The recent and first English judgment on generative artificial intelligence (AI) training and intellectual property (IP) infringement highlights a few key challenges and considerations for both those seeking to protect their IP and for AI developers and users of large language models (LLMs) from third party providers.
Getty had brought a claim against Stability AI alleging that Stability AI’s model had been trained on Getty’s images without their consent and therefore had infringed Getty’s IP rights. On initial review, the judgment could be seen as a win for AI developers, but there are still unanswered questions and this is unlikely to be the last high-profile case on AI and IP infringement.
There is now much anticipation for the Government’s responses to its Consultation on Copyright and AI, which seeks to create a balance between the creative industry and the AI sector – both of which are valuable to the UK.
The Case
There were three main issues:
Primary Infringement
The question as to whether the unauthorised use of IP content to train LLMs in the UK constitutes copyright infringement remains unanswered as there was a lack of evidence that Stability AI trained its models in the UK. This claim was withdrawn by Getty, so the court did not come to any conclusion on whether UK-based web scraping to train an LLM would infringe IP.
Secondary Infringement
Secondary copyright infringement occurs where an item, which is known to be infringing, is imported into the UK. The court was asked to decide whether Stability's AI model constituted secondary infringement. The court held that the imported work was not limited to something tangible and could include intangible electronic storage – i.e.an AI model.
Despite this determination, Getty’s claim failed because the court found that – whilst Stability’s AI model may have been trained using the images– the model itself did not store or reproduce those images, so there was no importation of infringing items. The mere fact that the model was exposed to and trained on infringing works did not make the model itself an infringing work.
Trademark Infringement
The court did find basis for trade mark infringement where some outputs generated by Stability’s AI model contained watermarks identical or confusingly similar to Getty’s watermark. However, the judge noted that this was "historic" and "extremely limited" in scope, so the claim was unsuccessful.
Conclusion
This judgment provides some guidance on how the courts will assess whether the use of AI infringes an owner’s IP. However, at present this area of law still requires much development and future cases or legislative developments are likely to provide more clarification, especially regarding primary infringement where AI developers train their models in the UK – as the court did not have to make a decision on whether UK-based web scraping which is used to train an LLM within the UK constitutes primary infringement – AI developers still face a significant risk that it might. On secondary infringement, the court was clear that an AI model could constitute an imported item for the purposes of secondary infringement, but the specific details as to where the infringed works sit will be relevant. Until a decision is reached in the English courts, there’s a potential gap for AI developers who can continue to train models outside the UK on infringing works and avoid infringement.
While we await the UK Government’s response to its Copyright and Artificial Intelligence consultation, caution should be exercised by AI and LLM developers and users, and IP owners should continue to protect and enforce their rights.
For AI/LLM developers – care should be taken in how the LLM is trained and how content is curated. If copyrighted content is scraped, it may constitute an infringement of IP. The safest approach will be to use freely available and royalty-free content wherever possible and agree licence deals with the author or owner if any copyrighted material is to be used.
For AI/LLM users – if using a third party LLM or AI service, there is a risk that the LLM and its outputs could infringe third party IP. Check what assurances and protections are given by the provider (if any) and consider carefully how you generate and use content, particularly if using a free or low cost service with few or no legal protections against IP infringement liability.
If your organisation is developing AI or integrating AI functionality into your products or services, we can assist with reviewing terms of use, checking for any risks relating to copyright infringement, drafting policies and procedures for how your organisation uses IP and advising on any specific application / industry / regulatory compliance issues.
Get in touch with our commercial team to find out more.