GEMA wins against OpenAI: Strengthening copyright in the age of AI

1. Background and key points of the judgement

GEMA, the collecting society representing the interests of around 100,000 music creators in Germany, filed a lawsuit against the OpenAI group of companies. The starting point of the proceedings was the use of song lyrics by nine well-known German authors (including “Atemlos” by Kristina Bach and “In der Weihnachtsbäckerei” by Rolf Zuckowski) in the training and output of the AI-based language model ChatGPT (link: https://haerting.de/wissen/do-you-hear-the-same/)

The Chamber ruled that OpenAI had infringed the copyright exploitation rights of GEMA members both by training its models with copyrighted song lyrics and by reproducing them at the request of users. The claims for injunctive relief, information and damages were essentially confirmed. By contrast, GEMA’s claims based on infringements of personal rights – for example due to the incorrect attribution of altered song lyrics – were rejected.

Significantly, the court ruled that not only the reproduction of the song lyrics in the chatbot output, but also the technical storage and reproducibility of these lyrics in the parameters of the language models constitutes a copyright-relevant reproduction. This means that the AI training itself is also subject to a licence.

Unlike OpenAI, the court is not of the opinion that “text and data mining” (TDM) within the framework of German Section 44b UrhG constitutes a sufficient barrier for this use. In the opinion of the court, the cases of “memorisation” found go beyond the extent provided for by law and infringe the economic interests of the rights holders.

2. The “memorisation” of copyright-protected content in AI language models

A central point of the ruling is “memorisation”: the court considers it proven that the OpenAI models can reproduce song lyrics in the original wording and make them accessible on request. From an information technology perspective, it is known that training data can be contained in AI models in such depth that the original text can be extracted in response to suitable user queries.

Through memorisation, there is an embodiment as a prerequisite for the copyright reproduction of the song lyrics in dispute through data in the specified parameters of the model. The song lyrics in dispute are reproducibly fixed in the models.

This technical embodiment (“fixation” in parameters) is sufficient under Art. 2 of the InfoSoc Directive and Section 16 of the German Copyright Act (UrhG) to be considered a “reproduction” within the meaning of copyright law. It does not matter whether the reproduction is made in conventional memory or merely as reproducible probability values – what is decisive is the practical reproducibility of the output. The song lyrics generated in response to user requests match the training data considerably in terms of scope and wording. The court therefore ruled out chance as a cause (such as statistical approximation). – The court apparently does not require a “bit-for-bit” copy: it is sufficient if the work can be technically reconstructed by the language model and can be retrieved on request.

3. Copyright restrictions and their limits: Text and data mining (Section 44b UrhG)

OpenAI argued that the training and operation of the AI fell under the barrier for “text and data mining” (TDM) created in the most recent copyright law (Section 44b UrhG). This provision was introduced with the European DSM Directive in order to enable automated analysis processes.

The court clarifies in this regard: – In principle, the training of AI-based systems falls under the TDM barrier, provided that it involves the automated extraction of information patterns and mere reproduction or analysis acts are carried out without economic impairment of copyright interests. – However, if there is “memorisation” – i.e. entire works can be reproduced from the model – this is no longer a mere analysis.

This infringes the economic exploitation interests and no longer constitutes a privileged act under Section 44b UrhG. An analogue application is also out of the question. Even if one wanted to assume an unintended loophole because the legislator was not aware of the memorisation and the associated permanent copyright-relevant reproduction in the models, there is a lack of comparable interests:

With the permissibility of preparatory acts of reproduction in the case of text and data mining, the limitation provision standardises a situation in which the exploitation interests of the authors are not jeopardised because mere information is extracted and the work as such is not reproduced. In the case of reproductions in the model, on the other hand, the exploitation of the work is permanently impaired and the legitimate interests of the rights holders are infringed as a result.

It is also not an “insignificant accessory work” pursuant to Section 57 UrhG, as the amount of lyrics used is neither incidental nor insignificant in relation to the overall data set.

Nor could it be assumed that the rights holders had given their consent,

as the training of models is not to be regarded as a usual and expected type of use that the rights holder must expect.

The reproduction of the song lyrics as output in the defendant’s chats also constituted unauthorised reproduction and making available to the public of the song lyrics in dispute.

4. Who is responsible for the output?

According to the Chamber, OpenAi was responsible for the output. The architecture and the training – and thus the ability to memorise lyrics – originate solely from the sphere of the AI provider. The generation of the copyright-protected outputs is largely controlled by the model and its parameters, not by the user. – The operators of language models with memorised works must therefore be liable for the resulting outputs.

5. Consequences: Signalling effect and business implications

For the first time, a court in Europe has ruled on the use of copyright-protected works by AI systems in a professional context. The Regional Court of Munich I is of the opinion that both the training of AI and the output of the AI are subject to licensing under copyright law – provided that works worthy of protection are not only analysed, but also defined in the AI models in such a way that they are “memorised” in the original and reproduced almost in the original.

The judgement is not yet legally binding. However, there are very good arguments in favour of AI providers having to conclude licensing agreements with collecting societies such as GEMA in future if their model is trained with protected content and memorisation cannot be ruled out. Users of AI solutions in companies should ensure that the systems used have been legally trained and that licences for copyright-relevant content are available, as injunctive relief and claims for damages can also be asserted against them. Rights holders and artists can benefit more in future if the AI is trained and works with their works.

OpenAI has announced an appeal. It is surprising that the Munich Regional Court has not referred the case to the European Court of Justice. Especially as the question of the copyright permissibility of the use of AI will also have to be clarified at European level.

6. Conclusion

The judgement of Munich Regional Court I underlines that innovative AI systems on the German and European market are subject to copyright rules. “Memorisation” represents a new form of use that is not covered by existing restrictions such as text and data mining, provided it goes beyond mere information analysis. With this decision, Gema has come closer to its goal of reaching a licence agreement with OpenAI and other AI providers.