imagen. Kneschke discovered that one of his watermarked preview images had been downloaded as part of LAION's dataset creation process. While LAION's final dataset consists only of URLs and metadata rather than the images themselves, the organization did temporarily download and analyze images to verify that text descriptions matched visual content. This verification step, essentially web scraping to filter and curate the dataset, became the central point of contention.

The photographer filed suit in April 2023, and the Hamburg Regional Court dismissed his claims in September 2024, finding that LAION's activities fell within the scientific research TDM exception under Section 60d of the German Copyright Act. The court determined that creating a dataset constitutes a fundamental preparatory step aimed at enabling future knowledge acquisition, and that LAION's non-commercial, research-oriented purpose was not negated by the fact that commercial AI companies later used the dataset.

The Hanseatic Higher Regional Court of Hamburg issued its appellate ruling on December 10, 2025, dismissing Kneschke's appeal and going significantly further than the first-instance court. VG Bild-Kunst, the German collecting society representing approximately 70,000 visual artists, had financed the appeal, underscoring the case's significance for the creative industry. The appellate court confirmed that LAION's activities qualified under both TDM exceptions, Section 60d for scientific research and Section 44b for general commercial TDM, providing substantially greater protection than the district court's single-exception approach. This dual affirmation gives LAION a robust legal foundation that dataset creators across Europe will look to as persuasive guidance.

Most significantly, the Higher Regional Court squarely rejected the lower court's controversial suggestion that natural language opt-outs could be effective. It held that the stock website's textual prohibition did not meet the statutory requirement of a machine-readable opt-out in the context of LAION's 2021 data collection. The panel emphasized that machine-readability is a technical standard: an opt-out must be in a form that computer systems can reliably identify and process automatically. General website terms or textual footers, even if clear in human language, do not inherently signal to web crawlers in a standardized way and therefore fail the opt-out test absent further technical measures.

The court adopted a technology-neutral approach with temporal implications that will shape compliance strategies for years to come. It acknowledged that if and when state-of-the-art crawling technology improves to reliably understand natural-language restrictions, a plain text opt-out might be considered machine-readable in the future. However, the burden was on Kneschke to prove that by late 2021 there were readily available tools to automatically detect such clauses, and he failed to establish this. The implication is that for the time being, only explicit technical measures count as opt-outs, but rights holders and AI companies should monitor evolving norms closely.

The appellate court strongly affirmed what might be called the third-party use doctrine: that commercial entities use LAION's open-source dataset for profit is irrelevant to assessing LAION's non-commercial purpose. The open-source model inherently makes results available to everyone, including commercial actors, and this does not transform non-commercial research into commercial activity. The court found no evidence that any private company exerted decisive influence or control over LAION's work, and importantly, since LAION made all results freely available to everyone, no preferential access existed that would disqualify the organization from the research exception.

The Higher Regional Court also examined the three-step test under the InfoSoc and DSM Directives to ensure that applying these exceptions did not unduly prejudice the photographer's rights. It found no conflict with normal exploitation of Kneschke's work: LAION's internal download for verification does not compete with or substitute for the licensing market for that photograph. The court specifically noted that at the point of use, LAION was not disseminating the photo or usurping any direct economic opportunity. Any downstream impact of AI outputs on the stock photo market was deemed too abstract and speculative to consider at this stage, and the strong public interest in enabling AI research and innovation supported the application of the exceptions.

Both courts carefully confined their holdings to the act of dataset creation—copying for TDM purposes. The judgments explicitly do not resolve whether using the dataset to actually train an AI model, or the model's reproduction of works in AI-generated output, would be permissible. This limitation is particularly significant given contrasting developments elsewhere. The Munich Regional Court in GEMA v. OpenAI reached dramatically different conclusions in November 2024, finding that ChatGPT memorizes song lyrics embedded in model weights and ruling that TDM exceptions do not cover permanent memorization or reproduction in outputs. The UK High Court in Getty Images v. Stability AI took yet another approach in November 2025, rejecting claims that model weights store protected works—directly contradicting the Munich court's findings on memorization and creating jurisdictional fragmentation that may ultimately require CJEU clarification.

The OLG Hamburg expressly granted leave to appeal to the Bundesgerichtshof, stating the matter has fundamental importance for legal development. Kneschke may seek Federal Court of Justice review, which could eventually lead to a CJEU referral on questions including the uniform interpretation of machine-readable opt-outs across the EU. VG Bild-Kunst has strongly criticized the ruling, calling it a template for circumventing Article 44b and urging political action during the DSM Directive's upcoming evaluation. Key questions remain unresolved that only the BGH or CJEU can clarify: what exactly constitutes "machine-readable" format given evolving technology, who bears the burden of proof for TDM opt-outs, and how courts should assess impacts of generative AI outputs on normal exploitation of original works.

