In a procedurally complex family law dispute about which school minor children should attend, an acting judge appears to have used an AI tool when drafting his judgment without verifying citation accuracy (L v O judgment). The result was that errors occurred in multiple places in the judgment, likely caused by the use of an AI tool without sufficient human oversight.

The matter was subsequently appealed to a Full Bench of the Gauteng Division. On 31 July 2026, the Full Bench (Wright J, Opperman J, and Mahosi J) delivered its judgment (under case number 2025/220239, but currently unreported), effectively agreeing with the merits of all the acting judge’s original decisions.

However, the appeal court confirmed that it found 11 discrepancies, including incorrect citations, muddled case names, fabricated quotations and at least one completely non-existent case in the judgment under appeal, all bearing the hallmarks of AI-generated mistakes. None of these errors were attributable to counsel in the matter.

So all’s well that ends well, then? That seems to be the approach taken by Wright J and Mahosi J. Not so for Opperman J.

The majority and minority approaches

The court was unanimous in confirming the correctness of the acting judge’s decisions on the merits, although the judges arrived at this conclusion differently. They differed, however, on how to address the issue of the citation errors.

The majority (Wright J, with Mahosi J concurring) took a pragmatic approach, holding that the acting judge’s original order was correct on the merits and declining to engage with the AI issues.

Opperman J felt that the citation discrepancies were serious enough to require a fresh reassessment of the application. After then conducting a full independent analysis, Opperman J concluded that the acting judge’s order was justified on the facts.

The practical distinction is this: under the majority approach, a court satisfied with the outcome may confirm it without engaging with AI errors; whereas under Opperman J’s approach, fabricated citations require de novo examination with verified authorities. Both routes led to the same result here but may diverge in less clear-cut cases.

Consequences for the acting judge

Opperman J held that the discrepancies constituted a prima facie case of professional misconduct, not curable simply by avoiding future acting appointments for the particular acting judge. She encountered some difficulty, however, in determining who should be responsible for considering the appropriate disciplinary action (if any). The fact that the presiding officer was an acting judge created jurisdictional questions about who should deal with a complaint about potential professional misconduct under these circumstances: the Judicial Conduct Committee (JCC) under the Judicial Service Commission Act 9 of 1994 (JSC Act) or the Legal Practice Council (LPC) under the Legal Practice Act 28 of 2014?

Opperman J identified a misconduct accountability gap for acting judges: while the JSC Act’s definition of “judge” includes acting judges, the constitutional removal mechanism under section 177 was designed for permanent appointments. No equivalent exists for acting judges, although the JCC has in practice entertained misconduct complaints against acting judges. Opperman J concluded that the LPC is the most effective forum to which the matter should be referred for investigation, because its jurisdiction over enrolled practitioners is continuous. This jurisdiction is not suspended during an acting appointment and survives after the acting appointment lapses. The LPC’s jurisdiction runs concurrently with, not in substitution for, that of the JSC.

Opperman J expressed her intention to forward the L v O judgment to the LPC chairperson for investigation. She opined that the acting judge may have potentially breached the Code of Judicial Conduct, including the requirements of integrity (Article 4(a)), honourable conduct (Article 5(1)), and diligence (Article 9).

Lessons from abroad

While this may be South Africa’s first such judicial incident, it is appropriate to consider whether other jurisdictions have responded to similar cases as Opperman J did in the L v O judgment. We found several examples of AI errors appearing in case outcomes in one way or another (all discovered and corrected), in addition to those referenced by Opperman J. These cases are, however, rarer than instances of legal teams’ misuse of AI.

In the Colombian case of Irma del Socorro Chadid Urueta v Civil, Family and Labour Division of the Superior Court of the Judicial District of Sincelejo (STC17832-2025), the Supreme Court of Justice set aside a ruling of a multi-judge bench after finding that the lower court had relied on quotations from prior decisions which, upon verification, did not exist. The offending quotations appeared to be conveniently and exactly on point for the issue under consideration. Special care should be taken when research using an AI tool presents what appears to be the perfect answer to a legal question. It was certainly a too convenient in this case, and turned out to have been fabricated. Ultimately, the Supreme Court of Justice corrected the error that a multi-judge bench had allowed to creep into their law.

