Currently, self-driving cars (SDCs) seem to be the golden goal for every automotive manufacturer. Companies such as Google, Tesla, and BMW have already invested millions of dollars in developing a fully autonomous and independent SDC. The operating systems of SDCs are almost solely based on utilising the principle of artificial intelligence (AI).
A great example of an SDC is Tesla's Full Self-Driving (FSD) system, which they offer on a subscription basis to clients. However, the introduction of SDCs to our roads will have its own particular challenges.
This blog will focus on investigating the principles of liability as it relates to SDCs.
Overview of the current landscape
AI is not uncommon in motor vehicles. In fact, it is often used to improve the experience of the driver. Blindspot monitoring, automatic steering, braking, and self-parking functions are all examples of AI systems integrated into motor vehicles.
Tesla is currently the forerunner in the race to produce an effective and autonomous self-driving car that is powered by AI. Tesla currently runs a database where they collect real-world feedback from sensors located on Tesla cars. These databases interpret the data from the sensors and use it to train algorithms by utilising the principles of big data and machine learning. Tesla thus teaches the AI by utilising real-world scenarios, creating a sophisticated, autonomous SDC. Furthermore, Tesla is not the only company incorporating AI into their cars. Google is also in the developing stage of an AI for autonomous cars called Waymo, while BMW has been using over 400 AI applications, including Monolith, since 2022.
Making split decisions
There are several instances where the AI must make decisions in split seconds in SDCs. For example, when to change lanes, when to apply emergency braking, and acceleration at stop signs and traffic lights are decisions that an AI will be confronted with daily. The basis of a successful SDC is AI maturity — the more miles that are collectively driven, the more data can be used to teach the AI how to react in certain situations. The goal is to create an autonomous car that can make better decisions in crucial moments by utilising the pool of knowledge from millions of collective miles driven by users. This can lower occurrences of motor vehicle accidents and prevent thousands of deaths annually all over the globe.
What are the levels of driving automation?
The term "driving automation" has been categorised into levels of autonomy by the United States National Highway Traffic Safety Administration (NHTSA), referred to as SAE levels, categorised below:
- Level 1: An advanced driver assistance system (ADAS) that assists a driver with braking, steering, and acceleration, but does not do so simultaneously.
- Level 2: A car that can steer and either accelerate or brake simultaneously, typically found in adaptive cruise control functions.
- Level 3: Enables a vehicle to drive itself under specific conditions) using cameras, radar, and other systems. While the car handles steering, braking, and monitoring the environment, the driver can engage in other tasks but must be ready to take control when prompted to perform other functions.
- Level 4: A car that can monitor all driving tasks and the driving environment so successfully that the driver does not even need to pay attention.
- Level 5: A completely autonomous vehicle that does not need or require any human intervention at any time. It can be controlled by either typing in prompts or with voice commands to another AI system, which digests information and makes decisions based on this information, without any human intervention after the prompt has been made.
However, this evolution of transport calls a myriad of legal problems to the fore. One of the main problems we face is that of liability. SDC accidents will inevitably occur, even though they will decrease as the AI systems used become more sophisticated.
The question then arises: who is liable for damage caused by an SDC that malfunctions? The answer does not seem to be straightforward.
Common law liability
An AI suitable for motor vehicles will be "built" by coders and IT professionals. This will then be fitted to a motor vehicle, where it will operate or take over some control that the driver usually had to control. If an AI malfunctions and causes a motor vehicle accident, it will be a defective product and will fall under the concept of product liability.
Product liability is a concept prevalent under the law of delict. Damages that flow from defective products are claimable under the actio legis aquilliae, provided that all the other elements of a delict are satisfied. These elements include those of conduct, harm, causation, fault, and wrongfulness. With a claim of product liability, two actions can be followed: A claim in terms of the common law or a claim in terms of the Consumer Protection Act (CPA). A common law claim will be discussed below.
In product liability cases, the elements of wrongfulness and causation are very hard to prove. When assessing the element of wrongfulness, it is necessary to determine what the legal convictions (boni mores) of society are. The boni mores in respect of products is that the manufacturer has the general duty to ensure defective products do not reach the market, withdraw them if he becomes aware that they are defective, and to take steps to ensure that no harm will be suffered should the product be released to the market.
