CAN AN AI SYSTEM PERFORM “CONDUCT” FOR PURPOSES OF THE SOUTH AFRICAN LAW OF DELICT?

Introduction

The first recorded pedestrian accident caused by an autonomous vehicle was in March 2018, where a self-driving Uber vehicle killed Elaine Herzberg as she crossed a road in Arizona. This moment forced the legal world to confront the question: when a machine causes harm, who is responsible in law? South Africa has not yet witnessed such an incident as autonomous vehicles are not yet popular, and fully automated systems have not yet been widely deployed in different industries. However, as artificial intelligence penetrates the South African economy, in financial services, healthcare, transport, and beyond; the prospect of harm caused by autonomous artificial intelligence (AI hereafter) is no longer a distant hypothetical but an approaching reality. When that harm arrives, courts will turn to the law of delict for an answer. This article therefore addresses the question of whether it constitutes conduct, for the purposes of the South African law of delict, when AI is used to perform an action, or when AI acts entirely autonomously.

The Conduct Element in the South African Law of Delict

A delict is a wrongful and culpable act that causes harm. For delictual liability to arise, the harm suffered by one person must have been caused by the conduct of another. Conduct may be defined as a voluntary human act or omission. Mere thoughts, however harmful in intention, do not carry delictual consequences until they are manifested in an act.

Human Conduct

The conduct has to be performed by a human being, and usually the defendant has to be a natural person. However, juristic persons may also be sued in a delict through the acts of its organs. Where an animal or an object is used in the commission of a delict, a human act is still present.

AI used as an instrument

Where a human being uses an AI system as an instrument to perform an act, the requirement of human conduct is satisfied. This is similar to when a driver is operating a vehicle or a person directs an animal. The instrument is merely a mechanism of harm, and the human’s will behind directing it remains the legally relevant human act. In Jooste v Minister of Police and Chetty v Minister of Police, the courts confirmed that where a human uses an animal as an instrument, a human act is still present. Similarly, in the case of using AI as an instrument, the human being remains the operating actor. Though AI may be a sophisticated tool, the human decision to deploy it, receive its output and act upon it is what constitutes conduct in the delictual sense. Therefore, the requirement of human conduct is satisfied.

Autonomous AI Systems

This position is fundamentally different where an AI system acts autonomously without the direction of a human. Autonomous AI systems make decisions based on their own processing of inputs without any human instructions. No human being directs the specific act that causes harm. One might thus be tempted to put blame on the developer of the AI system. In the case of autonomous AI systems, the developer of the system writes the underlying code in advance and does not necessarily direct the specific decision that caused harm. The deployer activating the system also does not necessarily imply that the system is being directed to deliver outputs. In a truly autonomous AI system, there is no active human role at the moment of harm, therefore no human to attribute the conduct to.

The question then arises whether the AI system itself can be regarded as performing conduct. An AI system is neither a natural person nor a juristic person in South African law. It has no legal standing, cannot bear rights or duties and cannot act as a defendant in court proceedings. One might thus be tempted to attribute the acts of an AI system to a juristic person. However, an AI system is not a human organ of a juristic body, therefore its actions cannot be attributed to a juristic person. If this were possible, an important question would then arise where it would have to be determined if the acts of the AI system have to be attributed to the company that developed the AI system or the company that currently uses the AI systems. Conclusively, the requirement of human conduct fails in the case of autonomous AI systems and thus it can be said there is no conduct present.

Voluntary Conduct 

Conduct is voluntary when it is subject to the actor’s will and control. This means that the actor must be able to direct muscular activity or to prevent such activity. Voluntariness does not mean that the person must have desired the conduct. A defence of automatism may be raised when someone claims they did not act voluntarily. The defence of automatism occurs when a defendant asserts that he or she behaved involuntary or mechanically. Conduct may be considered involuntary in the following instances: compulsion, reflex muscular movements and unconscious state. However the defence of automatism will not succeed when there is impulsive or spontaneous acts, intentional prior conduct and negligent prior conduct.

