Recent rulings in multiple jurisdictions highlight how courts are stepping in to clarify how copyright law applies to generative AI — particularly concerning the use of creative works in training AI models and the rights over outputs generated by those models. This emerging judicial guidance is shaping the conflict between creators and AI developers, underscoring deep tensions and uncertain balance.
In a landmark decision in the UK, the Getty Images’ lawsuit against Stability AI regarding its model Stable Diffusion was dismissed because the court found no evidence of unauthorized copying of images. The court held that the model’s internal data — its “weights” — do not store original images, but only statistical patterns derived from training data. As a result, those weights do not qualify as infringing copies under copyright law. Meanwhile, Stability AI was found liable, though only for limited trademark‑infringement linked to watermarking in certain outputs; the core copyright claims failed.
On the other hand, in Germany, a regional court ruled that training a language model on protected lyric texts (without permission) constitutes “reproduction” under copyright law — because the AI model’s parameters effectively encode the copyrighted works in a fixed form that can be later reproduced. Similarly, in the U.S., some cases have resulted in mixed outcomes: while one court found training an AI on pirated books to be “fair use,” another dismissed a fair‑use defense where an AI system used proprietary legal content, ruling this to be direct copyright infringement.
What emerges from these rulings is that results vary substantially depending on how courts interpret technical details — like how AI models store and generate content, what counts as “copying,” and whether the use is “transformative.” Legal experts are increasingly saying the industry may need a new middle‑ground: possibly a licensing or collective‑rights model where creators are compensated, and AI developers get access under regulated conditions.
