Authored By: Queen Thoriso Miraclina Emeana
University of South Africa
I. Introduction
In 2023, a South African hospital began using an artificial intelligence diagnostic system to help their radiologists identify early breast cancer. A patient sued the system because there was no detection of his/her cancer, which incurred the patient serious problems in their health.A patient complained when the system didn’t diagnose a cancerous growth, and therefore he/she was treated late and suffered from serious health issues. The hospital denied liability, the software company relied on the system’s ability to make autonomous decisions and the patient had no effective legal remedy under the current South African tortious regime. This is a global problem that is becoming common in the world of today: laws that were intended to regulate human behaviour and actions are failing to keep up with the world of machine behavior and actions that have real world consequences.
AI systems are being used in South Africa in a wide range of applications that impact on physical safety, financial security and fundamental rights. On the roads of South Africa, autonomous vehicles are being put to the test; algorithmic credit-scoring has been adopted for millions of South Africans to access financial services; and algorithmic diagnostics, based on artificial intelligence, is now commonplace in both public and private healthcare providers. Where these systems inflict harm, the issue of liability in South African law is not easily answerable and is far from settled.
This article contends that, the structure of the existing South African tort law is unsound to deal with liability for harm arising from artificial intelligence systems and that there is a need for legislative changes to create a coherent and just accountabilities regime. The article continues as follows: Part II explores the doctrinal problems arising out of the application of existing South African tort principles, such as negligence, product liability and vicarious liability to AI. In Section III, the comparative approaches of selected jurisdictions are examined. In Section IV a framework for the reform is suggested which is suitable for the South African legal and constitutional context. Recommendations for legislative action are included in Section V.
The current legal framework and its shortcomings.Current legal framework and its weaknesses.
The South African Law of Delict section 195.2
South African tort law is based on the principles of Roman-Dutch law, as developed by judicial decisions and inspired by English law, known as delict law . The basic components of delictual claims are: (i) conduct, (ii) wrongfulness, (iii) fault, (iv) causation and (v) damage. These are all challenging aspects when applied to harm resulting from AI systems that are autonomous.
The wrongfulness inquiry is whether the defendant’s conduct was socially unacceptable, which means recognizing a legal duty which the defendant had to the injured party. In the case of an autonomous AI system, the developer and the deployer may share some of the responsibility for the harm, while the end-user could have some responsibility as well in the specific circumstances – but this is not always clear. The South African Constitutional Court has made the point that wrongfulness is a normative judgement made in accordance with the mores of the society . But there is no clarity yet within the boni mores on standards related to harm caused by AI.
Another sizeable hurdle is the fault requirement. In the law of South Africa, fault includes both intentional and unintentional. A claimant must establish negligence by showing the defendant failed to live up to a reasonable man’s standard of care in the given circumstances. The difficulty is that, when harm is caused by the independent decision making of an AI system, whose internal processes are opaque, it may be very hard to determine which human actors are responsible, and what their actions are in line with what a reasonable person would have done. The lack of a “black box” and the ability to explain to itself the decision it makes undermines the traditional fault-based analysis.
Consumer Protection Act, B.The Consumer Protection Act, B.
The other option for redress is provided by the Consumer Protection Act 68 of 2008, which gives strict liability to manufacturers, producers and importers of harm caused by unsafe or defective goods. Under section 61 of the Act, a producer, importer or distributor or retailer is responsible for any damage resulting from the supply of goods that are defective, unsafe or fail to provide adequate instructions.
The use of AI systems, however, remains unclear in the context of the Act. Goods are defined in the Act as “any tangible object manufactured, produced, processed, assembled, packed or supplied, as well as any intangible product, such as software. This definition could include software that is based on AI, however, the term “defect” in a self-learning system with outputs that change over time is still not settled. The strict liability provisions of the Act also make the assumption that the harm-causing product can be identified, and that there is a causal relationship between the product and the harm. Fulfilling these requirements is a significant challenge in AI systems, where several actors (human and algorithmic) may impact the results.
Vicarious Liability
The doctrine of vicarious liability too, which imposes liability on the employer for the delicts of the employee in the course and scope of the employment, does not seem to be applicable in the case of AI harm. An AI system does not constitute a “workman” or an “agent” in a legally recognised sense. It is not possible to be directed, controlled and disciplined as a human worker. For non-human autonomous systems, it would be a radical extension of the doctrine of vicarious liability, which has been extended to independent contractors in some situations.
III. Comparative Perspectives
The European Union
Among EU member states, the EU has made the most advanced steps towards an AI liability regime. The proposed AI Liability Directive and the recently revised Product Liability Directive can complement each other in filling this accountability gap, by creating a rebuttable presumption of causality if a defendant does not meet required AI obligations, and the non-compliance is plausibly connected to the harm. This way, the claimant does not have to provide overwhelming evidence to establish liability, although it is not a complete departure from the fault-based model. The EU model, which is based on the precautionary principle and the protection of fundamental rights, is an example of a constitutional democracy that could be emulated in South Africa.
The United Kingdom
The UK has taken a more measured stance, primarily putting a trust in the existing tort doctrine, with sector-specific regulation. The Automated Vehicles Act 2024 contains a dedicated liability regime for AVs, with the primary liability that shifts to the entity who approved the use of an AV system on public roads. While this is a strengths-based strategy that has the advantage of specificity, it fails to tackle the issue of AI liability outside of the autonomous vehicle domain. The sector-specific model has been considered for South Africa, as the country already has a basis for having a legal framework for specific sectors.