There were no apparent consequences for the judges who made this error, other than that the Supreme Court ordered dissemination of the event to the judiciary to alert other judges to this risk. We could not determine from the judgment where the offending quotes first originated, and it is therefore possible that the litigating parties were responsible for the erroneous quotations, not the judges. It is, however, an instance in which AI hallucinations were not caught by the presiding judges and which could have changed the course of the law in that particular jurisdiction.

Will this happen again?

GenAI is now ubiquitous in legal work, but the well-publicised risks do not seem to have sunk in yet with all practitioners. Judges sit at the pinnacle of the legal hierarchy, and mistakes made at this level can actually alter the course of our jurisprudence. Judges simply cannot afford such errors given the impact on the legal system. If rigorous verification of AI-generated citations (or other AI content for that matter) is important for other legal role-players, it stands to reason that for the judiciary it is an indispensable safeguard for preserving the integrity of the judicial process. It is imperative that those who can pronounce on disputes and impact the development of our jurisprudence exercise the highest level of caution in using AI, so as to preserve public confidence in the administration of justice and ensure that decisions are based on objective, verifiable and legitimate criteria.

While less severe than the Colombian example, the L v O judgment raises legitimate concerns about uncritical AI use in judicial work. Courts worldwide are moving towards governance frameworks mandating verification, transparency and accountability. South African authorities would be well advised to be proactive, drawing on international precedents, rather than waiting for a case where AI errors irreparably prejudice a litigant’s rights.

Encouragingly, steps are clearly already being taken by the South African judiciary. The appeal judgment reveals that the Gauteng Division has a proactive induction programme cautioning acting judges about using AI with extreme circumspection.

At the 2026 South African Judiciary Conference, the judiciary resolved to expedite modernisation of court infrastructure. The Judiciary Administration and IT Steering Committee was tasked with finalising a draft AI policy distinguishing between administrative and adjudicative functions and articulating permissible and prohibited AI uses. The judiciary also committed to ongoing education on responsible AI use, including prompt design and verification of AI outputs. Until local formal guidelines are published, judges may refer to the UK Guidelines for Judicial Officers and UNESCO’s AI guidelines.

The appeal court’s involvement in the L v O judgment ensures that our legal system will not be polluted by doubtful authorities in this instance. However, lasting damage remains a concern: when judges pronounce on matters, their words have binding effect beyond the immediate parties.

The danger of repeating fake citations

Another aspect bears consideration by our judiciary when hallucinatory citations are submitted to court: namely, to what extent will repetition of these erroneous citations in subsequent judgments and commentaries cause additional harm as they become inadvertently incorporated into large language models’ search results?

UK courts have warned against replicating erroneous citations in subsequent judgments. As UK Barristers Matthew Lee and Jim Sturman KC observed: “Judges may be inadvertently exacerbating the issue because those AI-generated inaccuracies are being integrated into the established legal canon indirectly.”

Similarly, in JML Rose (Pty) Ltd v Jorgensen, Mrs Justice Wheatley in Australia adopted an approach of redacting false citations to prevent further AI propagation.

Practical governance comments

Consequences for legal practitioners are likely to harden as it becomes common knowledge that humans will carry responsibility for errors made by AI. It is becoming ever more important for practitioners, at all levels, to be very deliberate about their governance approach.

Practical governance steps include clear GenAI policies; mandatory verification protocols for AI-assisted research and drafting; training on GenAI limitations and risks; and quality assurance processes, including independent citation verification before finalising legal documents. Easier said than done. To get this right requires careful and constant attention. This is not a tick-box exercise that can be met by downloading an off-the-shelf solution. We are going to have to be innovative and attentive, and we will have to recognise this as a strategically important focus area for all lawyers.

The L v O judgment serves as a timely reminder that AI is a powerful tool, not a substitute for professional judgment. As GenAI becomes increasingly embedded in legal practice, the responsibility to verify, validate and exercise independent legal reasoning remains squarely with the human practitioner. The integrity of our legal system depends on it.

https://www.cliffedekkerhofmeyr.com/en/news/publications/2026/South-Africa/Knowledge-Management/Knowledge-Management-alert-11-august-When-AI-hallucinations-reach-the-bench-South-Africa-confronts-AI-in-the-courtroom