Munich ruling on GEMA v OpenAI reshapes AI copyright debate

In November 2024, German music rights society GEMA filed a closely watched lawsuit in Munich against OpenAI, accusing the US-based company of using protected song lyrics without permission to train its artificial intelligence models. Representing the creators of nine German songs and more than 100,000 composers, songwriters and publishers, GEMA alleged that OpenAI had reproduced lyrics drawn from its repertoire without purchasing licences or paying the artists. What had long been an abstract concern for the creative industries – that AI systems were ingesting vast catalogues of copyrighted works – was suddenly crystallised in a concrete legal challenge. The proceedings promised to test how far existing copyright rules could stretch to cover the internal workings of generative AI. When the written judgment arrived, it not only sided with GEMA but also set out a far-reaching view of what counts as infringement in the age of machine learning.

During the case, OpenAI offered a technical defence designed to distance its systems from traditional copying. The company told the Munich court that its language models do not store or duplicate specific song lyrics or other copyrighted texts. Instead, they were said to encode what they have learned as abstract patterns in their internal parameters, rather than keeping a retrievable database of works. From OpenAI’s perspective, this pattern-based learning meant that any overlap between training material and later outputs was incidental, not a direct reproduction of protected content. It argued that because the models contained only a “statistical impression” of the texts, they fell outside the ordinary bounds of copyright law. This explanation was presented as a core safeguard, intended to show that the company’s technology avoided the kind of literal copying that courts traditionally police.

OpenAI also advanced a responsibility-focused argument centred on who should be treated as the author of chatbot responses. The firm maintained that users are the true producers of outputs, because they decide what to ask, how to phrase their prompts and how to use the answers they receive. Any problematic reproduction of lyrics in a chatbot conversation, it said, would arise from user behaviour rather than corporate decisions. In that framing, the company acted as a neutral platform that merely facilitated interactions between users and an underlying model. By shifting emphasis towards user choice and control, OpenAI hoped to distance itself from direct liability for what the system might generate. The court, however, would ultimately take a very different view of where responsibility lies.

The Munich judges rejected both strands of OpenAI’s defence and instead drew an unusually sharp line around how German copyright law applies to modern AI systems. In their written ruling, they held that the mere fact of storing protected lyrics within a language model’s internal parameters amounts to an unlawful use, even before any user ever sees an output. Citing the training process directly, the court concluded that “both the memorisation in the language models and the reproduction of the song lyrics in the chatbot’s outputs constitute infringements of copyright law”. This reasoning directly challenged the notion that models retain only vague impressions of texts. Rather than treating memorisation as an incidental by-product of training, the court characterised it as a form of reproduction squarely covered by copyright rules.

For GEMA and the creators it represents, that recognition was pivotal. It acknowledged that the economic value of their works is engaged at the training stage itself, not only when lyrics are publicly performed or distributed. By locating infringement inside the black box of AI models, the judges signalled that the internal mechanics of machine learning systems are not beyond legal scrutiny. This approach also undermined OpenAI’s assertion that encoding data into model weights is fundamentally different from storing protected content. In the court’s view, what matters is that the protected texts are taken in and retained in a way that makes it possible for the system to reproduce them later. That is enough, the ruling suggests, to trigger authors’ rights under German law.

The judgment did not stop at the training phase. It went further by treating user-facing outputs as a separate and additional infringement, creating what amounts to a two-tier structure of liability. When the chatbot reproduces lyrics in response to a prompt, the court found that this visible output is not merely a technical side effect but a fresh act of unauthorised reproduction. OpenAI’s argument that users should be seen as the producers of such outputs, and therefore responsible for any legal breach, was firmly rejected. The judges emphasised that the system’s ability to produce protected lyrics was a direct consequence of how it had been trained and designed by the company. As a result, they placed the legal focus on the AI provider’s role in enabling those outputs, not only on the end-users who request them.

By drawing this distinction, the court effectively split potential compensation into two categories. First, the plaintiffs are entitled to payment for the reproduction of their texts within the language models themselves, reflecting the value of their works as training material. Secondly, they can claim remuneration for the reproduction of lyrics in chatbot outputs, which the court treated as a separate use of the same works. GEMA hailed this outcome as a landmark recognition of creators’ interests, describing it as the first major case of its kind in Europe. In practical terms, the ruling establishes a financial pathway for rights holders to seek rewards at both the training and output stages of an AI system’s lifecycle. That framework may now be tested in other disputes across the continent.

For OpenAI, the Munich decision lands at a moment when it is already facing a series of lawsuits in the United States. There, media organisations and authors allege that the company’s ChatGPT chatbot was trained on news articles, books and other works without permission. While the legal standards differ from Germany’s, the reasoning in the GEMA case could influence how courts and regulators elsewhere think about AI training. The Munich judges’ refusal to accept a purely technical distinction between “storing” and “encoding” content may complicate OpenAI’s reliance on similar defences abroad. At the same time, any appeal in Germany could produce a higher-court ruling that either confirms or narrows the scope of the initial judgment, adding another layer of uncertainty for the company and its peers.