The United States
In the United States, existing product liability doctrine, at state level, primarily governs AI liability. There are a number of courts that have started to address whether AI outputs are products or services, and the implications for the liability standard. There has been a growing push in academic literature for a strict liability regime for high-risk applications of AI, based on a similar approach to abnormally dangerous activities in the Restatement (Third) of Torts. A more appropriate alternative to the fragmented approach adopted in the USA with its state by state differences is, however, not suitable for South Africa’s centralised legal system.
The proposed reforms in the South African context.
The “Reformers” and Their Programs (1880s-1930s)
In South Africa, the Constitution of the Kingdom of 1996 offers a foundation on which to deal with the issue of liability for AI. The provision of the Bill of Rights in section 38 of the constitution that everyone has the right to approach a court for the enforcement of his rights, is understood to support measures in legislation that provide greater access to justice for persons who are impacted by the use of artificial intelligence systems. In the same way the constitutional principles of human dignity, equality and freedom demand that legal provisions are adapted to technological change affecting these basic rights.
The SALC is tasked to study and propose reform of specific areas of the law. The Commission has conducted in-depth studies on different areas of law before and has expertise to explore the liability of AI in a comprehensive context of technological governance and fundamental rights safeguarding.
A new and innovative approach for South Africa.
Based on the comparative survey in Section III, and applying the constitutional and legal context in South Africa, the article aims to propose a hybrid model of liability in relation to AI.
The South African Parliament must first adopt a specific AI Liability Act, which will define specific categories of high-risk AI applications, such as AI-powered autonomous vehicles, AI-powered medical diagnostic systems and AI-powered automated credit scoring systems, and impose strict liability for any harm caused by the operation of such systems. In these situations, the deployer will typically be the one best able to control the risk and strict liability is the correct approach because the injured party will often not be able to prove fault without access to the proprietary technical information.
Second, there are applications of AI outside the high-risk category where the Act should implement a rebuttable presumption of causality for the claimant to show that they have been harmed by an AI system and the deployer has not met applicable standards. This maintains the same fault-based model, but decreases the evidential burden placed on AI systems by their opacity.
Third, the obligations of developers and deployers should be strengthened, so that the Act places a record-keeping requirement on them to provide records to facilitate post hoc analysis of the AI for harms that occur. Records will be essential in resolving legitimate claims, otherwise, the black box problem will persist, even in a formal liability context.
Fourth, the Act must be in line with the constitutional duty of access to justice. This means provisions for class actions, particularly where AI systems may impact on large numbers of people in similar ways, and mechanisms for collective redress where there is likely to be a large number of people affected by the harms caused by AI.
Conclusion
AI isn’t the technology of the future. It is used today right throughout South Africa in ways which impact directly on the physical safety, financial security and dignity of citizens. As this article has shown, the legal system protecting the accountability for damage wrought by these systems is deficient. The doctrines of the South African law of delict, product liability under the Consumer Protection Act and vicarious liability were not developed with autonomous systems in mind, and their use in the context of AI brings with them the possibility of arbitrariness and sometimes injustice.
This article has proposed a hybrid solution of strict liability for the most dangerous uses of AI and a rebalancing of the burden of proof for other types of AI-related harm with the implementation of disclosure requirements and avenues to legal redress. Experience of the EU and the United Kingdom shows that it is possible to achieve targeted and principled reform while not blocking innovation. As a leading contributor to technological advances in Africa, and as a democratic state with a longstanding tradition of safeguarding its fundamental rights, South Africa has the opportunity, and the obligation, to be at the forefront of technological advances in Africa.
The remedy was due to the patient in the hospital in South Africa. The main challenge of AI liability in the 21st century is to make sure the law exists, is clear, predictable, and fair. The law has to adapt to the ever-changing society, and technology is changing the society too, which warrants appropriate and prompt legal response.
Reference(S):
Cases
Van Breda v Jacobs 1921 AD 330
S v Makwanyane 1995 (3) SA 391 (CC)
Pierce v Hau 1944 AD 175
Jacob Mathew v State of Punjab (2005) 6 SCC 1
Legislation
Constitution of the Republic of South Africa, 1996
Consumer Protection Act 68 of 2008
Automated Vehicles Act 2024 (UK)
Proposed AI Liability Directive, COM(2022) 496 final (EU)
Evidence Amendment Act 45 of 1988
Black Administration Act 38 of 1927
Secondary Sources
Nyathi M ‘Re-asserting the doctrinal legal research methodology in the South African academy: Navigating the maze’ (2023) 140 South African Law Journal 365
Van Eck M ‘An ethical framework for the use of artificial intelligence in the legal profession’ (2025) Tydskrif vir die Suid-Afrikaanse Reg 454
Badenhorst PJ ‘Unlawful and illegal mining in South Africa’ (2025) Tydskrif vir die Suid-Afrikaanse Reg 472
Edwards L and Veale M ‘Slave to the Algorithm? Why a Right to an Explanation is Probably Not the Remedy You are Looking For’ (2017) 16 Duke Law and Technology Review 18
Abbott R The Reasonable Robot: Artificial Intelligence and the Law (Cambridge University Press 2020)
South African Law Reform Commission, Issue Paper on Artificial Intelligence and Liability (forthcoming)
