This is the first article in our series on Artificial intelligence (AI) and intellectual property for South African financial services firms. It should be read together with the article by Carla Dennehy, which addresses ownership of AI-generated outputs.
This article considers whether training or using an AI system infringes copyright, and what the emerging global position means for your firm.
Every generative AI system must be trained on data, much of which is protected by copyright. The central question courts, regulators and legislators are grappling with is whether that training process infringes the rights of copyright holders, and whether AI-generated outputs may themselves be infringing.
For South African financial services firms, this is not abstract. If the AI tools you procure were trained on infringing training data, or if their outputs contain infringing material, your firm may be exposed.
The infringement framework
Under South African copyright law (as under United Kingdom law), infringement does not require intent. The question is whether a "substantial part" of a copyright work has been reproduced. This is a qualitative, not quantitative, test. The portion reproduced must reflect the intellectual creation of the original author.
Two primary risks arise in the AI context: First,
- the training process may involve copying copyright works; and
- AI-generated outputs may reproduce a substantial part of those works.
A third, often overlooked risk is that uploading third-party copyright material into an AI tool as a prompt may itself amount to copying. Firms whose employees routinely upload third-party research or reports into AI tools should take this seriously.
No safe harbour in South Africa
The United Kingdom's text and data mining exception applies only to non-commercial research and does not provide a safe harbour for commercial AI training.
South Africa's position is more restrictive. The Copyright Act 98 of 1978 contains no equivalent exception. There is currently no legislative safe harbour for AI training on third-party copyright works in this jurisdiction.
Getty v Stability AI: what we learned
The UK High Court decision in Getty Images v Stability AI was expected to resolve the AI training debate. It did not. Getty withdrew its primary copyright infringement claims before the full hearing, leaving those issues unresolved. On the secondary infringement claim, the court found that the Stable Diffusion model did not store Getty's works and therefore did not constitute an infringing copy. This finding was expressly fact-specific.
The Court of Appeal has granted Getty permission to appeal. The legal position remains unsettled.
The global picture
On 18 March 2026, the UK Government published a report on the use of copyright works in AI development, together with an economic impact assessment. These publications provide a clear indication of the Government's position but do not change existing law or propose specific reforms.
The Government emphasised the need to balance protection of the creative industries with the promotion of AI development, noting that there is insufficient evidence to justify immediate legislative reform. The Government's next steps will focus on monitoring international developments and gathering further information before considering any changes to the Copyright, Designs and Patents Act 1988. Importantly, the question of whether using copyright works to train AI constitutes infringement remains unresolved.
In the European Union, Article 53 of the AI Act requires general-purpose AI providers to publish summaries of training data and implement copyright compliance policies. It does not resolve the infringement question but introduces transparency obligations.
In the United States, the Copyright Office confirmed in January 2025 that not all AI training qualifies as fair use. Where AI outputs act as market substitutes for original works, fair use is less likely to apply. Most major AI systems are trained in the United States, and reliance on fair use has not yet been fully tested in the courts.
What to do now
- Ask AI providers what data was used to train their models and whether it was licensed;
- review internal AI usage policies, particularly in relation to prompt-related risks;
- ensure procurement contracts include indemnities for IP infringement arising from training data; and
- monitor the Getty appeal and UK developments closely.
