An ambitious policy, then a quick unraveling
South Africa's Draft National Artificial Intelligence Policy had every ingredient of a serious governance document. Approved by Cabinet on March 25 and April 1, 2026, and gazetted for public comment on April 10, the 86-page framework was meant to move the country past Fourth Industrial Revolution talking points and into actual binding regulation. It was modeled on the EU's AI Act, risk-based, multi-regulator, and built around six strategic pillars covering everything from talent development to cultural preservation.
The institutional architecture it proposed was genuinely interesting. Beyond the expected oversight bodies — a National AI Commission, an AI Regulatory Authority, an AI Ethics Board — the policy included an AI Ombudsperson where citizens could challenge harmful AI decisions, and an AI Insurance Superfund modeled on South Africa's Road Accident Fund. The fund was designed to compensate people harmed by AI systems in cases where liability couldn't be pinned on any single party across the developer-deployer-user chain. That's not a common idea in AI governance anywhere.
Two weeks after publication, the whole thing collapsed.
Six citations that didn't exist
The 86-page draft cited 67 academic sources. On April 24, 2026, Article One — a civic organization focused on participatory governance — sent a formal letter to Minister of Communications and Digital Technologies Solly Malatsi. Executive Director Tyronne McCrindle flagged that at least six of those citations were completely made up. Not misattributed. Not misquoted. Invented.
The next day, News24 called the editors of the actual journals named in the policy — the South African Journal of Philosophy, AI & Society, the Journal of Ethics and Social Philosophy — and asked whether the cited articles existed. They didn't. The editors confirmed the papers had never been published.
Two of the fabricated citations looked like this: Babatunde, O., & Mnguni, P. (2023). "Challenges and Opportunities in Regulating AI: Perspectives from South Africa." AI Policy Journal, 2(3), 143-156. And Burman, A., & Sewpersadh, K. (2022). "Legal Frameworks for AI in South Africa." South African Journal of Philosophy, 41(2), 207-217. Plausible authors. Plausible journals. Plausible volume numbers. None of it real.
The hallucinations clustered around sections on European legal scholarship and regional capacity building — exactly the areas where someone using an LLM to fill in supporting evidence would be most tempted to let the model run.
The withdrawal
Minister Malatsi pulled the document on April 26, two days after Article One's letter. His statement didn't minimize the damage. "This failure is not a mere technical issue but has compromised the integrity and credibility of the draft policy," he said. He acknowledged the likeliest explanation: AI-generated citations included without anyone checking whether they were real.
He also noted the obvious irony. "This unacceptable lapse proves why vigilant human oversight over the use of artificial intelligence is critical. It's a lesson we take with humility." He committed to a full redraft and to "consequence management" for the officials responsible for drafting and quality assurance.
Darren Olivier, a partner at Adams & Adams and head of their AI governance practice, said the swift withdrawal was the right call. His prescription for what should have happened: "Let AI do the heavy lifting as required. Let humans supply judgement as the work takes shape. Then, before anything leaves the building, let humans verify every reference."
Parliament wanted answers
The ANC Study Group on Communications and Digital Technologies called it "one of the most alarming failures of ministerial oversight and intellectual rigour in the recent history of South Africa's digital governance." That's a strong line for a parliamentary statement. ANC MP Shaik Subrathie demanded Malatsi appear before the Portfolio Committee to account for the debacle.
Portfolio Committee chair Khusela Diko had already been publicly calling for the withdrawal before Malatsi acted. The ANC's statement put it plainly: "We cannot have an AI policy drafted by AI to regulate AI. It's a cycle of doom."
The concern wasn't just embarrassment. A policy that touches manufacturing, energy, and the digital economy had apparently been built, at least in part, by outsourcing the intellectual groundwork to a generative AI tool that no one went back to verify.
It wasn't just the AI department
Here's what made this harder to dismiss as a one-off: the same problem showed up in another department at the same time. News24 found over 100 AI-fabricated references in the Cabinet-approved Revised White Paper on Citizenship, Immigration and Refugee Protection from the Department of Home Affairs. Two senior Home Affairs officials were also suspended. The department argued those fake references were in a standalone bibliography and didn't affect the policy text — which is a defense, though not a convincing one.
Two Cabinet-approved documents, two departments, fabricated citations in both. That's not a rogue employee. That's a workflow problem.
Leon Schreiber, the Democratic Alliance Coordinator in the National Executive, announced that ministers would require explicit AI verification checks before approving policy documents going forward — across the entire government.
The part that's hard to get past
The policy's own text stressed that responsibility for AI-assisted decisions sits with the deploying organization, and that meaningful human oversight is non-negotiable. The document said that. Then the document itself skipped both of those things.
AI models don't flag uncertainty when they invent citations. They produce confident, well-formatted references to papers that sound exactly like real scholarship. The problem isn't that someone used AI to help draft a policy document — that's increasingly normal. The problem is that no one checked the output before it went to Cabinet, got gazetted, and was released for public comment.
Tyronne McCrindle of Article One put the broader stakes plainly: a national policy governing transformative technology has to be grounded in real evidence, real public engagement, and real accountability — not the path-of-least-resistance output of an unverified model. The breach wasn't just procedural. It's the kind of thing that makes people stop trusting institutions.
What this actually costs
South Africa scores 53.94 on the Oxford Insights AI Readiness Index, well above the Sub-Saharan African average of 29.12. Other African governments watch how South Africa handles digital governance. The country was positioned to be a reference point for the continent on AI regulation.
The DCDT had planned to finalize the policy by late 2026 and begin implementing binding sector-specific regulations in the 2027/2028 financial year. The redraft, re-consultation, and resubmission process pushes that back by up to a year. Businesses and citizens are in regulatory limbo in the meantime.
That's the real cost — not the embarrassment, which passes, but the delay. Whatever framework eventually emerges will carry this episode with it and will be read more skeptically for it. That scrutiny might not be a bad thing. The revised policy will need to demonstrate the oversight the first draft claimed to require.
