This is an informal case summary prepared for the purposes of facilitating exchange during the 2026 WIPO IP Judges Forum.
Session 3: Copyright and AI Model Training
Munich I Regional Court (Landgericht München I), Germany [2025]: GEMA v OpenAI, Case No. 42 O 14139/24
Date of judgment: November 11, 2025
Issuing authority: Munich I Regional Court (Landgericht München I)
Level of the issuing authority: First Instance
Type of procedure: Judicial (Civil)
Subject matter: Copyright and Related Rights (Neighboring Rights)
Plaintiff/Appellant: GEMA
Defendant/Respondent: OpenAI, L.L.C.; OpenAI Ireland Ltd.
Keywords: Reproduction; Communication to the public; Memorization; Text and data mining; Large language model; Outputs; Adaptation; Artificial intelligence
Basic facts: The plaintiff is a collecting society which administers rights in musical works, including the lyrics of nine German songs. The defendants developed, trained and operate a generative language model. Use of the lyrics as training data by the defendants was undisputed. On prompts asking for the lyrics of the songs, the defendants’ chatbot returned outputs reproducing the lyrics in whole with some minor deviations. The plaintiff brought proceedings seeking injunctions against reproduction in the model and reproduction and communication to the public via the outputs, together with information regarding the extent to which reproduction and making available to the public had occurred and the income generated from these acts, a declaration of liability in damages, publication of the operative part of the judgment, and pre-litigation costs. German law applied under Article 8(1) Rome II Regulation. The oral hearing was held on September 29, 2025.
IP issues before the Court:
1. Whether the song lyrics were “memorized” by the model during training.
2. Whether memorization of song lyrics by the trained model constitutes a reproduction of the works within the meaning of § 16(1) of the German Copyright Act (Article Art. 2 of the Directive (EC) 2001/29 (InfoSoc Directive)).
3. Whether any such reproduction in the model is covered by the text and data mining exception in § 44b of the German Copyright Act (Article 4 Digital Single Market (DSM) Directive) or another exception.
4. Whether the chatbot outputs infringe the rights of reproduction, making available to the public and adaptation under §§ 15, 16(1), 19a and 23(1) of the German Copyright Act, notwithstanding deviations from the originals.
5. Whether the defendants are liable as direct perpetrators for the outputs, or whether the prompting user is to be regarded as the person making the reproduction and making available to the public.
6. Whether the outputs additionally infringe the authors’ personality right through incorrect attribution of authorship.
Held: The action succeeded in substantial part. The Chamber ruled as follows on each issue:
1. Yes. The lyrics were memorized by the model during training, as the trained parameters in the model contain not merely extracted information but coded the content of the training data.
2. Yes. The memorization of lyrics constitutes a copyright-relevant reproduction in the models within the meaning of § 16(1) of the Copyright Act.
3. No. Neither § 44b nor another exception under the Copyright Act applies to the memorized lyrics. The text and data mining exception covers the reproductions needed to assemble the training corpus (phase 1 of model development and use), not the further reproductions retained in the model (phase 2 of model development and use).
4. Yes. The outputs reproduce and make the works available to the public, and § 23(1) of the Copyright Act extends protection to adaptations. At least one element of the song lyrics at issue that gives rise to protection was recognizable in the outputs in the copyright sense.
5. The defendants are liable as perpetrators. They exercise control over the reproduction, as they selected the training data and the architecture of the model and they operate the model. Simple, open-ended prompts did not transfer that control to the user.
6. No. The claim concerning naming of the lyricists failed; for two outputs there was no incorrect attribution at all, and for the remainder the conditions of a personality right claim were not met. That claim, and the associated information and damages claims, were dismissed.
Relevant holdings in relation to copyright and AI model training: The Chamber first confirmed that the lyrics are literary works. Following the Regional Court of Hamburg in LAION, it then distinguished three temporally successive phases: (1) extraction and conversion of the training material into machine-readable form; (2) analysis of that material and training of the model; and (3) subsequent use of the trained model through prompts and outputs. It treated phase 1 reproductions as preparatory and covered by § 44b of the German Copyright Act, because analysis as such is not a copyright-relevant act and reproductions made for subsequent analysis purposes do not touch the author’s exploitation interests. Phase 2 is different: where training does not merely extract information but results in works being reproduced within the model, the act falls outside the definition of text and data mining in § 44b(1) of the German Copyright Act and Article 2(2) DSM Directive, as in this case the authors interest in compensation is affected. On the evidence, the Chamber was satisfied under § 286 Code of Civil Procedure that the texts are contained in the models in reproducible form, relying on the technical literature on memorization and extraction. On the outputs, it applied the recognizability test, being the recognizability of the original elements of the work in the other subject matter, drawn from Article 2 of the InfoSoc Directive, holding that no line need be drawn between § 16 and § 23 of the German Copyright Act because every adaptation fixed in material form is also a reproduction. It rejected the defense of independent parallel creation. Where memorized texts are returned on simple, open-ended prompts, the models reproduce the training data rather than creating it independently.
Relevant legislation:
German Copyright Act (DE311); Directive (EC) 2001/29 (InfoSoc Directive) (EU309); Directive (EU) 2019/790 (EU Digital Single Market Directive) (EU240); Regulation (EC) No 864/2007 (Rome II); German Code of Civil Procedure (DE326)