On 26 April 2026, South Africa's Minister of Communications and Digital Technologies, Solly Malatsi, withdrew the country's final draft National AI Policy just sixteen days after it was gazetted. The reason was extraordinary: the document meant to govern artificial intelligence had itself been compromised by it. News24 revealed that several of the draft's 67 references were fabricated — fictitious journal articles and authors credited with work they never produced. Editors at the South African Journal of Philosophy, AI & Society, and the Journal of Ethics and Social Philosophy independently confirmed the cited pieces had never appeared in their pages.
The minister did not equivocate. In a government statement, Malatsi conceded that "AI-generated citations were included without proper verification," and that "this failure is not a mere technical issue but has compromised the integrity and credibility of the draft policy." Two officials were placed on precautionary suspension. A seven-member expert panel, chaired by Wits University's Benjamin Rosman, was appointed to rebuild the evidence base. A revised draft is now expected to reach Cabinet by November 2026 and open for public comment only in January 2027 — leaving Africa's most industrialised economy without a formal AI framework for the better part of another year.
The case for moving fast — and why it backfired
It is worth steelmanning the impulse that produced this document. South Africa had real reasons to hurry. The continent is at risk of becoming a pure consumer of AI built elsewhere, and a national policy signals to investors, courts, and citizens that the state intends to set terms rather than inherit them. A coherent framework on data, liability, and public-sector deployment is a genuine public good, and the officials who pushed for one were not wrong to want it gazetted before the window narrowed. Ambition was not the failure here.
The failure was process. Sovereignty over a technology is not asserted by publishing a notice in the Government Gazette; it is earned through the institutional capacity to produce, verify, and defend the rules you write. A drafting team that fed prompts into a generative model and published the output without checking a single citation did not demonstrate command of AI — it demonstrated dependence on it, of exactly the unexamined kind a national policy is supposed to discipline. The irony is total, but the lesson is structural: a state that cannot quality-assure a policy document is not yet equipped to regulate the systems that policy targets.
Sovereignty is a capability, not a declaration
This is the distinction that should reshape how African governments approach AI. The most credible sovereignty efforts being mounted right now are resourced, not merely proclaimed. On 3 June 2026, the European Commission unveiled a tech sovereignty package — a Chips Act 2.0, a Cloud and AI Development Act with tiered sovereignty assurance levels, and an open-source strategy — built around a hard number: the EU relies on foreign suppliers for more than 80% of its key digital products. Brussels responded to dependence with industrial policy and law, backed by budgets and timelines. Whatever one thinks of CADA's protectionist edges, it is a plan to build something, not a press release.
Africa's own builders are already showing what capability-led sovereignty looks like from the ground up. As Rest of World documented this month, compute scarcity is driving genuinely local AI infrastructure across India, Brazil, the UAE, and Africa — stacks engineered to work around the concentration that lets Amazon, Microsoft, and Google account for nearly two-thirds of global cloud spending. Nigerian fintech Daya's stablecoin settlement corridor linking African and Middle Eastern markets is another instance of the same instinct: don't wait for permission, build the rails. These are sovereignty in its functional sense — control accrued through what you can actually do.
The proportionate path back
The right response to the South Africa episode is not regulatory retreat, and certainly not a moratorium on AI use inside government. The fix is proportionate: disclosure rules for AI-assisted drafting, mandatory human verification of any factual claim or citation in a public document, and procurement that funds real research capacity rather than outsourcing it to an unverified chatbot. The expert panel is the correct mechanism — independent, named, accountable — and its appointment suggests Pretoria has learned the right lesson rather than the reflexive one.
The deeper takeaway for the continent is that a good AI policy will be slower and duller than a gazette deadline rewards. It will rest on verifiable evidence, domestic expertise, and institutions that can be held to account when they cut corners. South Africa lost a year and some credibility. The cheaper way for everyone else to learn the same lesson is to treat digital sovereignty as a muscle to be built, not a flag to be planted — and to remember that the first test of whether you can govern AI is whether you can resist letting it govern you.