What happened
It was reported that on 11 November 2025 the 42nd Civil Chamber of the Regional Court of Munich I, which specialises in copyright, gave judgment substantially in favour of GEMA against two companies in the OpenAI group, in case 42 O 14139/24. Nine songs were at issue, by well-known German writers and composers, among them Atemlos and Wie schön, dass du geboren bist. The court found the lyrics reproducibly contained in the language models GPT-4 and GPT-4o. It described memorisation, where training material is not merely analysed for the information in it but taken into the model’s parameters completely enough that the works can be extracted again as output, and it ruled out coincidence as an explanation given the complexity and length of the texts.
It was reported that the court held GEMA entitled to the claims it asserted on two separate bases: the reproduction of the texts in the language models, and their reproduction in the chatbot outputs. It rejected the defendants’ argument that responsibility for the outputs lay with users rather than with them, on the ground that the models substantially influenced what was generated. It also rejected their reliance on the exception for text and data mining. The court accepted that the exception can cover the copying needed to assemble a training corpus, such as converting material into another digital format or holding it in working memory. What it did not accept is that the exception reaches memorisation, because in the court’s view the premise behind it, that the automated analysis of information does not itself touch the author’s exploitation interests, no longer holds once the works themselves have been taken into the model. The court granted GEMA’s claims for an injunction, for information and for damages, and dismissed a claim founded on personality rights concerning altered lyrics attributed to the writers. The press release states no damages figure, and records that the judgment is not final. Initiative Urheberrecht reported that OpenAI filed an appeal on 8 December 2025 under case number 6 U 3662/25 e, and that the Regional Court had declined a proposal from both parties to refer questions of European copyright interpretation to the Court of Justice. No ruling on that appeal was identified in the sources reviewed as of August 2026.
What an auditable version would have shown
The case was decided from the outside. A claimant prompted deployed models, recorded what came back, and the court compared those outputs with the works, which it found enough to establish memorisation without access to any training data inventory. That is testing from the outside, and it can only ever sample. An auditable version would answer from the inside. A record made at training time says which works entered the corpus, under what licence or claimed exception, and with what evidence of provenance. A record made at generation time says, for a given output, which model version produced it and what it returned. Neither record decides the legal question. What they do is spare a claimant, a regulator or a court from having to infer what went into a model from what the model can be induced to give back.
Where the gap was
It was reported that the court found the lyrics reproducibly contained in the models, and that the question was settled by external testing, with a claimant prompting a deployed product and the court assessing what came back. A ConductRecord created at generation time preserves the model version, the request and the output for each relevant response, which is what allows a specific output to be traced afterwards rather than reconstructed by a party trying to reproduce it. A ConstraintGate applies a standing check before delivery, so an output matching a registered protected work is held rather than delivered and argued about later. Neither control resolves the legal question now under appeal, which is whether and on what terms training on protected works is permitted. What they change is the evidence available to the parties and the court when that question arises.
What governance should have looked like
It was reported that the judgment is a first-instance decision and is under appeal, so the position in German law is unsettled and what is set out above is the view of one court. Where an organisation trains a model on works it does not own, good practice is for the composition of the training data to be recorded as it is assembled rather than reconstructed under litigation, for the basis relied on for each source to be recorded with it, and for outputs to be checked against a register of protected works before they reach a user. The court did not accept that the text and data mining exception covered what it found the models to be doing. Whether that holds is a question for the Higher Regional Court, and this entry will need revisiting when it rules.
Failure Pattern: complete works remained retrievable from a deployed model, and the public record identifies no training data inventory or output level record that would show which works entered the corpus or when one of them was reproduced.
Governance Principle: an organisation that trains a model on other people’s works should be able to show from its own records what went into the training data, and to establish before an output is delivered to a user whether it reproduces one of those works.
The reference implementation of ConductRecord and ConstraintGate is open source. It lives at github.com/saffronandindia/headlights-oss, Apache 2.0 licensed and free to install. The repository is public now.
Sources
- Press release 11/2025 on the judgment of 11 November 2025 in case 42 O 14139/24 (Landgericht München I, Bavarian State Ministry of Justice)
- OpenAI legt Berufung ein (Initiative Urheberrecht)
- Munich Regional Court upholds GEMA’s claim against OpenAI for unauthorised reproduction of song lyrics (IRIS Legal Observations, European Audiovisual Observatory)