While current discussions in the European Union suggest that the use of ChatGPT is just a stepping stone towards something bigger – the adoption of general-purpose AI for more sophisticated applications – it threatens to render copyright laws outdated. ChatGPT produces text that looks as if it was written by a very well-informed human. It is so convincing that it has been cited as a co-author in some research publications. 

This kind of AI model raises three copyright issues: should outputs from AI models be protected by copyright? Do AI models  infringe on human  authors’ copyright when the model’s outputs are based on the creative productions of human authors? And can these outputs be considered creative for legal purposes? 

The short answer to these questions is that the law is unclear. 

While some owners and users of AI have asserted copyright over outputs of the AI, this is neither clear-cut in law nor the likely motivation of the producers of AI such as ChatGPT. Simply put, lines of code that are not the direct result of human labour and creativity and, by extension, texts produced by those lines of code, are not subject to copyright since they lack human authorship. The creators of ChatGPT developed a machine made from software that produces outputs that mimic human creativity. They made the machine through a process that involved forcing masses of data through filters of increasing complexity. They did not write the resulting code – the system developed it, even if under their guidance. The users of ChatGPT do not generate the outputs themselves. They merely supply a prompt. However, if a human prompts a ChatGPT output and then edits the AI’s output significantly the answer could change, but copyright will apply only to the human’s contributions.

Read more in Daily Maverick: “AI can help us find some answers in the elusive quest for ‘quality’ in education”

There are other reasons that it is unwise to extend copyright protection to AI-produced works. These works are routinely developed using pre-existing information or texts, most of which are under copyright. This makes it difficult to consider AI-generated outputs truly original, a requirement in most jurisdictions for something to be copyright protected. The test for determining originality in South Africa is known as “sweat of the brow”, which asks how much effort, skill and creativity went into the development of a work. The work must also be creative (unique) and not a slavish copy of another as stated in the Moneyweb v Media24 case. AI-generative technologies do not pass this test. Technologies like ChatGPT scrape the web for texts and produce results based on probabilities and thus objectively lack the creative element. And, currently there is no consensus on whether the process of scraping could even be deemed “creative”. Nor do we know how unique and distinguishable AI-generated works may become. 

More questions

Whether ChatGPT’s outputs infringe copyright with respect to their inputs raises two further questions: was the use of the input in training permitted, and is enough of the input reproduced in the outputs so that it is infringing? If the entire dataset is in the public domain there is no infringement. However, if not, then permission may be required for use in training. Technologies like ChatGPT and Microsoft’s Copilot omit footnotes and credits for the original sources by design. Despite this, if it is possible to reverse-engineer the generating processes to discover the sources, and if the results incorporate enough of the source materials, they may infringe existing copyrights. 

Read more in Daily Maverick: “Why is everyone so angry at artificial intelligence?”

This is a problem for artists. Their collective work is used to create systems that make it easy to produce competing products. Even if they tried to enforce the remedies available to them through copyright, the territorial nature of laws means that enforcement across international lines would be an expensive and labour-intensive process. So, what can be done?

The most pragmatic intervention would be to deny AI-generated works copyright protection, thus reducing the incentive to generate them as replacements for artists’ works.

Restricting AI research in South Africa will not work since the major global powers are already competing for the most advanced AI. Restrictions will merely result in South Africa failing to build its own capacity. South Africa would fall behind in developing new applications and advancements, making it harder for the country to compete with nations which have invested in AI and are using it to generate economic growth and innovation. AI can have positive effects on social, economic and environmental issues such as healthcare, climate and agriculture. Hence, rather than forbidding AI research and use, South Africa should embrace it and enhance its capacity in this area. 

It’s unclear that a blanket prohibition on using copyrighted works to train AI makes any legal sense because that may also preclude important uses of AI such as folding proteins in drug development, which would be an infringement on the right to scientific research protected by the Bill of Rights. 

Read more in Daily Maverick: “Artificial intelligence has a dirty little secret”

Trying to write a law that distinguishes permissible from impermissible use for AI research while new generations of AI emerge every six months is like building on quicksand. Instead, the law should include a flexible, open-ended exception that makes it clear that using copyright works to train AI for important research is acceptable, except where it would unfairly impact on creative works – for example, if an AI was trained to imitate the style of an artist. This would both protect artists and still allow for the public benefits of AI research. 

The most pragmatic intervention would be to deny AI-generated works copyright protection, thus reducing the incentive to generate them as replacements for artists’ works. This may meet little resistance from the creators of AI – given the copyright uncertainty that exists, copyright is an unlikely motivation of their work – but it could result in resistance of those who wish to assert copyright over AI works, as argued here, without legal cause. 

Parliament is debating a copyright amendment bill introduced in 2017. While the delays to the bill are regrettable, it provides an opportunity to clarify that copyright applies only to human creativity. South Africa can be a world leader in protecting creatives, without blocking AI research. DM/MC

Hanani Hlomani is a researcher at Research ICT Africa and a doctoral candidate at UCT. Dr Andrew Rens is a senior researcher at Research ICT Africa.