AI used as an instrument

Where a human being uses AI system as an object, the voluntariness requirement is satisfied. This is because the human operator makes a conscious decision to deploy the AI system. The human actor could have chosen not to deploy the AI system. In the case of Van Wyk v Lewis, where the surgeon’s decision to perform the operation using the instruments at his disposal was itself a voluntary act, and it was in the exercise of that voluntary conduct that the negligence arose. The same logic applies where a professional deploys an AI tool in the performance of their duties. The sophistication of the instrument does not diminish the voluntariness of the human decision to use it. The conduct is therefore voluntary in the legal sense, and this requirement is satisfied.

Autonomous AI Systems

The voluntariness analysis in the case of autonomous AI is more complex. Where an AI system acts without real-time human direction, the specific act that causes harm is not susceptible to any human will at the moment it occurs. The developer’s will was exercised when the system was designed, and the deployer’s will was exercised when the system was activated, but neither directed the specific harmful output.

One might argue, similar to the automatism cases, that the prior voluntary acts of the developer and deployer are sufficient to ground liability, in the same way that a person who negligently creates the conditions for their own automatic state remains liable for the harm that follows. This principle, known as actio libera in causa, was affirmed in S v Chretien, where the Appellate Division recognised that a defendant cannot escape liability simply because the immediate act was involuntary, if that involuntary state was itself the product of an earlier voluntary and negligent act. In the case of automatism, prior voluntary acts are considered only where they negligently or intentionally created the conditions for the automatic state.² 

In the case of autonomous AI, this means that a developer’s or deployer’s prior conduct can only give rise to liability where it amounts to legally recognisable negligence, for example where the developer failed to adequately test the system before release, or where the deployer introduced the system into a context where harm was foreseeable. This is consistent with Wessels v Hall and Pickles (Coastal) (Pty) Ltd, where the court imposed liability on a diabetic driver who suffered a hypoglycaemic attack while driving, on the basis that he was aware of his condition and the precautions required, yet failed to take them before driving. Similarly, a developer or deployer who was aware of the risks posed by an autonomous system and failed to take reasonable precautions may similarly be held liable on the basis of negligent prior conduct. However, even accepting this, it remains a question of fault rather than conduct.

Furthermore, one cannot attribute voluntariness to the AI system itself. Voluntariness requires the capacity to exercise conscious control over one’s conduct. This capacity is traditionally seen as a distinctly human capacity. An AI system processes data and generates outputs through algorithmic pattern-matching and probabilistic inference. Regardless of its complexity, this process cannot be equated to the exercise of will in the legal sense. The outputs of an autonomous AI system are therefore a result of algorithmic computation rather than the voluntary exercise of human will. The voluntariness requirement therefore fails when an autonomous AI system acts.

In Molefe v Mahaeng, the Appellate Division confirmed that defences based on automatism must be scrutinised with great care. The court further held that the onus remains on the plaintiff to prove that the defendant’s conduct was voluntary. This is significant in the AI context because where harm is caused by an autonomous system, the plaintiff faces the considerable evidentiary burden of identifying a voluntary human act, which, as demonstrated above, may be impossible to discharge.

Commission and Omission

Conduct may take the form of a positive act or a failure to act. Liability for omissions is generally more restricted than liability for positive acts. For policy reasons, the law is hesitant to impose a legal duty to act positively, and an omission is not wrongful unless such a duty can be established. There are however some cases where it may be difficult to distinguish whether an omission or commission occurred. This is especially true in circumstances involving a continuous course of conduct, where the same action may be both a commission and an omission depending on the context. Many omissions are merely indications of legally deficient positive conduct, and the mere fact that conduct can be described in negative terms does not make it an omission in law.

AI used as an instrument

Where an AI system is used by a human as a tool, the commission or omission is that of the human who operates it. Thus, the ordinary principles of conduct as discussed earlier apply. This means that a person who uses an AI tool negligently, by failing to exercise adequate oversight of its output commits a positive act of conduct. Where an AI system produces an inaccurate output, the question arises whether such conduct is a commission or an omission. One may see this as an omission by the deployer, specifically a failure to take necessary measures against foreseeable harm from the autonomous system. According to Minister of Safety and Security v Carmichele, liability of omissions arises where a legal duty to act exists. One could claim that the deployer of an AI system owes such a duty, and that failure to avert foreseeable harm constitutes a wrongful act. This argument has merit in cases where the harm was predictable and the deployer took no measures. It does not, however, always address the conduct issue. In situations where the deployer took appropriate safety measures and the technology malfunctioned in a way that was unexpected, there is no human omission in this particular instance. Thus, the conduct element remains unsatisfied.

