A groundbreaking copyright ruling has emerged from Germany, with potential far-reaching consequences for the AI industry. The Regional Court of Munich has declared that copyright-protected works can indeed exist within AI models, a decision that has sparked intense debate and raised crucial questions about the future of AI development and intellectual property rights.
The case, brought by German collecting society GEMA, focused on OpenAI's ChatGPT and its use of copyrighted lyrics during the model's training process. The court ruled that OpenAI infringed the copyright of song writers by incorporating their lyrics into the model's parameters, allowing users to access and display those lyrics through specific prompts.
At the heart of the ruling is the court's finding that copyrighted lyrics were memorized within ChatGPT. This raises the question of whether the text and data mining (TDM) exception to copyright, as provided by EU law, allows developers to reproduce copyrighted content in their AI outputs without consent. The court suggested that this exception only covers preparatory measures for training data sets.
Dr. Nils Rauer, an expert in AI and IP law, emphasized the potential significance of this judgment. He highlighted the game-changing nature of the idea that copyright-protected works can exist within LLMs, leading to an obligation for model operators to prevent users from accessing and reproducing such works.
However, Rauer anticipates an appeal, believing that the legal questions at the core of this case will likely be considered and ruled upon by the EU's highest court. In the meantime, some AI providers have already taken precautionary measures, with ChatGPT and Copilot now displaying copyright notices when searching for the lyrics in question.
Alexander Bibi, a copyright expert on Rauer's team, added that the crucial question for the Court of Justice of the EU (CJEU) is whether large language models (LLMs) can be said to memorize content to the extent that the content used for training exists within the LLM. The answer to this question, Bibi explained, is critical for interpreting EU copyright law and determining whether copyrighted content has been 'reproduced' within the model and its outputs.
The Munich court's ruling expands the concept of 'reproduction' beyond identical copies, stating that it can take different forms. The court cited an example of copyrighted content being reflected in digital tokens within an LLM, with the key factor being whether the content is perceptible through user prompts.
Furthermore, the court considered that it doesn't matter if reproduction relies on probabilities or if copyrighted content is memorized as a puzzle within an LLM, as long as the entire puzzle can be revealed through user prompts. The rights holders, the court stated, don't need to understand the memorizing process within the LLM to claim unauthorized reproduction of their works.
Dr. Kai Welp, general counsel of GEMA, welcomed the ruling as a milestone, clarifying central legal questions regarding the interaction of new technology with European copyright law.
Rauer emphasized that this is just one of many AI copyright cases likely to come before EU courts, with potential divergence in case law presenting increasing uncertainty for both AI developers and rights holders. He highlighted the contrasting view taken by the High Court in London in the recent Getty Images v Stability AI case, where the court ruled differently on the question of whether an AI system could store and reproduce content.
The Munich court drew a clear line between preparatory acts covered by the TDM exception and the memorizing of lyrics within the AI model, applying the exception up to the point where the analysis of training data ends and the shaping of LLM parameters begins.
This ruling, Rauer said, will spark intense debate on balancing the need for well-trained AI tools that benefit society and the protection of intellectual property rights. If this judgment sets the precedent for understanding memorized content within AI models, it could lead to a reconsideration of the scope of the TDM exception by legislators.