Implications for creators and licensing models

For composers, songwriters and other creators, the ruling marks a potential turning point in how their work is valued in the era of generative AI. By confirming that both the memorisation of lyrics in language models and their later reproduction can infringe copyright, the court has strengthened the argument that training on protected works is not a cost-free resource. GEMA spokesperson Kai Welp said after a hearing in September that the case could have “vital implications for the remuneration of creative artists”, capturing the sense that this is as much about economics as legal doctrine. If similar views take hold in other courts, AI developers may be forced to treat lyrics and other creative outputs more like traditional broadcast or streaming uses that require permission and payment. That would move rights holders and collecting societies from the margins of AI development to the centre of a new licensing layer.

The decision also suggests a future in which collective rights bodies play a more assertive role in negotiations with AI firms. Societies such as GEMA are structured to license large repertoires of works on behalf of thousands of members, a model that could translate to AI training if courts continue to identify it as a copyright-relevant use. Developers might seek blanket licences for access to lyric catalogues or other content, in much the same way broadcasters obtain performance rights today. For many artists, this shift offers not only legal vindication but the prospect of sharing in the economic value that AI companies derive from their catalogues. It could turn what has often been experienced as uncompensated data extraction into a more structured marketplace, even if the exact pricing and terms remain to be defined.

On the industry side, the judgment pressures AI firms to reassess how they source and handle training data, particularly when it includes music, literature or journalism. Companies may need to negotiate licences directly or through collecting societies when using protected repertoires, adding a new category of royalties to their cost base. That, in turn, could encourage the emergence of standardised licensing frameworks designed specifically for AI, mirroring the way performance and mechanical rights are administered in music. Firms may also choose to segment their models, reserving fully licensed datasets for products offered in stricter jurisdictions such as Germany, while pruning unlicensed content that risks triggering similar claims. Any failure to adapt could expose developers to compensation demands wherever courts, as in Munich, find infringement in both internal model representations and user-facing outputs.

Operationally, the ruling points towards a future in which dataset curation and traceability become core compliance functions for AI companies. To demonstrate that they have respected authors’ rights, firms may be pushed to log the provenance of training material and filter out works controlled by societies such as GEMA unless they are covered by an appropriate licence. They may also need to monitor and redact content that appears likely to surface verbatim in responses. That could involve building tools capable of detecting and suppressing near-exact reproductions of songs and other creative works, or tightening usage controls when users request full lyrics or long passages of text. Model “hygiene” processes, once considered largely technical, now carry clear legal and financial stakes as courts look closely at what was ingested and what can be reproduced.

Some AI services may respond by limiting features that risk delivering copyrighted content in full, while highlighting capabilities that summarise, transform or analyse works instead. Such shifts might reduce the chances of direct reproduction while preserving the models’ usefulness for tasks such as drafting, translation or education. For creators, these changes would not only raise the prospect of new revenue streams but also offer greater transparency into how and where their works are used. Increased documentation of training data and licensing arrangements could make it easier for authors to track the role their material plays in AI development. Over time, that visibility may help rebuild trust between technology firms and cultural industries that have often felt sidelined by rapid advances in machine learning.

A wider legal test for AI training

The Munich judgment is expected to echo well beyond the dispute between GEMA and OpenAI. As the first major European ruling on the use of copyrighted works to train AI models, its reasoning will be closely scrutinised by lawyers across the EU. Rights organisations are already assessing how it might bolster their own claims on behalf of journalists, authors, visual artists and others whose works may have been used without explicit consent. An appeal by OpenAI would prolong the legal uncertainty, but it would also give a higher court the opportunity to clarify how far copyright extends into the inner workings of large language models. That clarification will matter not only for music publishers and songwriters, but for any sector whose works underpin AI services.

Beyond the courtroom, the ruling is likely to feed into broader negotiations between AI companies and collective rights bodies over licensing frameworks. If judges in future cases continue to treat both training and output as copyright-relevant uses, technology firms could face growing pressure to reach structured agreements with societies representing large repertoires of works. At the same time, regulators and policymakers in Europe are expected to study the Munich reasoning as they refine rules on data provenance, transparency obligations and what constitutes “safe” training practices. Any resulting guidance could influence how models are built, documented and audited, particularly in markets that have been cautious about the impact of AI on cultural ecosystems.

Future cases will test the reach of the Munich decision in several directions. One question is whether its logic extends beyond song lyrics to other creative formats, including journalism, books and images. Another is whether it applies equally to different types of generative models, from conversational chatbots to image generators and voice synthesis tools. As those disputes unfold, courts will be asked to balance the promise of innovation against the rights of authors whose works fuel the AI revolution. The GEMA v OpenAI ruling does not settle that balance, but it signals that European law is prepared to look inside the black box of AI training – and to demand that creators are paid when their works are used to build the next generation of machines.