Greg Troxel <[email protected]> writes: > This issue of whether LLM models and LLM outputs are derived works of > LLM training data is *not at all* settled (well established) law. > However, Anthropic settled a suit for $1.5B (1.5E9 US dollars!):
A question that has been decided (for now) in Germany is whether it constitutes copyright infringement when an LLM reproduces an existing work with “easy” prompts: https://www.gesetze-bayern.de/Content/Document/Y-300-Z-GRURRS-B-2025-N-30204 Answer: if the work was part of the training data and it doesn’t take very complicated prompts to get it re-created (even if not completely unchanged), then this is a copyright violation by the one providing the LLM. That then means, it’s memorized (in a changed form), so it violates the copyright. The training corpus used was Common Crawl. The court also says that training on data without ensuring that there’s no memorization of copyright protected works is a violation of due diligence. What this likely means for software: if a model easily re-creates an amount of code from the training data that crosses the threshold of originality (⇒ similar to "non-trivial"¹ in the US), then this is a copyright violation (if the court getting the case follows the decision of the Munich court). ¹ The rule of thumb by the GNU project is that 15 lines or more is non-trivial: https://elpa.gnu.org/packages/doc/compat.html#Development-1 By that rule, it would be a copyright violation if even a single 15 line function from the training data is reproduced with only trivial changes. But keep in mind that Germany does *not* have case law, so later decisions can differ. Somehow I feel compelled to say “please don’t hate me for stating the current legal situation”, but looking deeper I realize that this is a reflex and I actually don’t care. Best wishes, Arne -- Unpolitisch sein heißt politisch sein, ohne es zu merken. https://www.draketo.de
signature.asc
Description: PGP signature
