South Africa's Department of Communications and Digital Technologies (DCDT) withdrew its draft National Artificial Intelligence Policy in May 2026 after the news outlet News24 found that at least 6 of the document's 67 academic citations did not exist. Editors at the South African Journal of Philosophy, AI & Society, and the Journal of Ethics and Social Philosophy independently confirmed that articles attributed to their pages had never been published. The references, it emerged, had been lifted from a Spanish-language Chilean document translated into English, and ChatGPT had been used to edit the draft. No one verified a single citation before Cabinet approved it on 25 March and the policy was gazetted for public comment on 10 April 2026.
The irony writes itself: a government document meant to govern artificial intelligence was undone by unverified AI output. But the response matters more than the embarrassment, and on that score Pretoria has handled the fallout reasonably well.
What actually happened
Director-General Nonkqubela Jordan-Dyani placed two officials on precautionary suspension and called the episode "highly, highly regrettable." Minister Solly Malatsi was blunter, citing "massive oversight and nondisclosure around the use of AI in the formulation of the policy." Rather than quietly re-paper the document, the DCDT stood up a seven-member National AI Expert Review Panel chaired by Professor Benjamin Rosman of Wits University's Machine Intelligence and Neural Discovery Institute. Its members include Professor Vukosi Marivate (University of Pretoria), Professor Alison Gillwald (Research ICT Africa), competition lawyer Heather Irvine, Dr Tshepo Feela, Dr Jabu Mtsweni of the CSIR, and cyberlaw specialist Lufuno Tshikalange.
The panel is to deliver a consolidated report by August 2026, with a revised draft and discussion paper targeted for late 2026 and a fresh public-comment window opening in January 2027 — effectively pushing a finished national AI framework into 2027.
The steelman for caution
There is a serious case that this delay is the responsible outcome, and it deserves to be stated plainly. A national AI policy sets the terms for how a mid-income economy of 60 million people will handle automated decision-making in credit, policing, hiring, and public services. Shipping that on a foundation of fabricated scholarship would have been far worse than shipping it late. A policy that cannot survive a bibliography check cannot be trusted to weigh the harder trade-offs underneath it. Slowing down to rebuild the evidentiary base is not bureaucratic timidity; it is the minimum due diligence the subject demands.
But the lesson is process integrity, not AI abstinence
Where caution curdles into the wrong lesson is the temptation to treat generative AI itself as the villain. It isn't. The failure here was entirely human and procedural: no citation check, no disclosure of tooling, no editorial sign-off that anyone owned. A junior analyst fabricating footnotes by hand, or copying them unverified from a translated foreign document, would have produced the identical disaster — and that second failure mode is exactly what happened alongside the AI use. The corrective is mundane and well understood: mandatory source verification, disclosure of any AI assistance, and a named accountable reviewer. Those are controls every newsroom and law firm already runs. None of them requires banning the tool.
A proportionate response treats this as a quality-assurance failure to be fixed with checklists, not as proof that AI has no place in government drafting. The opposite reaction — reflexive prohibition — would leave South Africa's public service less capable precisely as it tries to regulate the technology for everyone else.
The harder context: regulating Big Tech while fixing your own house
The timing is awkward because South Africa is, separately, one of the more assertive AI-era regulators in the Global South. On 13 November 2025, the Competition Commission released the final report of its Media and Digital Platforms Market Inquiry, a 24-month probe into how Google, Meta, Microsoft, TikTok and X intermediate South Africans' access to news. The Commission found Google dominant in search — where news represents an estimated 5–10% of queries — yet paying nothing for the journalism it surfaces and increasingly displaces through AI-generated summaries and "zero-click" answers.
The headline remedy: Google committed to a R688-million package over five years for South African media, spanning content licensing, a Digital News Transformation Fund, vernacular-language training via the Media Development and Diversity Agency, and tools to prioritise local sources. Whatever one thinks of state-brokered transfers from platforms to publishers — and we are sceptical of remedies that entrench incumbents rather than open markets — the inquiry was evidence-led, ran a transparent interim-report process, and named its findings.
That contrast is the real story. A government credibly demanding rigour and disclosure from the world's largest platforms cannot afford to be caught publishing hallucinated scholarship in its own flagship document. Regulatory authority is borrowed against institutional credibility, and credibility is spent fast when your AI policy can't pass the check you'd impose on others.
The takeaway
South Africa got the recovery roughly right: suspend, disclose, convene genuine experts, rebuild on verifiable sources. The country should resist the louder lesson that AI tools are the problem. The problem was the absence of verification and ownership — failures that predate ChatGPT and would persist without it. Build the checklist, keep the tool, and finish the policy on a foundation that survives scrutiny.