Keep up-to-date with the latest IP insights and updates as well as upcoming webinars and seminars via HLK’s LinkedIn page, or simply subscribe to our updates.
In this article, Roland Weede explores the rise of design protection inGermany, from a neglected afterthought to a recognised necessity of property law, exploring how shifting legal priorities and industrial pressures shaped its evolution to today.
Get in touch
Landmark AI Decision: German Court rules AI training on Song Lyrics infringes Copyright
In a landmark decision, the Munich Regional Court I (case No. 42 O 14139/24) ruled that OpenAI’s use of copyrightes song lyrics to train its “chatbot” ChatGPT constitutes unauthorized reproduction under German copyright law. The court action had been filed by Germany’s copyright collective GEMA, an agency aiding musicians in the monetization of their works by enforcing royalty payments (akin to the UK’s PRS).
Training the software behind GPT and similar chatbots, which are based on so-called Large Language Models (LLM), requires exposure to enormous amounts of text, which are used for building a model of statistical relationships between words. This model is then used for predicting the “most probable” next word in a reply to user input.
The Internet offers an ideal source for such text, which includes, among large amounts of freely distributable text (bulletin boards, forums, Wikipedia articles etc.), many texts which are protected by copyright, for example song lyrics. Website like AZLyrics or Genius make those lyrics readily available for almost any song. Those freely accessible lyrics are part of the training corpus for OpenAI’s LLMs.
For this reason, GEMA filed a lawsuit against OpenAI for copyright infringement, and for violation of personality rights at the Munich Regional Court I.
What makes the case interesting is the fact that LLMs do not copy text as such and store the contents directly, but only adjust parameters based on the whole corpus of the training texts. German copyright law contains an exception allowing data mining, i.e. the automated analysis of copyrighted works for extrapolation of trends and correlations (see sec. 44b and 60d Copyright Act). OpenAI claimed in its defence that its use of song lyrics for LLM training falls under this data mining exception. The actual output itself could not constitute infringement by OpenAI, since it was a result of, and informed by, previous user input (the prompts).
GEMA, on the other hand, showed that it was possible (at least in certain versions of GPT) to obtain a full reproduction of song lyrics, using very simple prompts. In those cases, repeated exposure to the same lyrics in training data had resulted in memorization of the actual text.
In yesterday’s press release, the court announced that its 42nd chamber, specialized in copyright cases, had dismissed the claims for the violation of personality rights, but ruled in favour of claims for copyright infringement by unauthorized reproduction of protected works. The memorisation of texts in LLMs as well as the reproduction in the output were deemed infringing.
Since reproduction can happen “by any means and in any form, in whole or in part” according to Art. 2 of the Information Society Directive (Directive 2001/29/EC), it did not matter that the reproduction was done by repeated adjustment of statistical parameters instead of direct encoding. This reproduction was not covered by the data mining exception, since it is not merely a reproduction as part of the training corpus, but the result of the training process.
The defence argument that the output is just the result of user prompts was dismissed by the court pointing out the simple and generic nature of the prompts which had resulted in repetition of the memorised text.
The decision is not yet final, and OpenAI is expected to file an appeal. The questions about the exact boundaries of the data mining exception, and reproduction by memorization in electronic models will certainly remain relevant, considering the immense potential impact on this rapidly evolving field of technology.
Need assistance?
This is for general information only and does not constitute legal advice. Should you require advice on this or any other topic then please contact hlk@hlk-ip.com or your usual HLK advisor.