Wrongfulness
Wrongfulness depends on whether a product is unreasonably dangerous. This provides a massive problem for prospective litigants, due to the fact that determining whether an AI is unreasonably dangerous would require expert evidence from coders, IT professionals, and experts in the integration of AI with motor vehicles. Litigation will thus become very costly and drawn-out. Even then, it could be hard for an IT professional to explain to a judge, a layperson (when it comes to coding), where exactly the mistake in the coding led to the faulty decision-making of the AI.
Causation
The element of causation is difficult to prove for the same reason. An AI is written using certain baseline coding, whereafter the AI artificially gathers information and uses that newly (self-acquired) information to drive decision-making. It would be exceptionally difficult to determine causation or establish a causal link between a coder’s work and the decision-making the AI performed independently. For conduct, proving negligence has the same practical difficulties that causation has — it has to be proven that the coder is negligent for the independent action of the AI. Negligence utilises a subjective test for the manufacturer, meaning the court will determine whether the harm caused was reasonably foreseeable and preventable by the manufacturer. Note that this differs from the normal objective "reasonable person" negligence test. It is submitted that the independent functioning of an AI will, as a matter of foreseeability, be very hard to prove, especially in light of information gathering after the initial coding.
As discussed above, a product needs to be defective, and that defect must render the product "unreasonably dangerous" to the average consumer. This is a prerequisite that must be met before a manufacturer can be held liable for the defect. However, some products are inherently or inevitably dangerous. A knife and a gun, for example, are inherently dangerous products. Neethling, in his textbook Law of Delict in South Africa (p. 384) states:
“In determining what qualifies as defective, the state of human science and technology and the need for experimentation must not be lost sight of. Many human products are inevitably dangerous at a certain stage…”
Is an SDC inherently or inevitably dangerous?
The question is whether an SDC is "inherently" or "inevitably" dangerous. In the festive season of 2022, 1,451 people died in fatal car accidents all over South Africa. It is well known that driving is an extremely dangerous activity.
The Seattle Times reports that 120 people die from vehicle-related crashes in the US every day and that it equates to four major airline accidents a week. Driving a motor vehicle is not any safer in South Africa. In fact, there are many external factors that contribute to accidents in South Africa. Poor quality roads and unroadworthy cars are just two factors that may cause more accidents than in a developed country. A court will not be wrong in ruling that driving a motor vehicle is "inherently" dangerous and that an AI driving such a motor vehicle does not remove the inherently dangerous nature of such. However, this does not account for a mistake made by an AI as a result of the poor condition of a road.
If a product is inherently dangerous, the consumer or user of such a product, by default, accepts the inherent risk associated with the product. By accepting the risk, the consumer waives the right for that same risk to be "unreasonably dangerous". For example, if the user of a gun knows that the gun is inherently dangerous (in the sense that it could injure or kill people), but still uses it, then by virtue of his use, the product is not unreasonably dangerous in his mind. The same could possibly apply to the acceptance of the risk of an AI driving on your behalf, with the knowledge that artificial intelligence can (and will) make mistakes. What if a warning about the risk of using an SDC is presented to the consumer? The answer is not clear.
Conclusion
Common law thus does provide for a remedy, but the chances of succeeding are very slim. The need for expert evidence in AI coding, car manufacturing, and any other relevant supply chain experts would drive up the cost of litigation.
In addition, the elements of fault and causation would be very hard to prove, and even one varying opinion from an expert witness may either break the causal link or disprove that the harm was reasonably foreseeable. It would be very hard to prove causation and negligence (fault), which leaves the injured party with an avenue of recourse, but one that is difficult to pursue. However, one could look to the Consumer Protection Act for another avenue of recourse, which will be discussed in a follow-up blog.
About the author
Hein Steenberg completed his BCom Law degree, followed by an LLB and LLM in mercantile law, with a specific focus on space law at the North-West University. Hein is currently pursuing his LLD in space law.
Last updated on 28 April 2026.