Getty sues over use of images to train AI

James Talbot, Mark Marfé and Bella Phillips of Pinsent Masons said that where third party data or IP is being used, a licence should be obtained to use that data or IP and that organisations that do not acquire appropriate licences risk infringing copyright or other IP rights that subsist in the data.

Talbot and Phillips were commenting after Getty Images announced that it had commenced legal proceedings against Stability AI before the High Court in London, alleging infringement of its IP rights. Stability AI’s ‘Stable Diffusion’ system automatically generates images. According to the company’s website, the tool has been trained using data compiled from “a general crawl of the internet” by the Large-scale Artificial Intelligence Open Network, a German charity. It said the training was carried out by academics at the University of Heidelberg “in compliance with German law”, and that the data was “not filtered to exclude or include any specific group”.

Getty said: “It is Getty Images’ position that Stability AI unlawfully copied and processed millions of images protected by copyright and the associated metadata owned or represented by Getty Images absent a license to benefit Stability AI’s commercial interests and to the detriment of the content creators.”

“Getty Images believes artificial intelligence has the potential to stimulate creative endeavors. Accordingly, Getty Images provided licenses to leading technology innovators for purposes related to training artificial intelligence systems in a manner that respects personal and intellectual property rights. Stability AI did not seek any such license from Getty Images and instead, we believe, chose to ignore viable licensing options and long‑standing legal protections in pursuit of their stand‑alone commercial interests,” it said.

Out-Law asked Stability AI for its response. In a statement, a spokesperson for the company said: “Please know that we take these matters seriously. It is unusual that we have been informed about this intended legal action via the press. We are still awaiting the service of any documents. Should we receive them, we will comment appropriately."

Talbot said that using data to train AI tools is vital for ensuring the effectiveness of those systems. He said IP rights can subsist in even raw, unstructured data, but that they are particularly likely to subsist in images. He said UK law places restrictions on how such works can be processed by software. Talbot also cautioned that even where IP protection does not apply, the data’s use may be limited by contractual terms.

The Copyright, Designs and Patents Act 1998 provides limited exceptions to copyright infringement. One exception allows text and data mining (TDM) of copyrighted works for non-commercial purposes, provided that the AI developer has lawful access to the work via, for example, a licence, subscription or permission in terms and conditions. Last year, the UK government confirmed its intention to expand the scope of that exception to enable TDM of works protected by copyright and database rights for any purpose. Those reforms have not yet been implemented, however.

“If copyrighted works are used without the rights owner’s permission or without an exception applying, this will constitute copyright infringement,” Talbot said. “When using someone else’s IP, businesses need to hold a licence to do so. The rights of use will be subject to the terms of that licence. Often, rights holders will not issue a general use licence, meaning businesses wishing to use their IP will need to review the granular detail of the licensing agreement or terms and conditions for the software in question, to understand what IP applies, what rights of use have been licensed, and ultimately what use is allowed.”

“Stability AI’s website refers to their model having been trained in compliance with German law. It will be interesting to see how this statement will impact its arguments in the UK case. This case highlights the importance of doing due diligence on the data and other IP rights that are intended to be used for training AI tools. It is not uncommon for IP rights holders to robustly monitor use of, and then enforce their rights in, that data and IP,” he said.

Marfé said that the English courts are well placed to hear disputes in relation to AI. “This case is yet another example of the English courts’ flexibility and innovation in its ability to apply the existing law to new and significant technologies such as AI," he said. 

Phillips said sometimes there can be a misconception among users of AI technology and of the outputs of AI technology that they do not need a licence to use IP in background development work if that IP does not feature directly in their own AI systems or in the output from those systems.

Phillips said: “Valuable background IP, such as data or other materials used to train an AI tool, can necessarily become integrated with foreground IP – which can be IP that subsists in the AI tools themselves or in their development – so that the background IP is then present in the developed tool and perhaps the output results that the developed tool creates. Some businesses may wish to use dummy or test data instead of valuable confidential data or IP to train the tool, but the type of AI tool being used may mean this will not always work.”

“Licensing agreements often contain wider restrictions on use beyond those that relate to IP rights, so AI developers using data or works obtained from a third party need to be aware of those terms to avoid a breach of contract and/or IP infringement. Some businesses may consider that licensing-in AI tools or using external parties to develop such tools on their behalf is preferable to developing them in-house. Each model has different financial, commercial, data protection and IP implications that businesses will need to weigh up,” she said.