Higher Regional Court Hamburg Confirms AI Training was Permitted (Kneschke v. LAION)

The Higher Regional Court of Hamburg has delivered a landmark judgment in the appeal of Kneschke v LAION. The court dismissed the photographer's appeal, confirming that the creation of the LAION-5B dataset did not infringe copyright.

What was the case about?

Robert Kneschke, a photographer (“Plaintiff””), initially sued LAION, a non-profit association that compiles open-source datasets for AI training. LAION downloaded an image taken by the Plaintiff from a stock photo agency website to analyse it and correlate it with a text description.

The stock agency’s Terms of Service (“ToS”) and the website's source code contained a clause in natural language prohibiting the use of "automated programs" for scraping or downloading. The Plaintiff argued this was a valid "opt-out," meaning LAION could not rely on the commercial text and data mining (“TDM”) exceptions.

The District Court of Hamburg had dismissed the Plaintiff’s claim on 27 September 2024 (see Long-awaited German judgment by the District Court of Hamburg (Kneschke v. LAION) on the text and data mining exception(s)). The Plaintiff then appealed to the Higher Regional Court against the decision of the District Court.

What did the court decide?

The Higher Regional Court upheld the dismissal of the claim.

Crucially, the Higher Regional Court went significantly further than the first instance. The District Court had relied solely on the scientific research exception, under Sec. 60d German Copyright Act (Urheberrechtsgesetz, “UrhG”) transposing Art.  3 RL (EU) 2019/790 (Copyright in the Digital Single Market Directive, „CDSM-Directive“) and indicated a tendency to reject the TDM exception (Sec. 44b UrhG transposing Art. 4 CDSM-Directive), considering the opt-out likely valid. 

By contrast, the Higher Regional Court explicitly held that Sec. 44b UrhG applies, thereby justifying the data processing under the general TDM exception. This was because no valid, machine-readable opt-out had been declared. Specifically, the Higher Regional Court ruled that a reservation of rights expressed in natural language was insufficient to meet the standard of machine-readability during the relevant period of use in 2021.

Furthermore, the court confirmed the District Court’s view that LAION qualifies as a research organisation and acted for non-commercial research purposes, so that they would also benefit from the scientific research exception under Sec. 60d UrhG.

Why is this a significant decision?

Ambiguity remains: Although this is one of the first appellate court decisions in Europe to confirm that the TDM exception under Sec. 44b UrhG (Art. 4, CDSM-Directive) applies to the training of generative AI models, it leaves substantial uncertainty regarding the opt-out mechanism. 

While the Higher District Court rejected the validity of the specific natural language opt-out in this case, it based its decision heavily on the state of technology in 2021. The Higher District Court reasoned that the plaintiff failed to prove that technology capable of reliably interpreting natural language restrictions in ToS or source code, was widely available at the time of the download (late 2021).

At the same time, the Higher District Court implied a dynamic, evolving standard and explicitly adopted a "technology-neutral" approach, suggesting that a natural language reservation could be considered machine-readable today or in the future, provided state-of-the-art technology can reliably interpret it. Consequently, whether a simple text warning constitutes a valid opt-out remains a case-by-case assessment depending on the date of use and the technology available.

What Are the Key Takeaways?

The judgment also provides further guidance on the interplay between generative AI and copyright.

1. Applicability of General TDM for Preparatory Acts (Sec. 44b UrhG)

The Higher District Court confirmed that downloading images to correlate them with text descriptions falls under the TDM exception under Sec. 44b(1) UrhG.

• Applicability to Generative AI: The court clarified that the TDM exception extends to the training of Generative AI. The judges expressly referenced Art. 53 EU AI Act, noting that the EU legislator operates on the assumption that the TDM exception (and consequently also the opt-out mechanism) applies to providers of general-purpose AI models.
• Broad Scope: The Higher District Court rejected the view that TDM exception is limited to extracting "hidden" information from the training data. The automated analysis of correlations between an image and its textual description is sufficient to fall within the definition of TDM (Sec. 44b(1) UrhG).
• Preparatory Acts Included: The Higher District Court explicitly stated that the TDM exceptions cover the preparatory acts (such as compiling datasets) necessary for training AI.

2. The "Opt-Out" Mechanism (Sec. 44b(3) UrhG)

The general TDM exception applies, unless the rights holder effectively reserved their rights (opt-out). The court’s findings here are vital for compliance:

Burden of Proof: The burden of proof is split. The user (AI developer) must generally prove that no opt-out exists. However, if an opt-out exists the rights holder bears the secondary burden of proving that their specific opt-out was legally valid – specifically, that it was "machine-readable" at the time of use.

