The High Court of Justice in London has issued a ruling in the dispute between Getty Images and Stability AI. At the heart of the case were allegations of copyright infringement (both primary and secondary), database rights infringement, trade mark infringement and passing off, all linked to Stability AI’s open-source generative model, Stable Diffusion.
Getty argued that Stable Diffusion had, without authorisation, acquired millions of works (copyright-protected images and videos), many of them owned by Getty or licensed to it on an exclusive basis, and had used them to train the model through scraping carried out also on websites and platforms controlled by the claimant.
During the proceedings, however, the claim for primary copyright infringement relating to training was abandoned. It emerged that the infrastructure used for training (AWS) was located outside the United Kingdom, that the scraping and materialisation of images took place on foreign servers, and that no part of the training process was carried out on UK territory. Since copyright is territorial in nature, UK law can only sanction acts of copying carried out within the United Kingdom.
The fact that training takes place abroad, however, does not “immunise” the model provider with respect to the outputs. If, once placed on the market, the model generates images that recognisably reproduce protected works or distinctive signs, this may amount to a new primary infringement of copyright or trade mark rights in the country where those outputs are made available to users.
By contrast, the claim for secondary copyright infringement was dismissed on the merits. The Court held that Stable Diffusion cannot be classified as a “counterfeit copy” of Getty’s works, finding that the model weights do not contain images nor do they reproduce any protected work, but are numerical parameters describing patterns and features learned during training. In the absence of any copy embedded in the “article” (the model), the essential requirement for secondary infringement is not met.
On the trade mark front, Getty achieved partial success. The Court held that the normal use of the model in the United Kingdom may, in some cases, generate synthetic images bearing Getty’s trade marks, including watermarks. In particular, in the initial versions of the model (v1.x) ISTOCK watermarks were detected, while in the later versions (v2.x) GETTY IMAGES watermarks were generated, with a sufficient risk of confusion to amount to infringement.
The Getty Images / Stability AI judgment is significant because, on the one hand, it rules out that a generative AI model which does not store copies of protected works can be classified as a “counterfeit copy” for the purposes of UK copyright law; on the other hand, it confirms that the generation of watermarks and other distinctive signs can amount to trade mark infringement, even where the underlying image is entirely synthetic.