South Africa gazetted a Draft National Artificial Intelligence Policy on 10 April 2026, following Cabinet approval in late March, with written comments due by 10 June 2026. On Sunday 26 April 2026, Communications and Digital Technologies Minister Solly Malatsi withdrew it. The reference list contained fictitious sources.
"This failure is not a mere technical issue but has compromised the integrity and credibility of the draft policy," Malatsi said. He drew the obvious conclusion out loud: "This unacceptable lapse proves why vigilant human oversight over the use of artificial intelligence is critical."
The most plausible explanation, and the one the department has not disputed, is that a generative model produced citations that were never checked before the document went to Cabinet and then to the Government Gazette. A policy intended to govern the technology was undermined by an unverified use of it.
The minister ordered consequence management for those responsible for drafting and quality assurance, and said the framework would be redeveloped before resubmission. Commentary from South African law firms tracking the process expects a revised version to reach Cabinet around November 2026, with public release in early 2027.
The substance of the withdrawn draft was not fringe. It adopted a risk based structure modelled on international frameworks including the EU AI Act, with heavier requirements for high risk systems in healthcare, financial services and law enforcement. That architecture is likely to survive into the next version. What was lost was time, and a measure of credibility at the moment South Africa was positioning itself as the continental reference point for AI governance.
The episode has a practical lesson that generalises well beyond Pretoria. The failure mode was not a model producing an incorrect analysis. It was a model producing something that looked exactly like a verified reference, inside a workflow where nobody was assigned to verify it. Document review processes built for human error do not catch fabricated citations, because human authors rarely invent plausible looking sources at volume.
For any organisation drafting regulated or published material with model assistance, the control is narrow and cheap: every citation gets resolved to a live source before the document moves to the next stage, and the person who resolves it signs for it. South Africa is now the case study for what happens when that step is missing.