The jurisdiction problem in South African AI copyright
The infringement happened in a data centre in Oregon. The infringing output landed on a phone in Durban. The writer carrying the loss sits in neither country.
Sign up for my Substack daily AI newsletter here.
Follow me on LinkedIn: https://www.linkedin.com/in/johanosteyn/
A South African writer discovers that a chatbot can reproduce the cadence of her novel, in full paragraphs, in her own voice. A Pretoria composer hears his melody rephrased in a generative music tool. A Rhodes academic finds a model summarising her monograph with a precision that only reading the book itself could produce. In each case, the creator has a plausible copyright grievance, and in each case the grievance has nowhere to go. The infringing act, the training run, happened on a server in a jurisdiction that will probably never see her case. The infringing output appears on a phone in Durban, Johannesburg or Cape Town, which cannot by itself anchor a claim. The writer is left with every element of a legal case except the one that matters. A route to court.
CONTEXT AND BACKGROUND
That absence of a route is not for want of a copyright debate. The Constitutional Court’s recent judgment on the Copyright Amendment Bill reopened the public debate about what counts as a protected creative work in South Africa, including the long-contested fair use clause. The judgment matters, but it was written for a world in which copying was still a human activity. Independent academic analysis of the ConCourt record has focused on the three-step test and the human user. The ingestion of entire bodies of South African literature, journalism and music by foreign-trained generative models sits outside that frame. Investigations by the Sunday Tribune earlier this year documented how South African music is being harvested at scale and reused in AI-generated tracks with no compensation attached.
INSIGHT AND ANALYSIS
The jurisdiction problem has three strands. The first is where the infringing act happened. Training reproduction occurs on infrastructure outside South Africa, usually in the United States or Western Europe, and foreign courts have been reluctant to let South African claimants in through the service address alone. The second is where the output lands. A model’s answer appearing on a phone in Durban is an act of consumption, not reproduction, and local copyright law does not easily anchor there. The third is who could be sued. The foundation model publisher, the platform integrator, the hosting cloud provider and the end-user application are often four different companies in four jurisdictions with four liability positions.
South African legal commentary this year has argued that the country currently has no settled framework for licensing or opting out of AI training, which means the only route is litigation. The international response is instructive. The European Union offers a transparency obligation and an opt-out signal under its text and data mining regime, which gives creators somewhere to stand. The United States is litigating the question through fair use defences that may or may not settle. South Africa has neither.
IMPLICATIONS
The practical consequence is already visible. South African creators are not, in the main, bringing claims against foreign AI labs, because the cost of doing so exceeds the realistic recovery. A continental study of the same problem reaches the same conclusion for African music licensing against generative tools. The absence of enforcement is read as consent. Every month South Africa does not build a workable remedy regime, more South African creative output enters a foreign training corpus with no price attached and no route for the creator to object.
Three practical steps would open the door. A registration channel through the Companies and Intellectual Property Commission that lets a creator mark a work as closed to AI training. A domestic small-claims track for AI copyright disputes under a prescribed value, so a Johannesburg writer does not need senior counsel to be heard. A role for the Information Regulator in connecting training-data cases to the Protection of Personal Information Act where the training set includes identifiable local individuals.
CLOSING TAKEAWAY
The copyright question in South African AI is not whether local work is being used. The question is whether the law gives the person who made it a route to a hearing. Without that route, the formal rights do no work. The Constitutional Court’s judgment on the Copyright Amendment Bill is a necessary first act in a longer play. The next act belongs to Parliament, to the Department of Trade, Industry and Competition, and to the collecting societies and author associations who can organise what individuals cannot. The absence of a remedy is being read as consent, and the industry is operating on the consent it has inferred. Until the route exists, every South African creative work is on an open licence to the AI industry, and the licence pays nothing.
Author Bio
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
Comments