Opt-Out declared by licensee: The Higher District Court confirmed that an author can rely on an opt-out declared by a licensee (here, the stock agency), regardless of whether they have an exclusive or non-exclusive license. This confirmation was based on the wording of Sec. 44b(3) UrhG, which refers to a reservation of rights by the rights holder rather than the author of the work. The court did not need to address whether and to what extent authors will also be able to rely on opt-outs declared by third parties, who make their content available without a license with the content otherwise being lawfully accessible. 

Attribution & Agency Conflict: Accordingly, the Higher Regional Court held that any opt-outs included in the licensee’s ToS are generally attributable to the author. Importantly, the court noted that in this specific case, the agency's opt-out within the ToS did not contradict the photographer's interests; on the contrary, the plaintiff explicitly relied on it. This leaves the question unanswered how courts might treat cases where an agency's policy explicitly conflicts with the author's intention.

Machine-Readability is Mandatory for Online Content:

• Natural Language vs. Technical Standards: The Higher District Court ruled that a reservation in natural language (e.g., a sentence in Terms of Service) was not sufficient to constitute a "machine-readable" opt-out for online content in 2021 (the time of the download).
• The "Time of Use" Argument: The Higher District Court emphasised that the technology available at the time of the copyright-relevant act is decisive for determining whether the opt-out is declared in a machine-readable manner or not. The plaintiff had argued that modern tools (like ChatGPT) can read natural language ToS. The court rejected this specifically because such technology was not widely established in 2021 when the download occurred.

Dynamic Standard (Technological Neutrality): The Higher District Court adopted a technology-neutral approach, implying a dynamic standard for opting out in a machine-readable manner. Crucially, though, the Higher District Court rejected the plaintiff's respective reasoning solely on procedural grounds (late submission): The plaintiff had argued that a tool to detect natural language opt-outs could have been easily programmed in 2021. The Higher District Court also explicitly left open whether users might be obliged to develop their own software to read opt-outs if standard tools fail. This suggests that if state-of-the-art technology (e.g., LLMs) can reliably interpret natural language today, or if a plaintiff can prove that detection tools are easily programmable, a plain text opt-out might be deemed valid in future cases. Both authors and licensees should consider verifying that licensing agreements sufficiently address potential opt-outs, potentially obliging the licensee to declare opt-outs and the form in which they need to be declared to ensure they are machine-readable.

3. The Three-Step Test

The Higher District Court confirmed that applying the TDM exception passes the "Three-Step Test" (Art. 5(5) Directive 2001/29/EC, “InfoSoc-Directive”).

• Step 1 (Special Case): The Higher District Court found this condition met by the specific statutory scope of Sec. 44b UrhG.
• Step 2 (Normal Exploitation): The Higher District Court ruled that the mere creation of a dataset for internal analysis does not compete with the primary market of the plaintiff (licensing/sales). While the plaintiff argued that future AI models might compete with him, the Higher District Court noted that this abstract future risk does not invalidate the exception for the preparatory training step. The EU legislator explicitly intended to foster AI development.

• Step 3 (Legitimate Interests): Finally, the Higher District Court held that the legitimate interests of the rights holder were not unreasonably prejudiced. In balancing the interests, the court emphasised that the reproduction was technically necessary for the analysis and that LAION acted for non-commercial purposes. Crucially, the Higher District Court argued that the law provides the rights holder with the possibility of an opt-out. Since the plaintiff failed to declare a valid, machine-readable opt-out, his interests were not considered unduly harmed.

   

4. Scientific Research Exception (Sec. 60d UrhG)

The court also upheld the primary defence under the specific scientific research exception, providing detailed guidance to Sec. 60d UrhG (Art. 3 CDSM-Directive), in particular:

• Definition of Research: The Higher District Court clarified that compiling a dataset using novel filtering methods (such as CLIP) is a scientific activity (applied research). It is not merely a technical preparatory step for third parties but a fundamental part of data science.
• Commerciality: The Higher District Court confirmed that LAION acts as a non-commercial research organisation. The fact that commercial entities later use the open-source dataset does not render LAION commercial. Furthermore, the court held that receiving donations or third-party funding does not automatically negate non-commercial status.
• No "Decisive Influence" (Sec. 60d(2) sent. 3 UrhG): The judgment clarifies the high threshold for the "public-private partnership" exclusion. The plaintiff failed to prove that a private company exercised "decisive influence" over LAION.
  • Funding vs. Control: The court found that providing computing power or making a minor payment ($5,000) for data access does not equate to structural control over the research agenda.

  • Personnel: The mere fact that some LAION members are simultaneously employed by major tech companies (e.g., Google, Stability AI) does not automatically create a conflict under Sec. 60d UrhG, provided they do not exercise control over the non-profit in their capacity as corporate representatives.

     

Next Steps

The Higher Regional Court has granted leave to appeal to the Federal Court of Justice (Bundesgerichtshof – BGH). Given the fundamental importance of these questions for the European digital economy, a final clarification in the highest instance (or a referral to the CJEU) is likely.