The Lawyers Just Rejected the AI Governance Playbook

South Africa's legal profession consulted for a year and then declined the registers, sign-offs and disclosure rules that every consultant is selling.

Sign up for my Substack daily AI newsletter here.

See my AI Training course portfolio for corporate Business Leaders here.

Article link: https://johanosteyn.substack.com/p/the-lawyers-just-rejected-the-ai?r=73gqa&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true

Follow me on LinkedIn: https://www.linkedin.com/in/johanosteyn/

The profession with the most to lose from unverified artificial intelligence has just published its considered position on how to govern it, and the document is remarkable for what it leaves out. No mandatory disclosure to clients. No risk-tier matrices. No artificial intelligence registers, no two-lawyer sign-offs, no compulsory impact assessments. For executives who have spent the past two years assembling exactly that apparatus, the reasoning deserves attention.

CONTEXT AND BACKGROUND

This is not complacency. A public database maintained by a researcher at HEC Paris now records 1,922 court decisions worldwide in which a party relied on hallucinated material, and a court responded, including 28 involving judges. South Africa appears six times. On 31 July 2026, a Full Bench of the Gauteng Division confirmed eleven discrepancies in an acting judge’s written reasons, including muddled case names, fabricated quotations and at least one wholly non-existent case, none of it attributable to counsel. Opperman treated the discrepancies as a prima facie case of professional misconduct and indicated she would refer the judgment to the Legal Practice Council. The profession is writing from inside the problem, not observing it from a distance.

INSIGHT AND ANALYSIS

The revised guidelines, published in the attorneys’ journal after a year of consultation with members of the Bench, senior advocates, legal academics and the Law Society’s own artificial intelligence committee, reject a duty to disclose machine assistance to clients. The reasoning is worth following in substance rather than in outline. A generic notice that artificial intelligence was used communicates nothing that would let a client assess quality. Clients are entitled to results rather than methods. Most damningly, no reliable means exists of establishing that a practitioner used the technology in the face of a denial, so a rule whose breach cannot be established is a gesture rather than a rule, and enforcing it breeds stylistic policing. The drafters were equally firm on procedural machinery, holding that it would dull the value of the technology and impose costs the market for legal services will not bear.

IMPLICATIONS

What survived is more demanding than what was discarded. Every citation confirmed from a primary source. Every quotation verified verbatim. Reliance on foreign law expressly identified as persuasive. Tool selection treated as the practitioner’s responsibility rather than a matter for client consent, with free and consumer-tier services excluded from confidential work. The charging paragraph was rewritten to reject the assumption that the technology saves time at all. Scepticism about vendor assurance is warranted. When Stanford ran the first preregistered evaluation of the closed commercial systems marketed as eliminating hallucinations, it found each of the leading legal research tools producing false information between seventeen and thirty three percent of the time.

CLOSING TAKEAWAY

The distinction the guidelines draw is between responsibility and paperwork, and it is the distinction most corporate artificial intelligence policies quietly avoid. A register records that a tool exists. A sign-off records that somebody looked. Neither establishes that the person whose name is on the document read the underlying source. The board question is uncomfortable in its simplicity: strip out everything in your policy that produces a record without producing verification, and consider what remains. If the answer is very little, the policy was never a control. It was documentation of a control you had not built.

Johan Steyn is a prominent AI thought leader, speaker, and author with a deep understanding of artificial intelligence’s impact on business and society. He is passionate about ethical AI development and its role in shaping a better future. Find out more about Johan’s work at https://www.aiforbusiness.net