Autonomous AI Systems

In the case of autonomous AI systems, the commission or omission analysis is more nuanced. Where an autonomous AI system causes harm, one might first ask whether such conduct amounts to a commission. In a descriptive sense, the autonomous AI system did act, by producing an output. However, characterising this as a commission for the purposes of delictual liability encounters the obstacles identified above. As established, an autonomous AI system is neither a human being nor a voluntary actor. The positive act would thus be belonging to the machine, and the machine is not a legal subject capable of committing a delict. The commission analysis therefore fails.

In terms of omission, if the autonomous AI system fails, one might claim that the developer had a legal obligation to build a safe system, and their failure to do so is the relevant omission. In situations where the harm was easily predictable, the argument has some merit. However, a basic challenge in the context of autonomous AI is that it is characterised by its ability to make judgements that no human could have predicted or stopped. Thus it becomes extremely challenging to determine what action the developer should have taken to ensure that the system does not fail.

Recommendations

This article reveals a gap in the South African law of delict. Where an autonomous AI system causes harm without any human direction, the conduct element fails entirely, leaving the injured plaintiff without a delictual remedy. Since autonomous AI satisfies none of the existing requirements of conduct, and since no human actor can be identified as performing the legally relevant act, interpretation of existing doctrine cannot adequately bridge this gap. It is therefore submitted that a statutory strict liability framework is the only adequate solution. Such a framework would bypass the conduct element entirely, imposing liability on a designated party, whether the developer, the manufacturer, or the deployer at the time of activation, without requiring proof of voluntary human conduct. This is the only mechanism that closes the accountability gap without distorting existing doctrine beyond recognition.

Conclusion

In conclusion, this article has examined whether an AI system can perform conduct for purposes of the South African law of delict. Where AI is used as a tool under human control, the conduct element is satisfied since the human operator’s decision to deploy the system and act on its output is a voluntary human act. Where AI functions autonomously, the conduct factor fails completely. No human being committed the exact act that produced harm, the AI system lacks legal personality and human will, and neither the commission nor omission analysis gives an adequate remedy. Because no human actor can be identified as committing the legally relevant act, the injured plaintiff is left without redress under present South African law. This gap cannot be resolved through judicial interpretation alone, and legislative intervention in the form of a strict liability framework is therefore necessary.

BIBLIOGRAPHY

Primary Sources

Legislation

Road Accident Fund Act 56 of 1996

Cases

Chetty v Minister of Police 1976 (2) SA 450 (N)

Jooste v Minister of Police 1975 (1) SA 349 (E)

Minister of Safety and Security v Carmichele 2004 (3) SA 305 (SCA)

Minister of Safety and Security v Van Duivenboden 2002 (6) SA 431 (SCA)

Molefe v Mahaeng 1999 (1) SA 562 (SCA)

S v Chretien 1981 (1) SA 1097 (A)

Van Wyk v Lewis 1924 AD 438

Wessels v Hall and Pickles (Coastal) (Pty) Ltd 1985 (4) SA 153 (C)

Secondary Sources

Books

Loubser M and Midgley R (eds), The Law of Delict in South Africa (3rd edn, Oxford University Press 2017)

Neethling J, Potgieter JM and Visser PJ, Law of Delict (8th edn, LexisNexis 2020)

Pasquale F, The Black Box Society (Harvard University Press 2015)

Van der Walt AJ and Midgley GF, Principles of Delict (4th edn, LexisNexis 2016)

Journal Articles

Calo R, ‘Robotics and the Lessons of Cyberlaw’ (2015) 103 California Law Review 513

Scott TJ and Visser SJ, ‘Artificial Intelligence and Legal Personality’ (2021) 84 Tydskrif vir Hedendaagse Romeins-Hollandse Reg 1

Official Reports

National Transportation Safety Board, Collision Between Vehicle Controlled by Developmental Automated Driving System and Pedestrian, Tempe, Arizona, March 18, 2018 (NTSB/HAR-19/03, 2019)

[1] National Transportation Safety Board ‘Collision Between Vehicle Controlled by Developmental Automated Driving System and Pedestrian, Tempe, Arizona, March 18, 2018’ (NTSB Highway Accident Report NTSB/HAR-19/03, 2019).