Our Perspective

The Kneschke v. LAION decisions carry profound implications for businesses operating at the intersection of AI development and content creation. For organizations building or utilizing AI training datasets, the ruling provides welcome legal clarity—but with important operational caveats.

Commercial AI developers should recognize that while the German courts endorsed TDM applicability to AI training, the machine-readable opt-out requirement creates significant compliance burdens. Article 53(1)(c) of the EU AI Act, effective August 2025, requires general-purpose AI providers to implement policies respecting TDM opt-outs using state-of-the-art technologies. The Hamburg court's citation of this provision suggests that natural language processing may now be required for opt-out detection. Companies training foundation models face potential penalties up to 3% of global turnover or €15 million for non-compliance.

For content platforms and rights holders, the message is unambiguous: natural language terms of use provide insufficient protection. Effective opt-outs require technical implementation through robots.txt directives, TDM Reservation Protocol metadata, or similar machine-readable mechanisms. Organizations managing significant content libraries should audit their current opt-out implementations and upgrade to standardized technical measures. The time-of-use standard established by the appellate court means compliance assessment may shift as technology evolves, what was sufficient in 2021 may be inadequate in 2026.

The case also highlights an emerging business opportunity in the AI compliance space. As the technology-neutral, time-sensitive interpretation of machine-readable creates ongoing uncertainty, demand will grow for tools and services that help both rights holders implement effective opt-outs and AI developers detect and honor them. The development of common opt-out vocabularies and interoperable standards represents a nascent but potentially valuable market segment.

Research institutions and non-profit AI initiatives gain the most direct benefit from this ruling. The decisive influence test sets a high bar for disqualification—isolated business relationships with commercial entities do not transform non-profit research into commercial activity. This preserves the open-source AI development model while acknowledging that commercial downstream use is an inherent feature, not a bug, of open research.

Key Takeaways

• Europe's first appellate court has ruled that text and data mining exceptions under the DSM Directive cover AI training dataset creation when conducted for non-commercial scientific research purposes.

• Natural language terms of service prohibiting web scraping do not constitute valid "machine-readable" opt-outs under current EU copyright law, at least as assessed against 2021 technology standards.

• The "machine-readable" requirement is interpreted dynamically—what was insufficient in 2021 may become adequate as technology evolves, creating ongoing compliance uncertainty.

• Non-profit research organizations can make datasets freely available to commercial third parties without losing their non-commercial status, validating the open-source AI development model.

• Rights holders must implement technical opt-out measures such as robots.txt directives or TDM Reservation Protocol metadata to effectively reserve their works from data mining.

• The rulings address only dataset creation (the input stage), leaving actual AI model training and output generation for future litigation or legislation.

• Contradictory rulings across German courts (Hamburg vs. Munich) and between jurisdictions (Germany vs. UK) create legal fragmentation requiring eventual harmonization.

• The EU AI Act's August 2025 requirements for respecting opt-outs "through state-of-the-art technologies" may raise the compliance bar for commercial AI developers.

• The case may proceed to Germany's Federal Court of Justice and potentially to the CJEU for definitive EU-wide interpretation of key concepts.

• Businesses should prepare for evolving standards by investing in both technical opt-out implementations (for rights holders) and detection capabilities (for AI developers).

References

• Hamburg Regional Court (Landgericht Hamburg), Case No. 310 O 227/23, Judgment of September 27, 2024

• Hanseatic Higher Regional Court of Hamburg (OLG Hamburg), Case No. 5 U 104/24, Judgment of December 10, 2025

• Kluwer Copyright Blog, "Kneschke vs. LAION - Landmark Ruling on TDM Exceptions for AI Training Data," Parts 1 & 2

• Grünecker, "AI and Copyright: Hamburg Court of Appeal Rejects Appeal in LAION Case," December 2025

• Norton Rose Fulbright, "Machine-Readable Opt-Outs and AI Training: Hamburg Court Clarifies Copyright Exceptions," December 2025

• German Copyright Act (Urheberrechtsgesetz) 

• EU Directive 2019/790 on Copyright in the Digital Single Market (DSM Directive)

• EU AI Act, Article 53(1)(c)

Keywords: AI training data copyright, LAION dataset, text and data mining exception, TDM opt-out, EU DSM Directive, machine-readable copyright reservation, open-source AI datasets, German copyright law