The High Court recently delivered its judgment in Getty Images v Stability AI. The decision marks a significant development in the use of generative AI and the enforcement of intellectual property.
Originally, Getty Images claimed that Stability AI's Stable Diffusion model, which creates images from text prompts, infringed copyright, trade marks and database rights. Before trial, Getty however abandoned some of its claims, so that the judgment ultimately focused on i) trade mark infringement; and ii) secondary copyright infringement. issues.
Key facts
The Stability AI model operates (much like other AI models) by scraping millions of images without permission in order to train its model. Released in 2022, Stable Diffusion generates synthetic images based on text prompts and has multiple versions (v1.x, v2.x, SD XL, v1.6). Getty alleged this training process infringed its intellectual property, with some AI-generated outputs containing Getty and iStock watermarks.
The judgment
The Court addressed secondary copyright and trade mark infringement and held that:
- Secondary Copyright Infringement: Getty's claim that Stable Diffusion was an "infringing copy" under the Copyright, Designs and Patents Act 1988 (Act) was rejected. It was found that Stable Diffusion was not trained in the UK, and therefore Stability AI's importation and distribution within the UK, was not dealing in an article which was an "infringing copy" under the Act. Further, whilst Stable Diffusion was trained using Getty's copyright protected images, the model itself did not store or reproduce these original images. Instead, it learned patterns from training data and generated outputs based on those patterns.
- Trade mark Infringement: Getty's trade mark infringement claims, based on GETTY and ISTOCK appearing as watermarks on AI-generated output images, partially succeeded. The Court ruled that the older Stable Diffusion versions (v1.x and v2.x) infringed Getty's and iStock's trade marks when outputs were generated by UK users via Stability's API and DreamStudio platforms. However, claims for newer models (SD XL and v1.6) were dismissed due to lack of evidence of UK users generating watermarked images.
The Court also dismissed Getty's Section 10(3) claim, it found use of the signs did not amount to detriment to distinctive character or reputation nor had there been any unfair advantage (free-riding) and citing insufficient evidence. The court gave no judgment on the passing off claims.
Takeaways
The decision provides some clarity on how intellectual property law applies to generative AI, in particular in the context of secondary copyright infringement. Training AI models on copyright protected data or materials outside the UK does not automatically constitute infringement.
For AI developers, the case emphasises ensuring AI models do not store or reproduce copyright protected material, both crucial elements for defending against secondary copyright infringement claims. In respect of trade mark infringement, developers should implement safeguards to prevent generating infringing outputs, particularly those that include trade mark watermarks.
The case provides valuable insights into the legal treatment of AI-generated content, underscoring the importance of robust risk management strategies and staying informed as the legal landscape around AI and intellectual property evolves. The decision does however leave some unanswered questions around AI models and primary copyright infringement. Getty abandoned this claim before trial, and we await further guidance from courts on: i) copyright infringement in any inputs used to train AI models; and (ii) in outputs generated by such AI models. Balancing AI developers and content creators' interests will no doubt continue to be a key priority for the government, and we await further guidance on how the government responds to the consultation on artificial intelligence and copyright.
