AI Training and Copyright: Text and Data Mining

Text and data mining and AI training: the legal question

Training an AI model requires copying large amounts of text, images, music and code. Under German copyright law, those copies need either the rightholder's consent or a statutory exception. The relevant exceptions are the rules on text and data mining in the Copyright Act (Urheberrechtsgesetz, UrhG), which implement Articles 3 and 4 of the EU Directive on Copyright in the Digital Single Market (EU) 2019/790. § 44b UrhG permits text and data mining for any purpose, including commercial AI training, unless the rightholder has reserved its rights — for works available online, in machine-readable form.

§ 60d UrhG adds a broader exception for scientific research. German courts have now tested both exceptions in the first AI cases, and the AI Act adds transparency duties for providers of general-purpose AI models.

For businesses, the issue arises on both sides. Rightholders — publishers, photographers, music companies, software owners — want to control and monetise the use of their content. AI developers and companies fine-tuning models want legal certainty about their training data. And companies deploying AI need to know whether outputs that reproduce protected content expose them to claims. The law is developing quickly: within roughly a year, German courts have issued decisions on dataset creation, on language models and on music generators, and the Federal Court of Justice is expected to rule on the first of these cases.

The two text and data mining exceptions

§ 44b(1) UrhG defines text and data mining as the automated analysis of one or more digital or digitised works to obtain information, in particular about patterns, trends and correlations. Under § 44b(2), reproductions of lawfully accessible works for text and data mining are permitted and must be deleted when no longer required. Under § 44b(3), the use is permitted only if the rightholder has not reserved it; for works accessible online, a reservation is effective only if made in machine-readable form.

§ 60d UrhG allows reproductions for text and data mining for the purposes of scientific research by research organisations that pursue non-commercial purposes, reinvest all profits in research or act under a state-recognised public-interest mandate, and by cultural heritage institutions and individual non-commercial researchers. Research organisations cooperating with a private company that has decisive influence and preferential access to results are excluded (§ 60d(2) sentence 3). There is no opt-out for § 60d.

The LAION case in Hamburg

The first German judgment on AI training data concerned LAION, a non-profit association that built an image-text dataset used to train image generators. A photographer sued because one of his photographs, offered via a stock photo agency whose terms reserved rights against automated use, had been downloaded and analysed. The Regional Court of Hamburg dismissed the action (judgment of 27 September 2024 – 310 O 227/23), relying on § 60d UrhG. The Hanseatic Higher Regional Court dismissed the appeal on 10 December 2025 (5 U 104/24).

According to the court's press release, the Higher Regional Court confirmed that creating the dataset was scientific research within § 60d, even though commercial providers could also use the dataset, because no private company had decisive influence over LAION.

Going further than the regional court, it also held that LAION could rely on § 44b: the agency's reservation, attributed to the photographer, was not in the machine-readable form required by § 44b(3) sentence 2 at the time of the download. The court admitted an appeal on points of law to the Federal Court of Justice (Bundesgerichtshof, BGH), so the question of what counts as a machine-readable reservation may receive a final answer.

Memorisation: the GEMA cases in Munich

The Regional Court of Munich I took a different angle in GEMA v OpenAI (judgment of 11 November 2025 – 42 O 14139/24). GEMA, the German music collecting society, argued that song lyrics were memorised in the language models and reproduced almost verbatim in chatbot outputs. According to the court's press release, the court found that the lyrics were reproducibly contained in the models; both the memorisation in the models and the reproduction in the outputs interfered with exploitation rights, and these acts were not covered by the mining exception. The court granted injunctive relief, information and damages in substance.

In GEMA v Suno (judgment of 31 July 2026 – 42 O 763/25), the same chamber largely upheld claims against the provider of an AI music generator concerning well-known musical works, according to the press release of 31 July 2026. The decisions draw a line between analysing works to extract information — which the exceptions are designed to permit — and models that contain and reproduce the works themselves. Both judgments are first-instance decisions; appeals are possible.

Why the two lines of case law fit together

The Hamburg and Munich decisions are not contradictory. Hamburg concerned the compilation of a dataset — downloading and analysing images to extract information about their content — which is the core of what the exceptions permit. Munich concerned the end product: models that had absorbed specific works so completely that they could reproduce them. The exceptions authorise copies made in order to analyse; they do not authorise a model that functions as a copy of the work. For AI developers, this suggests that compliance needs to cover both the collection stage and the behaviour of the trained model.

