Beware the Minefield of Intellectual Property Law in the Race for Generative AI

While not a new invention in itself, generative artificial intelligence (AI) has recently come to the forefront of the public imagination with tools such as ChatGPT and DALL-E making generative AI available to the person on the street. The launch of ChatGPT in particular saw millions of people flocking to the chatbot for answers to all sorts of questions.

ChatGPT is able to answer questions using natural, human-like language and has helped people with various tasks, such as coding computer programs and composing music and documents, to name a few. When asked to define AI, ChatGPT explained that “[a]t its core, AI aims to replicate and mimic human cognitive abilities, such as perception, reasoning, learning, problem-solving, and decision making”.

However, with all the excitement comes serious questions relating to the interplay between generative AI and existing copyright laws, and the precautions that both developers and users of AI need to take to avoid infringement claims. This article will consider a few of the potential intellectual property rights issues arising out of the use of generative AI based on existing laws.

Almost all publicly available AI tools such as ChatGPT, Bard and Midjourney are considered generative AI. A key characteristic of generative AI is that it is pre-trained using large amounts of data from the internet and other sources. The output of any such tool is thus heavily dependent on the data used to train the model. Inevitably, questions arise as to what source material is used to train the generative AI, and to what extent the output of the generative AI is influenced by or is a copy of said source material.

Given the ability of generative AI tools to create new content, the question of whether copyright subsists in such output has increasingly become of concern. The apex court in Singapore has categorically held that copyright protection will only arise where the work was created by a human author or authors. The United States Copyright Office has similarly held that human authorship is required for copyright to subsist in a work. This may be contrasted with the position in the UK, where computer-generated works without a human author enjoy protection for a term of 50 years from the date the work is made, with the author of such a work taken to be the person who undertakes the arrangements necessary for the creation of the work.

At what point can copyright subsist in works created by a human but using computer software? The Singapore courts have observed that the mere fact that a literary work is reduced into its final form of expression through the instrumentality of computer software does not in itself divorce the work's origination from the person operating the machine, and that there had to be an authorial creation that was causally connected with the engagement of the human intellect for copyright to subsist, which meant the application of intellectual effort or creativity, or the exercise of mental labour, skill or judgement by the human author. As such, the position at the moment is that copyright may subsist in a work created by a person using software where the software acts as nothing more than a tool in the work's creation, much like the writing of a work with a pen.

While one might argue that generative AI is far from being anything as passive as a pen, an industry is developing over the creation of “prompts”, which are the inputs a human enters into a generative AI to generate a specific output. Proponents might argue that the generative AI would not create the desired output without the creativity and effort put into coming up with the right prompt. This is an argument that has yet to be tested in court.

Earlier in 2023, Getty Images filed a lawsuit against Stability AI, creators of AI art generator Stable Diffusion, alleging that the company copied 12 million images to train its AI model “without permission … or compensation”. In support of its complaint, Getty Images adduced images showing original photographs and similar images (complete with the Getty Images watermark) generated by Stable Diffusion.

Because AI tools are trained using large amounts of data points resulting in the AI tool producing output that closely resembles the works they are trained in, there may be prima facie infringement if the consent of copyright owners is not obtained before such works are used to train the AI tool. This is because the output (whether this is an image, literary work or otherwise) may, in some cases, be considered to be a copy of a substantial part of the original work. Infringement may also arise if copies of the original work are made and stored as part of the AI's training.

Would the use of original works in AI tools fall within the scope of fair use or other permitted uses under copyright laws? In Singapore, whether the copying of copyrighted works to train generative AI tools falls within the scope of fair use would depend on the facts of each case. In particular, the court would likely consider whether any such use is fair and sufficiently “transformative”, having regard to the relevant matters set out in Section 191 of the Singapore Copyright Act 2021.

Companies that are developing, using and looking to procure licences in generative AI tools should be aware of the potential legal pitfalls presented by existing intellectual property laws. These considerations, together with those relating to the use of personal data and confidential information, should be taken into account when designing or licensing an AI tool. It is also important to remember that intellectual property laws are territorial and vary across jurisdictions. Companies should therefore consider jurisdictional issues carefully, in addition to the various applicable laws.