[2] M Loubser and R Midgley (eds) The Law of Delict in South Africa (3rd edn, Oxford University Press 2017) 7.

[3] J Neethling, JM Potgieter and PJ Visser Law of Delict (8th edn, LexisNexis 2020) 27.

[4] J Neethling, JM Potgieter and PJ Visser Law of Delict (8th edn, LexisNexis 2020) 27.

[5] M Loubser and R Midgley (eds) The Law of Delict in South Africa (3rd edn, Oxford University Press 2017) 93.

[6] M Loubser and R Midgley (eds) The Law of Delict in South Africa (3rd edn, Oxford University Press 2017) 95.

[7] M Loubser and R Midgley (eds) The Law of Delict in South Africa (3rd edn, Oxford University Press 2017) 95.

[8] J Neethling, JM Potgieter and PJ Visser Law of Delict (8th edn, LexisNexis 2020) 28.

[9] Jooste v Minister of Police 1975 (1) SA 349 (E); Chetty v Minister of Police 1976 (2) SA 450 (N).

[10] Ryan Calo ‘Robotics and the Lessons of Cyberlaw’ (2015) 103 California Law Review 513, 530.

[11] Ryan Calo ‘Robotics and the Lessons of Cyberlaw’ (2015) 103 California Law Review 513, 533.

[12] M Loubser and R Midgley (eds) The Law of Delict in South Africa (3rd edn, Oxford University Press 2017) 95.

[13] M Loubser and R Midgley (eds) The Law of Delict in South Africa (3rd edn, Oxford University Press 2017) 95.

[14] J Neethling, JM Potgieter and PJ Visser Law of Delict (8th edn, LexisNexis 2020) 30.

[15] J Neethling, JM Potgieter and PJ Visser Law of Delict (8th edn, LexisNexis 2020) 31.

[16] M Loubser and R Midgley (eds) The Law of Delict in South Africa (3rd edn, Oxford University Press 2017) 96.

[17] M Loubser and R Midgley (eds) The Law of Delict in South Africa (3rd edn, Oxford University Press 2017) 97.

[18] M Loubser and R Midgley (eds) The Law of Delict in South Africa (3rd edn, Oxford University Press 2017) 98.

[19] Van Wyk v Lewis 1924 AD 438.

[20] J Neethling, JM Potgieter and PJ Visser Law of Delict (8th edn, LexisNexis 2020) 33.

[21] S v Chretien 1981 (1) SA 1097 (A) 1104.

[22] Wessels v Hall and Pickles (Coastal) (Pty) Ltd 1985 (4) SA 153 (C).

[23] J Neethling, JM Potgieter and PJ Visser Law of Delict (8th edn, LexisNexis 2020) 32-33.

[24] Frank Pasquale The Black Box Society (Harvard University Press 2015) 3.

[25] Molefe v Mahaeng 1999 (1) SA 562 (SCA) 568.

[26] Molefe v Mahaeng 1999 (1) SA 562 (SCA) 568.

[27] J Neethling, JM Potgieter and PJ Visser Law of Delict (8th edn, LexisNexis 2020) 34.

[28] J Neethling, JM Potgieter and PJ Visser Law of Delict (8th edn, LexisNexis 2020) 36.

[29] J Neethling, JM Potgieter and PJ Visser Law of Delict (8th edn, LexisNexis 2020) 36.

[30] AJ van der Walt and GF Midgley Principles of Delict (4th edn, LexisNexis 2016) 92.

[31] AJ van der Walt and GF Midgley Principles of Delict (4th edn, LexisNexis 2016) 92.

[32] Minister of Safety and Security v Carmichele 2004 (3) SA 305 (SCA).