South Africa Let AI Write Its AI Policy. It Did Not Go Well.

This has to be the most embarrassing story in African tech this year—and honestly, it’s one of the most absurd moments in global AI policy so far. South Africa tried to lead the way by drafting its own national AI law, but they ended up being the first country on the planet to have their AI legislation completely sabotaged by the very tech they were trying to regulate.

Here’s what happened: In April, the South African government published their Draft National Artificial Intelligence Policy. Eighty-six pages, all official-looking. But within weeks, they yanked it offline. Why? Researchers realized the policy was filled with references to academic papers and journals that don’t exist at all. We’re talking fabricated citations. Just flat-out hallucinated sources—apparently courtesy of the very AI tools the officials used to write the thing.

The fallout was fast. Communications Minister Solly Malatsi quietly withdrew the document on April 26. Four days later, two senior officials from the Department of Communications and Digital Technologies got suspended. And it didn’t end there. Investigators then dug into a White Paper on Citizenship and Immigration—boom, another hundred or so “references” that cannot be traced. Now, there’s a massive audit of every government policy paper written since ChatGPT showed up in 2022.

So yes—South Africa is the first country to have its shot at AI policy sabotaged by AI itself.

But the bigger point is, this isn’t just South Africa’s mess. It’s a wake-up call for every government racing to slap together flashy AI policies to look modern, without doing the real work. Generative AI is dazzling at sounding smart while being flat-out wrong. If you let these tools write the paper instead of supporting human experts, you’re asking for trouble.

What stings even more? South Africa isn’t struggling to keep up—they’re actually one of the most AI-ready countries in Africa according to the Global AI Readiness Index. They’ve got serious institutional know-how. And if it happened there, anywhere’s fair game. Here’s the takeaway: AI can help with policy, sure, but it shouldn’t be the one writing it. Every fact, every source, every number—humans need to check it all before it goes out into the world. The tech is powerful enough now to be both convincing and completely wrong at the same time. Governments need to sharpen their editorial instincts, and fast.