The opt-out in practice

For rightholders, the machine-readable reservation is the key tool for content online. The statute does not define the technical format. Common approaches include instructions in robots.txt directed at AI crawlers, metadata in files and on websites, and emerging protocols designed to express reservations for automated mining. Reservations in terms of use written in natural language carry the risk, illustrated by the Hamburg case, that a court finds them insufficient for online works. Rightholders should therefore:

• implement technical, machine-readable reservations on all websites and platforms where their content is published;
• align terms of use, licence agreements with agencies and platforms, and metadata;
• document when reservations were put in place, because the relevant time is the moment of copying.

For AI developers, the corollary is to respect reservations at the time of collection, document crawling practices and keep records of sources — obligations reinforced by the AI Act.

The AI Act: copyright policy and training data summaries

Providers of general-purpose AI models must put in place a policy to comply with EU copyright law and, in particular, to identify and comply with rights reservations under Article 4(3) of the DSM Directive, including through state-of-the-art technologies (Article 53(1)(c) of the AI Act).

They must also draw up and publish a sufficiently detailed summary of the content used for training, according to a template provided by the AI Office (Article 53(1)(d)). These obligations have applied since 2 August 2025, with transitional arrangements for models already on the market. The Commission's general-purpose AI code of practice includes a copyright chapter describing how providers can implement these duties. See our insight on general-purpose AI provider duties.

Fine-tuning, retrieval and in-house use

Most companies do not train foundation models; they fine-tune existing models on their own data or connect models to document collections through retrieval-augmented generation (RAG). The copyright analysis differs. Fine-tuning on a company's own documents raises few issues where the company owns the rights, but third-party content — industry standards, licensed databases, purchased reports, customer materials — may be subject to licence terms that prohibit such use.

RAG systems copy documents into an index and reproduce passages in answers; this is closer to traditional reproduction than to analysis for patterns, and the mining exceptions may not cover the output. Licence terms for databases and publications should therefore be reviewed before content is connected to AI systems.

Data licences and contractual allocation of risk

Where training or fine-tuning data is licensed, the contract should state the permitted AI uses (training, fine-tuning, retrieval, output generation), territorial scope, duration and what happens to models trained during the licence term after it ends. Licensors typically seek attribution, usage reporting and restrictions on memorisation; licensees seek warranties that the licensor holds the necessary rights and indemnities against third-party claims. Agreements with AI providers should similarly address whether customer inputs may be used for training, and who bears the risk if outputs infringe.

Training outside the EU

Copyright is territorial. Training that takes place outside the EU may be assessed under foreign law — for example US fair use — but German courts may still have jurisdiction where the model is offered or its outputs are reproduced in Germany. The Munich court's decision in GEMA v Suno addressed training in the United States and use in Germany. In addition, the AI Act requires GPAI providers placing models on the EU market to respect EU rights reservations irrespective of where training occurs, as recital 106 explains. Providers cannot avoid EU copyright standards simply by training elsewhere.

Practical steps for companies

Rightholders

1. Audit where your content is published and implement machine-readable reservations.
2. Consider licensing models for AI training where you want to monetise access.
3. Monitor training data summaries published by GPAI providers for evidence of use.
4. Preserve evidence of memorisation and reproduction, including prompts and outputs with timestamps, before taking action.

AI developers and companies training or fine-tuning models

1. Map training data sources and their licence or exception basis.
2. Respect machine-readable reservations and document compliance.
3. Test for memorisation and implement output filters.
4. Use licensed data for fine-tuning where possible, and include warranties and indemnities in data licences.
5. Prepare the AI Act training data summary and copyright policy where you are a GPAI provider.

How TRW & Co helps with text and data mining and AI training

We advise rightholders, AI developers and companies deploying AI on copyright questions around text and data mining and AI training. For rightholders, we design reservation strategies, licensing models and enforcement actions. For AI developers, we review training data pipelines, data licences and AI Act copyright policies and training data summaries. For companies using AI, we assess output risks and negotiate indemnities with providers. See our pages on the AI Act and IP and patents, our insight on copyright in AI-generated content, or contact us.