Machine-Readable opt-outs and AI training: Hamburg Court clarifies copyright exceptions

A recent decision by the German Hanseatic Higher Regional Court (OLG Hamburg, 5 U 104/24, 10 December 2025) in the case of Robert Kneschke v. LAION e.V. provides the first appellate-level guidance in Germany on questions surrounding copyright exceptions for text and data mining (TDM) and scientific research in the context of AI training.

The ruling has significant implications for rights holders, AI developers and research organisations.

Background

In 2024, in the case of Robert Kneschke v. LAION e.V. we examined the open questions surrounding Germany’s copyright exceptions for text and data mining and scientific research in the context of AI training (see our blog, Germany: landmark court decision deals with AI training and copyright).

Robert Kneschke, a seasoned German photographer, initiated legal proceedings against LAION e.V. (“Large-scale Artificial Intelligence Open Network”), a non-profit organisation known for creating vast datasets used in AI training.

Kneschke claimed that LAION included his photographic images in their "LAION 5B" dataset without his consent. This dataset has been instrumental in training popular AI models like Stable Diffusion. The lawsuit challenged the legality of using copyright works in this manner and demanded the removal of his images from the dataset.

The German Copyright Act (UrhG), updated by the EU Directive on Copyright in the Digital Single Market in 2021, introduced provisions for text and data mining, allowing such activities for commercial purposes under specific conditions. LAION argued that their use of metadata, text data and URLs fell within these exceptions.

At first instance, the Hamburg Regional Court ruled in favour of LAION, finding that the use by LAION of Kneschke’s photo for AI training benefited from the exception to copyright infringement for text and data mining for the purpose of scientific research under section 60d UrhG (implementing Article 3 of the EU Copyright Directive) – for more information about the first instance decision, see our blog, Testing the boundaries: AI training models and copyright laws.

On appeal by Kneschke to the Hanseatic Higher Regional Court (OLG Hamburg, 5 U 104/24, 10 December 2025), Kneschke failed in his bid to overturn the lower court decision, the Hanseatic Higher Regional Court ruling that LAION could rely on copyright exceptions for text and data mining and that Kneschke’s TDM opt-out was invalid because it was not machine-readable.  In doing so, the Court provided various guidance on such issues. 

What is  protected from infringement claims – and what is not?

This decision draws a clear line:

✅ Protected under sections 44b and 60d UrhG

• Temporary reproductions for automated analysis (e.g., image-text matching, filtering, validation).
• Creation of datasets for research purposes, including applied research and development of AI models.
• Internal processing steps that do not result in public distribution of the protected work itself.

❌ Not protected

• Any use beyond TDM, such as publishing or redistributing the actual copyrighted content.
• Outputs generated by AI that reproduce protected works or infringe adaptation rights.
• Commercial exploitation by entities that do not meet the criteria for research organisations under section 60d.
• TDM where a valid, machine-readable opt-out was in place at the time of use.

This distinction matters: while the Hamburg court confirmed that pre-processing steps for AI training fall under TDM (and so are protected), it also stressed that rights holders retain control through effective opt-outs, and that downstream uses (e.g., AI-generated outputs) remain subject to copyright scrutiny.

The Three-Step Test explained

Under EU law (Article 5(5), InfoSoc Directive), all copyright exceptions must satisfy three conditions:

1. Special case

The use must fall within a clearly defined exception. Here, sections 44b and 60d UrhG provide that framework.

2. No conflict with normal exploitation

The use must not significantly undermine the primary market. The court emphasised that internal downloads for dataset validation do not compete with licensing markets. Potential future impacts from generative AI outputs were considered too abstract.

3. No unreasonable prejudice to rights holders

A proportionality test: weighing research and innovation interests against property rights. The court found that non-commercial research and internal processing outweighed any harm. Especially since rights holders can deploy machine-readable opt-outs.

Practical implications

For rights holders

• Implement machine-readable opt-outs (robots.txt, TDM Reservation Protocol, metadata tags).
• Document opt-out deployment and maintain evidence for the relevant time period.
• Consider contractual clauses with agencies and platforms to enforce technical compliance.

For AI developers and research organisations

• Ensure compliance with opt-out detection obligations (see Article 53 of the EU AI Act).
• Maintain transparency on non-commercial purpose and governance to avoid losing section 60d protection.
• Document research activities and publish validation results to strengthen the scientific character of the work.

Open questions for the revision

The Court allowed a further appeal to the Federal Court of Justice (BGH), highlighting unresolved issues:

• Who bears the burden of proof for TDM opt-outs?
• What qualifies as “machine-readable” in practice - especially for historical periods?
• How should courts weigh the impact of generative AI outputs on the normal exploitation of works?

Final observations

The Hamburg decision confirms that German copyright law accommodates AI innovation, but under clear conditions. Rights holders must act proactively to protect their works, while AI developers must respect opt-outs and maintain research integrity. The next round before the BGH will likely set the tone for Europe-wide debates on copyright and AI.