South Africa's AI Policy Learns Hard Lesson: Never Let the Robots Grade Their Own Homework

There’s a particular flavor of irony baked into South Africa’s national AI policy disaster. The government tasked itself with drafting a visionary blueprint for artificial intelligence governance—and then discovered the document had been partly written by the very technology it was supposed to regulate.

Communications Minister Solly Malatsi withdrew the country’s AI policy last month after News24 uncovered a troubling problem: at least 6 of the document’s 67 academic citations were complete fabrications, references to journal articles that never existed. The culprit? AI hallucinations. Tools trained to sound authoritative had simply invented scholarly sources—from the South African Journal of Philosophy, AI&Society, and the Journal of Ethics and Social Philosophy—without anyone catching the fraud before the policy went public.

Malatsi didn’t mince words. This wasn’t a glitch, he insisted. This wasn’t a minor copy-editing miss. This was a fundamental breach of credibility. As he put it,“This failure is not a mere technical issue but has compromised the integrity and credibility of the draft policy.”Translation: we blew it. The government handed over critical policy work to a machine, failed to verify the output with actual human eyes, and embarrassed itself in front of the world.

The National Artificial Intelligence Policy was meant to position South Africa as a continental leader in AI innovation while grappling with the thorny ethical, social, and economic challenges the technology raises. Instead, it became Exhibit A in why that leadership requires something AI can’t provide: judgment. The policy has been yanked for revision, and Malatsi promised“consequences”for those responsible—a pointed reminder that delegating without due diligence has a cost.

The real lesson here transcends South Africa’s stumble. As institutions worldwide rush to adopt AI for everything from policy drafting to legal briefs to academic research, this case study in failure carries weight. You can’t automate away the need for human scrutiny. And when you’re writing rules about AI itself, the stakes of getting it wrong aren’t just embarrassing—they’re deeply corrosive to public trust.