That Zimbabwe has a far better AI policy than South Africa -– and that SA still has nothing approaching a plan – is far worse than the intellectual dishonesty behind two instances of hallucination-contaminated government policies.
That two different South African government departments based key policies on “proof” that turned out to be LLM hallucinations is, you know, not ideal.
But if you think that is the biggest insult to governance, the democratic process and your intellect, then I submit that you have not in fact read the Draft South Africa National Artificial Intelligence Policy.
Because the fake references in that document aren't even in the top 10 biggest problems with it.
To avoid (entirely understandable) confusion, the AI policy is not the policy that saw two officials suspended on Friday after hallucinated references were found in the document. That's a whole different thing, the White Paper on Citizenship, Immigration and Refugee Protection. The AI policy is the one that saw two officials suspended on Thursday because hallucinated references were found in the document.
Now you might be tempted to think that the biggest insult lies in the combination of those two instances, in the implication that an entire layer of senior technocrats is literally faking it.
Granted, that is top-10 stuff, but it should be downranked because that is something we knew before.
Whereas the AI policy offers a whole host of brand-new horrors.
Choosing the worst of the worst insults in the AI policy is a bit like choosing between the firing squad and the rope, but I lean towards the fact that Zimbabwe did it better.
Nobody in Cabinet seems to have noticed that the AI policy does not contain any policy.
Instead of turning to some stone-age LLM for help, Zimbabwe brought in Unesco to deliver a pretty clear-headed analysis of where it is and where it needs to be going on AI. Zimbabwe's policy document recognises its few strengths (such as a high literacy rate) alongside its disadvantages (such as its never-ending brain drain), then does a creditable job of trying to figure out how to pull off a leapfrog while mitigating the risk of digital colonialism.
Zimbabwe's policy sets actual KPIs and deadlines for stuff that is achievable. It weaves ubuntu throughout the document, giving practical expression to that philosophy.
Whereas South Africa's policy says in a throwaway line that ubuntu should be a guide in all things AI – and then never mentions it again. Much in the same way that South Africa's AI policy explains how intergenerational equity is a central pillar – yes, to AI policy – and then utterly ignores that concept for 24 000 agonising words.
I would not judge you harshly, though, if you were more insulted by the fact that South Africa's AI policy calls for the creation of six brand new regulatory-type bodies while simultaneously assigning the work of regulating AI to existing regulators such as ICASA.
Because that sounds made up, the six are: an AI Ethics Board, a National AI Commission, an AI Regulatory Authority, an AI Ombud, an AI Insurance Superfund and a National AI Safety Institute.
That doesn't include a possible but not (yet) actually recommended AI accreditation authority to accredit AI developers, AI organisations and, let's not forget, AI end-users.
But wait, there's more! South Africa's policy also envisages layering a new bureaucracy on top of existing bureaucracy, by way of an Integrated AI-Powered Monitoring Centre that is supposed to help increase efficiency not only in every government sector but also throughout society, in some way that is never defined.
That's some pretty insulting stuff.
Is it as insulting, though, as the nonsensical word salad sprinkled through the document? After whittling down the long list, this is my favourite sentence: “Development of shared supercomputing centres such as AI giga-factories facilitates startups and researchers and promotes sector-specific regulation compliance.”
It is hard to ignore the insult inherent in that sentence passing through an entire formal process that included Cabinet approval.
Then again, nobody in Cabinet seems to have noticed that the AI policy does not contain any policy. At some points it suggests that the best policy is to not have a policy, by presenting every possible policy approach and saying all must be implemented equally. At other points, the policy is to make a policy at some unspecified point in the future.
Therein lies what amounts to a sort of meta-insult, the umbrella insult that stretches across this whole episode of government policy poisoned by AI hallucinations: South Africa still has no AI policy. Not because the LLM-infected AI policy has now been withdrawn, but because it never amounted to a policy in the first place.
After supposedly working on one since August 2024 – way back when “AI agents” were just wrappers around chatbots – there's nothing to see. Government has made not a single decision of consequence, has made no plan of substance, has not so much as settled on a broad philosophical approach that could lead to the development of a policy.
Compared to that, including hallucinated references doesn't even register.
Share
