South Africa Has No Lawful Route To Train An AI Model

Neither the Copyright Act nor the pending Bill contains a text and data mining exception, and the practice that a Californian court legalised would likely infringe here.

Sign up for my Substack daily AI newsletter here.

See my AI Training course portfolio for corporate Business Leaders here.

Article link: https://johanosteyn.substack.com/p/south-africa-has-no-lawful-route?r=73gqa&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true

Follow me on LinkedIn: https://www.linkedin.com/in/johanosteyn/

There is a question that governs whether an entire industry can legally operate in this country, and it has not appeared in a decade of copyright debate. May a machine read a copyrighted book in order to learn from it? In South Africa, the answer is probably no, and almost nobody has noticed.

CONTEXT AND BACKGROUND

In Bartz v. Anthropic, a United States federal judge in the Northern District of California held that training an AI model on lawfully acquired copyrighted books was fair use. The company had sourced books two ways: by downloading millions from pirate sites and by lawfully buying physical copies, scanning them and storing them digitally. The court drew a line between the two, finding the piracy could not qualify while the lawfully acquired books could, describing the purpose and character of using copyrighted works to train models as quintessentially transformative. South African law reaches a different conclusion.

Our Copyright Act of 1978 follows a closed-list fair dealing approach permitting limited, purpose-specific exceptions such as private study, criticism, review and news reporting, and legal analysis concludes that training an AI system on copyrighted materials would likely not qualify under any of them, meaning such training would likely infringe.

INSIGHT AND ANALYSIS

The Copyright Amendment Bill was supposed to change that by introducing a fair use clause modelled on American doctrine, and many commentators read it as doing exactly what Bartz did. On 26 June the Constitutional Court handed down judgment in Ex Parte President of the Republic of South Africa, the first time it has undertaken an abstract review of a bill’s substance before enactment. It declared copyright to be property protected from arbitrary deprivation, upheld the shift from fair dealing to fair use, and struck down the broad educational exceptions in sections 12D(1) to (5). The President may therefore not sign the Bill until Parliament cures the defect. Critically, the Court upheld the surviving exceptions by reading them narrowly, and that reasoning binds their future scope.

Analysis by Adams & Adams, who acted for the seventh amicus in the matter, argues that those constraints effectively foreclose any claim that the Bill’s exceptions could authorise AI training, since such use is not personal use, does not share the attributes of the named educational and research purposes, and fails the mandatory substitution-effect test. The same analysis notes that neither the Act nor the Bill contains a text and data mining exception and that there is no statutory safe harbour for AI training, and that neither the Bill nor the judgment contemplates generative AI at all.

IMPLICATIONS

So South Africa has arrived at one of the more restrictive positions in the world on AI training, through a judgment about textbook photocopying, in legislation drafted before ChatGPT existed. Nobody chose this. It is the residue of a decade-long process aimed at other problems. For local businesses, the practical consequence is immediate. Any South African firm fine-tuning a model on local material, whether legal texts, medical records, news archives or literature, is operating without a clear legal basis, and the standard advice is to confine training data to public domain material, openly licensed content, non-copyrightable information and properly licensed proprietary content. For policymakers, the position is worse than restrictive. It is unconsidered.

CLOSING TAKEAWAY

Parliament must reopen the Bill regardless, because the constitutional defect leaves no alternative. That creates a narrow and unearned opportunity to ask a question the process never asked, which is whether a country that wants an AI industry can have one under a copyright framework designed in 2015. Other jurisdictions have answered deliberately, some permissively and some restrictively, and each did so knowing what it was deciding. We are about to do it by omission. The question is not whether South Africa should copy California. It is whether we intend to decide this at all, or simply discover in five years what we accidentally chose.

Johan Steyn is a prominent AI thought leader, speaker, and author with a deep understanding of artificial intelligence’s impact on business and society. He is passionate about ethical AI development and its role in shaping a better future. Find out more about Johan’s work at https://www.aiforbusiness.net