Getty v. Stability: United Kingdom Appeal Revives Debate Over Whether AI Models Store Images

A major court ruling found that AI models do not include copyrighted training data. Now, this decision is being appealed, and the industry is left wondering about possible territorial loopholes.

Getty Images sued Stability AI in London, claiming 12 million of its photos were scraped without permission. They wanted to set a UK legal precedent against AI training. However, the November ruling was more complicated. Getty now has permission to appeal, according to IP firm Murgitroyd. The case will decide if UK creators are protected when their work is used in datasets processed in other countries. The appeal centers on whether AI model weights, which are billions of numbers that determine what the AI can do, are legally considered "copies" of the original training data.

Mrs Justice Joanna Smith's original judgment drew a sharp technical distinction. "An article which is an 'infringing copy' must have at some point in time consisted of, contained, or stored a copy of a copyright work," she ruled. She found that statistical parameters—the numbers derived from images to describe patterns or features—are different from the images themselves.

The ruling was a major win for Stability AI, according to Osborne Clarke. It confirmed that owning or sharing Stable Diffusion's weights, which are the numerical settings that let the AI generate images, does not count as secondary copyright infringement under UK law. Mayer Brown explained, "An AI model, by contrast, contains statistically trained parameters—not stored copies or reconstructions of photographs." In this context, 'secondary infringement' means being held responsible for copyright infringement for actions like possessing or distributing infringing material, even if you did not make the original copy.

Getty dropped its main infringement claim because Stability trained Stable Diffusion in the United States. Ken Clark, an IP partner at Aird & Berlis, said: "Acts that occur outside the U.K. fall beyond the statute's reach, even if their effects are felt domestically."

This situation creates a major loophole. A UK company could legally own and share an AI model trained on millions of works by British creators, as long as the training happened outside the UK. The weights, which result from that training, are still considered separate from copyright infringement under the current rules.

Despite these bigger challenges, Getty did have a small win on trademark issues. The court found limited infringement when Stable Diffusion outputs showed altered versions of Getty's watermark. Still, this narrow victory does not change the overall situation.

The main technical issue in the appeal is more than just wording. If model weights are not legally seen as part of the training data, then secondary liability does not apply. The industry has already adjusted. Midjourney keeps its weights private, avoiding distribution issues. Stability AI, which shares its weights openly, now has a UK precedent saying those weights are not a problem, at least until the appeal court decides.

Getty's appeal will be difficult. They need to persuade judges that mathematical parameters actually contain the millions of images used to create them. This idea goes against how machine learning works. If Getty does not succeed, Parliament may have to step in, since current copyright law does not match the reality of AI's cross-border training.

In the end, the case has already created what Ken Clark calls "the first framework for AI-related IP litigation in the UK." For now, though, this framework means British copyright law stops at the border, even though AI training data can move freely between countries.