German Researchers Argue AI Training's Legal Shields Are Quietly Crumbling
A new arXiv paper argues that training generative AI on copyrighted data likely isn't text-and-data-mining or fair use, with memorization creating separate liability even for legally trained models.
Takeaways
- Paper argues generative AI training fundamentally differs from TDM and does not qualify for EU exceptions.
- Memorization creates copyright liability independently, since model weights encode training data.
- Munich I court ruled against OpenAI in November 2025, rejecting the TDM defense for memorized lyrics.
- Larger models memorize 2 to 5 times more than smaller ones in the same family.
- Neural audio codecs like EnCodec fall under the same legal analysis as generative models.
- Authors propose a tiered documentation framework using URLs, ISRC codes, and fingerprinting.
A new overview paper from researchers at Otto-von-Guericke-University Magdeburg and Leibniz University Hannover makes an uncomfortable argument for anyone building generative models on scraped data: the legal shields most labs are hiding behind may not actually cover what they are doing. The paper synthesizes an interdisciplinary study originally published as a German open-access book, and it sharpens the technical case against treating GenAI training as ordinary data mining.
In a legal statement filed as part of an RIAA copyright lawsuit against Suno AI, the company said its training data includes essentially all music files of reasonable quality accessible on the open internet, arguing this qualifies as fair use. Similar disputes are now producing actual rulings, and the paper walks through what they mean for anyone shipping a generative model.
Two legal shields, both leaking
Two frameworks dominate the debate. In the EU, the 2019 DSM Directive carves out a Text and Data Mining (TDM) exception. Article 3 permits reproductions by research organizations for scientific research on lawfully accessed works, while Article 4 permits commercial TDM but lets rightholders opt out through machine-readable means such as a robots.txt file. In the US, the fair use doctrine applies a four-factor test around purpose, nature of the work, amount used, and market effect.
Neither cleanly fits generative training. TDM is traditionally defined as identifying patterns and extracting useful knowledge from data repositories, while GenAI models are trained to approximate the probability distribution of their training data with the aim of producing outputs that look like that data. One approach extracts insight; the other reproduces style. That difference matters legally because TDM exceptions were written with statistical analysis in mind, not systems designed to output substitutes for the input works.