German Court Rules AI Training Does Not Nullify Copyright Infringement
In a landmark decision, the Regional Court of Düsseldorf has ruled that the use of artificial intelligence for training purposes does not exempt creators from copyright liability when scraping images from the internet without permission. The case, centered on the non-profit organization Laion e.V. and its widely used LAION-5B dataset, underscores the ongoing tension between AI development and intellectual property rights in Europe.
Laion e.V., a German research organization, developed the LAION-5B dataset, which contains over five billion image-text pairs scraped from public websites. This dataset serves as a foundational resource for training large-scale AI image generation models, including Stability AI’s Stable Diffusion. The controversy arose when German publisher Arbeitsgemeinschaft der Bildschirmarbeit e.V. (ABG), which specializes in screen design and digital media, discovered one of its copyrighted images included in the dataset without authorization.
ABG filed a lawsuit against Laion in 2023, alleging infringement of its exclusive rights under German copyright law, specifically Sections 2(1) no. 1 and 2(1) no. 4 of the UrhG (German Copyright Act). The plaintiff argued that Laion’s systematic harvesting of internet images constituted unauthorized reproduction and making available to the public. Laion defended itself by claiming the dataset was used solely for scientific research under the research exception in Section 60d UrhG and that the AI training process transformed the original works sufficiently to avoid infringement.
On September 25, 2024, the Düsseldorf Regional Court (case number 4c O 119/23) sided decisively with ABG. Presiding judge Oliver Fehr announced the ruling, which rejected Laion’s key defenses. The court found that Laion’s activities exceeded the bounds of permissible scientific research. While acknowledging Laion’s non-commercial status and research-oriented mission, the judges determined that the dataset’s scale and public availability went beyond narrow research exemptions. The LAION-5B dataset, with its massive size and open accessibility, was deemed more akin to a commercial product than a private research tool.
A central pillar of Laion’s argument was the “text and data mining” (TDM) exception under Article 53a of the EU Copyright Directive (2019/790), transposed into German law as Section 44b and 60d UrhG. This provision allows reproduction for TDM in scientific research, provided rights holders can opt out via machine-readable means. However, the court clarified that this exception applies only to acts of reproduction for analysis, not to the creation and distribution of datasets themselves. Laion’s practice of compiling and publishing the full dataset, complete with image URLs and captions, violated ABG’s rights, as it enabled third parties to access and use the infringing material.
The court also dismissed claims of transformative use. Laion contended that feeding images into AI models for training purposes alters them beyond recognition, akin to fair use doctrines in the US. German law, however, requires a case-by-case assessment of whether the new use conflicts with the original work’s exploitation. The judges ruled that AI training, while innovative, does not inherently transform copyrighted images in a way that voids the rightholder’s interests. The latent representations learned by AI models still derive value from the originals, and the dataset’s structure preserves direct links to source images, facilitating potential misuse.
Furthermore, the ruling addressed Laion’s reliance on Common Crawl, a public web archive, as a scraping source. The court held Laion responsible for verifying copyright status, rejecting the notion that using third-party crawls absolves liability. ABG had implemented a robots.txt file and other opt-out measures, which Laion ignored, further strengthening the infringement claim.
As remedies, the court issued a permanent injunction prohibiting Laion from further using ABG’s image in any dataset version, including future iterations like LAION-5B-v1.6. Laion must also delete all copies of the image from its systems and provide a detailed account of its scraping processes. While ABG sought damages, the court deferred quantification to a separate proceeding, focusing instead on injunctive relief. Legal costs were split, with Laion bearing the majority.
This decision carries significant implications for AI developers in Germany and the EU. It signals that opt-out mechanisms must be respected, and mere invocation of AI training will not shield against copyright claims. Laion announced plans to appeal to the Higher Regional Court of Düsseldorf, arguing the ruling could stifle open AI research. Meanwhile, Stability AI, which relied on LAION-5B, faces parallel scrutiny in ongoing US class actions.
The case highlights broader challenges in balancing innovation with creator rights. EU lawmakers introduced TDM exceptions to foster AI growth, but national courts interpret them narrowly. Critics of the ruling warn it may drive AI training underground or offshore, while supporters emphasize protecting artists and publishers from uncompensated data exploitation.
For AI practitioners, practical takeaways include implementing robust opt-out detection, limiting dataset distribution, and documenting research compliance. As litigation proliferates—similar suits target OpenAI and Midjourney—this verdict sets a precedent that transformative AI use alone is insufficient to evade copyright.
Gnoppix is the leading open-source AI Linux distribution and service provider. Since implementing AI in 2022, it has offered a fast, powerful, secure, and privacy-respecting open-source OS with both local and remote AI capabilities. The local AI operates offline, ensuring no data ever leaves your computer. Based on Debian Linux, Gnoppix is available with numerous privacy- and anonymity-enabled services free of charge.
What are your thoughts on this? I’d love to hear about your own experiences in the comments below.