IP and generative AI: what you need to know

Generative AI systems were virtually unheard of among the general public until the launch of ChatGPT-3.5 in 2022. But, as we looked at in our previous post, they are now being used widely by individuals and businesses alike. The latest versions of gen AI systems like Chat-GPT, Dall-E, Adobe Sensei, Midjourney, GitHub Copilot, and AlphaCode (among many others) are able to write prose, produce images, and code to a level that was still the stuff of science fiction fantasy only a few years ago. 

In the 2 years since GPT-3.5 first seized the public consciousness, businesses have understandably been scrambling to procure and deploy generative AI tools to get ahead of the ever-steepening curve and avoid falling behind. For any business looking to actively use generative AI, intellectual property considerations should be baked into their intended use case and choice of generative AI tool. 

The risks and key considerations from an IP perspective are similar to when procuring content from a human third party, whether a freelancer or another business, but there are some significant differences. Applied to gen AI, they take 3 main forms: the risk of ‘input infringement’; the risk of ‘output infringement’; and ownership issues. 

A. Input infringement

Generative AI systems can be trained in different ways but they tend have one thing in common: the use of vast amounts of third-party data such as images or text. If the developer of the AI model has not secured the relevant permissions to use that data, they risk infringing relevant IP rights such as copyright or database right. 

Gen AI may still be in its infancy but we are already seeing content owners bring ‘input infringement’ claims against AI developers:

1. In the UK, Stability AI – the developer / operator of the text-to-image tool ‘Stable Diffusion – has been sued by Getty Images. Among other things, Getty claims that Stability AI unlawfully trained Stable Diffusion on millions of images on Getty’s website, infringing Getty’s copyright in those images. 
   
   A key issue in the claim is where that training took place: in the UK or abroad. If Stability AI’s location-based defence succeeds, it could pave the way for certain jurisdictions to introduce commercial ‘web-scraping’ exemptions to IP infringement and bill themselves as ‘AI training havens’ (though international IP conventions make this unlikely in major jurisdictions).
   
   UK copyright legislation only provides a defence for web-scraping for text and data analysis for the sole purpose of non-commercial research. In 2023 the UK Government, as part of its White Paper for ‘A Pro-innovation Approach to AI Regulation’, proposed expanding the defence to apply to text and data mining for any purpose. The proposal resulted in a significant backlash from the creative community and was replaced with a proposed industry consultation. At the time of writing that consultation also appears to have been parked.    
    
2. In the US, a number of authors including George RR Martin and John Grisham have brought a class-action lawsuit against OpenAI and Microsoft, claiming that OpenAI’s language model was trained on authors’ works without their consent. Similar claims have been brought against the developers of other generative AI models.   

The risk of input infringement is predominantly an issue for the developers and operators of generative AI models but users should still keep it in mind, particularly from a PR perspective. 

B. Output infringement

‘Output infringement’ is where the content generated by an AI system – i.e. its output – infringes a third party’s intellectual property rights. Copyright poses the greatest risk but AI content is also capable of infringing other IP rights, such as trade marks and design rights.  

Businesses using AI-generated content should take at least the same steps to protect themselves from liability for infringing third party IP rights as they would if they were producing content themselves or procuring it from a non-AI source. 

1. First, the risk of infringement can be significantly reduced simply by not trying to copy the works of others in the first place (assuming copyright has not expired). Internal staff training and policies on the use of generative AI play an important role and offer the opportunity to promote AI best practice on related issues such as cybersecurity, confidentiality, and data protection. 
2. Second, the risk of inadvertent infringement should be covered off by the terms and conditions of the AI tool, which should warrant that the content produced by the AI tool will not infringe third party rights. Ideally they should also indemnify the user from losses or costs flowing from any such infringement. Some of the major operators of generative AI systems – such as Microsoft’s Copilot – include these protections in their standard terms and conditions but they may be limited in scope and subject to carve outs or exclusions, so it is important that businesses carefully consider whether they offer the protection they need for their generative AI use case. 

C. Ownership issues

Users should not assume that they own the copyright in the content they produce using generative AI systems, if copyright even exists in that content at all. This is particularly important for businesses looking to sell or license AI-generated content. 

On ownership, at first glance UK copyright legislation appears somewhat ahead of its time in providing that the owner of copyright in computer-generated works is the person who made the “arrangements necessary for the creation of the work”. But in the case of generative AI it is unclear whether this would be the user entering prompts or the developer of the AI system itself. The answer may vary from work to work, depending on the detail and importance of the user’s inputs and prompts. 

Ownership can – and should be – dealt with by the AI system’s terms and conditions, but it doesn’t matter what the contract terms say if copyright does not exist in the AI content in the first place: the AI operator cannot assign ownership of any copyright to the user if  there is no copyright to assign. 

Copyright is a creative IP right and only comes into existence when something ‘original’ is created. The bar for originality under UK and EU law is relatively low but at its core is the concept of personal creativity: a work is original if it bears the author’s formative touch. This concept is arguably uniquely human, meaning there is a real risk that works produced by generative AI tools in response to user prompts may not benefit from copyright protection in the UK or EU at all. 

As for the issue of ownership, the position may vary from case to case, depending on the level of human involvement in the creation of the content (a distinction can be drawn between computer-generated content and computer-aided content) but with no relevant court decisions or legislation on the horizon the answer may remain unclear for a number of years. 

If you would like to discuss any of the IP issues raised in this article please contact Harry Jewson or Emily Roberts in our IP team. For wider AI queries please contact David Varney, Tom Whittaker, or Liz Smith in our Technology team.

For the latest updates on AI law, regulation, and governance, see our AI blog at: AI: Burges Salmon blog (burges-salmon.com). 

Subscribe to our Concept newsletter and receive the latest intellectual property legal updates, news and event invitations direct to your inbox.

Users shouldn't assume content generated by AI will be protected by copyright. Without copyright protection, users may not be able to stop third parties reusing their AI-generated content.