Munich court rules OpenAI infringed German song lyric copyrights

The Munich Regional Court ruled that OpenAI’s language models infringed the copyrights of German songwriters, in a case brought by the collecting society GEMA over lyrics to nine well-known songs, including Helene Fischer and Kristina Bach’s “Atemlos,” Herbert Grönemeyer’s “Männer” and Reinhard Mey’s “Über den Wolken.” It is the first European court decision to hold an AI developer directly liable for how training data is retained inside a model, rather than only for what a chatbot outputs.

GEMA’s case rested on the fact that ChatGPT could reproduce the lyrics near-verbatim from simple prompts. The court treated this as evidence that the underlying model had “memorised” the copyrighted text — stored it in a retrievable form within its parameters — and found that this memorisation itself constitutes a reproduction under German copyright law, independent of whether a user ever prompts the model to output it. OpenAI had argued its training fell under the EU’s text-and-data-mining exception, which permits copying works to extract information from them. The court rejected this, reasoning that the exception covers the transient analytical phase of training, not a model that permanently retains protected content and can reproduce it later. Output of the lyrics by ChatGPT was treated as a separate, further infringement — an adaptation and public communication of the work.

The court largely upheld GEMA’s claims for an injunction, disclosure of relevant information and damages (to be quantified separately), while dismissing a secondary claim over alleged personality-rights violations. Because the judgment concerns memorisation as a category, rather than a specific output, lawyers described it as potentially reaching any large language model trained on copyrighted text without a licence, not just song lyrics.

The ruling is not final; OpenAI can appeal to Germany’s Higher Regional Court, and a reference to the Court of Justice of the EU on interpretation of the text-and-data-mining exception remains possible. It gave rights-holders across Europe a legal theory — memorisation as reproduction — that sidesteps the harder question of what a model does with data at output time.