UK Court Rules AI Models Trained on Copyrighted Works Don’t Constitute Infringement

Here in the US, there is a continuing battle in the courts over how copyright law applies to generative AI. There are serious questions around whether AI training or outputs violate copyright law. AI firm Anthropic has already settled a major copyright case for $1.5 billion. Which was just one of several copyright cases pending around the country. However, it isn’t just US copyright law that AI firms must contend with. Copyright law can vary significantly from country to country, and recently, a major decision was handed down in the UK. Getty Images sued Stability AI over the use of Getty’s copyrighted works in Stability’s AI training. The court found that if an AI model is trained on copyrighted data, but that data isn’t present in the final model, then there is no infringement. A recent Sidley memo discusses this point, stating:

“The Court held that an article can be an infringing copy only if it actually contains or embodies the copyright work — at least transiently. The Court found that Stable Diffusion’s model weights did not store, reproduce, or contain any of Getty’s images; they were sets of numerical parameters derived from statistical training. Accordingly, the Court held that merely using infringing copies in the course of creating another artifact does not make that artifact an infringing copy and that the Stable Diffusion model ‘has never consisted of or contained a copy’ of Getty’s works.”

While the court didn’t rule substantively on several issues raised by Getty, this finding is an important one. Moving forward, we’ll see if US courts adopt a similar stance, or if we’ll see variety in how different countries apply their copyright laws to AI.