UK High Court Rules Stable Diffusion AI Model Does Not Infringe Getty Images’ Copyrights

The High Court of England and Wales recently delivered a significant judgment in the case of Getty Images v Stability AI Ltd, addressing whether the Stable Diffusion artificial intelligence (AI) model infringes copyright. Getty Images alleged that Stability AI unlawfully used millions of its copyrighted photographs to train the Stable Diffusion model and that both the AI-generated outputs and the model itself infringed copyright.

Getty Images operates a global licensing business for photographs and visual content created by professional photographers and contributors. Its business model depends on licensing millions of images to media organizations, advertisers, publishers, and other commercial users. Stability AI develops generative AI systems, including Stable Diffusion, which generates synthetic images based on user-entered text prompts.

The lawsuit claimed that Stability AI copied millions of images from Getty Images’ websites without permission to train different versions of the Stable Diffusion model. Getty Images contended that many of these images were original artistic works protected by copyright, owned or exclusively licensed by Getty Images. The initial claims included allegations of unlawful copying during training, infringement by AI-generated outputs, and that the AI model itself constituted an infringing copy.

During the court proceedings, Getty Images acknowledged there was no evidence that the training and development of Stable Diffusion occurred within the United Kingdom. Additionally, Stability AI blocked the prompts used to generate the allegedly infringing outputs, leading Getty Images to abandon claims related to the training process and the AI-generated images.

Consequently, the court’s focus narrowed to whether the Stable Diffusion AI model itself could be considered an infringing copy under the secondary infringement provisions of the Copyright, Designs and Patents Act 1988 (CDPA). The key legal questions were:

1. Can the Stable Diffusion AI model be treated as an infringing copy of copyrighted works under the CDPA?

2. Could importing, distributing, or possessing the AI model amount to secondary copyright infringement?

3. Do the model weights created during training constitute copies of copyrighted works?

Getty Images argued that the training process created model parameters embodying information derived from the copyrighted photographs, effectively representing transformed copies of those works. It asserted that these model weights should be treated as infringing copies under the CDPA.

Stability AI countered that Stable Diffusion consists of numerical parameters encoding statistical relationships learned from the training data, not copies of the original images. The company maintained that the model parameters cannot be reconstructed into the copyrighted images and thus do not constitute infringing copies.

The court examined the secondary infringement provisions of the CDPA, which apply to articles embodying or containing infringing copies of copyrighted works. It observed that Stable Diffusion is a machine learning model composed of mathematical parameters created during training, which do not store or contain copies of the copyrighted images.

The court explained that while the model parameters influence how the AI generates images, they do not reproduce or store the original images themselves. Therefore, the AI model cannot be treated as containing copies of the copyrighted works.

Notably, Getty Images did not claim that Stable Diffusion stored copies of the images but argued that the creation of model weights should itself be considered an infringing copy. The court rejected this argument, holding that model weights do not constitute copies of the copyrighted works.

Ultimately, the court concluded that Stable Diffusion does not contain copies of Getty Images’ copyrighted photographs and thus cannot qualify as an infringing copy under the CDPA 1988. As a result, Getty Images’ claim that the AI model itself constituted an infringing copy failed.

This ruling clarifies the application of UK copyright law to AI models trained on copyrighted images, emphasizing that model parameters encoding learned information do not equate to infringing copies. The decision marks a significant development in the intersection of copyright and AI technology.