AI Agents: Who is liable when things go wrong?

Although South Africa does not yet have AI specific legislation, existing legal principles already provide a clear framework for understanding liability.

South African law does not treat AI agents as legal persons. This means an AI agent cannot be sued or held legally responsible for what it produces. Liability instead rests with the persons or organisations behind the agent, those who develop it, deploy it, supervise it, or rely on its output. This applies whether the agent was built using proprietary software or open source software.

Open source software can make accountability more challenging because it is often difficult to identify the original programmers. Contributors can be located anywhere in the world and may participate anonymously. If the harmful behaviour comes from this earlier code, it may be impossible to find the person responsible. However, South African law does not require every contributor to be identified. Instead, it focuses on the party with actual control over how the AI agent is built, customised, and deployed. Usually, this will be the organisation that assembled the software, integrated it into its operations, and enabled others to rely on it. As a result, the deploying business is typically the party most likely to be held accountable.

Problems become more serious when an AI agent produces hallucinations (false or fabricated statements presented as fact). Where such statements are defamatory, the matter is handled through ordinary delictual principles, not special AI legislation. Courts will examine whether a person or business acted reasonably in allowing the harmful information to be created or shared. Key considerations include whether the risk of false statements was foreseeable, whether the organisation had proper safeguards in place, and whether anyone published or relied on the information without verifying it. Given the well-known risk of AI hallucinations, failing to check accuracy may readily be regarded as negligent.

It is not a defence to argue that the harmful statement came from AI or that the agent was built using open source software. South African law is concerned about who allowed the harmful statement to reach the public, not where in the software the error started. Even an end user who repeats defamatory AI generated content may face liability if it was unreasonable to rely on the information.

Many platforms attempt to rely on disclaimers to distance themselves from AI generated content, but these offer only limited protection. While disclaimers can help parties allocate risk between themselves, they cannot remove their legal duty to avoid reasonably foreseeable harm. South African courts generally do not permit companies to escape fundamental legal obligations through fine print, and disclaimers cannot override delictual principles.

Ultimately, the central question is not whether the AI agent acted independently, but who was in the best position to prevent the harm. Although open source software may make it difficult to identify original contributors, it does not remove legal responsibility. The organisation that deploys the AI agent typically has the most control and oversight and is therefore the party most likely to be held liable when something goes wrong.