A credibility collapse in sixteen days
South Africa's Draft National Artificial Intelligence Policy cleared Cabinet on March 25, 2026, and was gazetted for public comment on April 10 (Government Gazette Notice 3880), opening a comment window that was due to run through June 10. It did not survive to that deadline. On April 26, Communications and Digital Technologies Minister Solly Malatsi withdrew the draft, saying the document's failures were "not a mere technical issue" but had "compromised the integrity and credibility of the draft policy" (SAnews). Cabinet formally ratified the withdrawal on June 5 (SAnews).
The cause was almost too on-the-nose: at least six of the document's 67 academic citations were AI-generated hallucinations — nonexistent journals and studies attributed to real scholars who never wrote them (TheNextWeb). Director-General Nonkqubela Jordan-Dyani later confirmed that "ChatGPT was used in as far as the editing of the document itself," and that the fabricated references traced back to a source document from Chile that had been machine-translated from Spanish before officials folded it into the South African draft. Two officials were placed on precautionary suspension for failing to disclose the tool's use (EWN).
The steelman: this is not a fringe problem
It would be easy to dismiss this as a one-off embarrassment. It isn't. A Nature-published study cited in the same reporting found that 2.6% of academic papers published in 2025 contained potentially hallucinated citations, up from 0.3% in 2024 (TheNextWeb) — a trend serious enough that courts in multiple jurisdictions have already sanctioned lawyers for filing AI-hallucinated case law. A national policy meant to set standards for trustworthy AI failing its own citation check is a legitimate governance failure, not a technicality. Requiring departments to disclose when generative AI tools are used in drafting official documents, and to run independent verification passes before publication, is a proportionate and defensible response — the kind of rule critics of AI regulation should be glad to concede.
Where the response has been proportionate so far
Malatsi's own handling deserves credit on those terms. He pulled the draft within sixteen days of gazettal rather than letting it sit for public comment, and Cabinet backed the reversal without dragging out the political cost. The Independent Expert Review Panel appointed to rebuild the draft, chaired by Wits University's Prof. Benjamin Rosman alongside Prof. Vukosi Marivate, Prof. Alison Gillwald, attorney Heather Irvine, Dr. Tshepo Feela, CSIR's Jabu Mtsweni and cyber lawyer Lufuno Tshikalange, is a credible bench with real technical and legal depth (TechCentral). The new timeline — Cabinet resubmission expected by November 2026, public comment reopening in January 2027 — is a roughly seven-month delay, not an indefinite shelving.
Where the caution should stop
The risk is that the scandal becomes a pretext to make the eventual policy heavier rather than more careful. The withdrawn draft already proposed a dense institutional stack modeled loosely on the EU AI Act's tiered approach: a National AI Commission, an AI Ethics Board, an AI Regulatory Authority, an AI Ombudsperson, a National AI Safety Institute, and a proposed AI Insurance Superfund. None of that architecture was the source of the credibility failure — bad citation hygiene was. Yet a panel rebuilding a policy under reputational pressure has every institutional incentive to over-correct: to add review layers, expand mandates, and multiply new bodies as a hedge against being blamed for the next embarrassment, regardless of whether South Africa's existing regulators — ICASA, the Competition Commission — could absorb equivalent functions without new agencies.
As a widely-read analysis of the episode argued, the real governance lessons are about accountability, transparency and explainability in the drafting process itself, not the substantive scope of what gets regulated (The Conversation, full analysis). A policy that can't verify its own footnotes has no standing to mandate algorithmic audits for others until it fixes that first. The panel's most useful output in January 2027 would be a leaner framework with a transparent, human-verified evidence base — not a heavier one padded with new institutions to compensate for the department's own embarrassment.
The scandal exposed a verification failure inside government, not a flaw in the case for AI regulation. Proportionate regulators should fix the former without over-building the latter.
South Africa still has a genuine policy gap: no binding AI-specific statute exists yet, and the underlying 2024 National AI Policy Framework needs an operational successor. But the fastest way to lose public trust in that eventual framework is to let a citation scandal justify a bigger bureaucracy than the evidence supports.