Civil Rights

South African courts caution lawyers on AI citations

By Kesya Wulandari · · 4 min read
South African courts caution lawyers on AI citations - ai citations
South African judges have begun to warn that AI-generated authorities should be independently verified.

Generative AI tools such as ChatGPT, Claude and Gemini are now routine aids for South African attorneys drafting opinions or advocates preparing heads of argument. The promise is clear: research that once took hours can be produced in seconds.

Court decisions signal caution

South African judges have begun to warn that AI-generated authorities should be independently verified. In Parker v Forsyth NO and Others (Johannesburg Regional Court, unreported case no 1585/20, 29 June 2023), the magistrate criticised reliance on unchecked AI citations.

The High Court in Mavundla v MEC, Department of Co-Operative Government and Traditional Affairs (2025 (3) SA 534 (KZP)) went further, describing the use of AI-produced material without verification as “irresponsible” and incompatible with professional duties.

A later judgment, Northbound Processing (Pty) Ltd v South African Diamond and Precious Metals Regulator and Others (unreported case no 2025/072038, 30 June 2025), found two fictitious authorities in the applicant’s heads of argument. Both counsel admitted the citations came from a subscription-based AI tool marketed as trained on South African law.

In each case the courts emphasized that the duty not to mislead the court applies regardless of the technology used, and they referred the matter to the Legal Practice Council. The rulings stop short of banning AI, but they reaffirm a long-standing principle that ultimate responsibility rests with the practitioner.

Fragmented digital market hampers AI accuracy

South Africa has made progress in digitising legal material. Open-access platforms such as SAFLII and LawLibrary now host many judgments, while court websites and university repositories add to the pool.

Nevertheless, the digitisation effort remains uneven. Case law is scattered across commercial databases and free sites; legislation appears on government portals and subscription services; academic commentary is split among university libraries, journal sites and private publishers.

No single repository contains the entire South African legal corpus. As a result, lawyers often toggle between several databases to assemble a complete picture of the law.

AI systems inherit the same limitations. They can only draw on material that is digitally available and accessible to their training data or retrieval mechanisms. If a judgment, statute or scholarly article is absent from those sources, the AI cannot incorporate it into its output.

Related Post: South Africa’s Takeover Regulation Panel Explained

Materials that are openly online are more likely to shape AI-generated responses than those hidden behind paywalls or poorly digitised archives. Consequently, the visible portion of the law may acquire disproportionate influence in machine-produced summaries.

Historical judgments—especially older High Court decisions and lower-court rulings—are often missing from digital collections. Government gazettes, legislative histories and many specialist commentaries remain difficult to locate in searchable formats.

Because South African law relies heavily on precedent, omitting such authorities can lead AI to produce seemingly persuasive answers that silently exclude critical sources. An absent citation does not prove the authority does not exist; it may simply be outside the system’s reach.

In practice, a lawyer who encounters a “hallucinated” case may quickly spot the error, but the omission of a relevant precedent can go unnoticed, leaving a gap in the argument that is harder to detect.

Limitations of Generative AI for Legal Research and Education

Generative AI systems do not conduct legal research like lawyers. They do not read judgments, evaluate authority or select persuasive precedent; instead they generate responses by identifying patterns within vast digital text. Their output reflects statistical correlations rather than legal judgment.

When lawyers hear “hallucination,” they often picture a technological malfunction. Hallucinations are a predictable consequence of large language model operation. AI-assisted research may encourage lawyers, students and researchers to rely on summaries rather than the law itself. Such hallucinations often mislead users who assume the cited authority exists.

Legal research focuses on locating and interpreting primary authorities, not merely consuming summaries. Students who rely exclusively on AI summaries may never develop skills to identify missing authorities, distinguish primary from secondary sources, or challenge flawed reasoning.

Regulatory guidance and professional standards

Regulators are beginning to issue advisory notes that clarify how AI may be used within the bounds of existing ethical rules. Those notes stress that any citation generated by an algorithm must be cross-checked against an authoritative source before submission to a court.

Practical steps for practitioners

Practitioners can mitigate the risk of unseen omissions by adopting a layered research approach. First, they should run the AI query, then identify each cited authority and retrieve the original document from a trusted repository.

Leave a Reply

Your email address will not be published.