South Africa's Draft National AI Plan: Not Good Enough

South Africa's government published its Draft National Artificial Intelligence Plan ("the Draft AI Plan") discussion document yesterday (it does not seem to be available on the DCDT's website today). It is made up of three parts:

• An introduction that gives an overview of the AI landscape and how it is perceived from a South African perspective;
• A second section that considers how the South African landscape has been impacted by the pitfalls and opportunities of applied and generative AI, the Presidential Commission on the 4th Industrial Revolution's ("PC4IR") approach to AI and discusses localisation of AI; and
• The third section outlines South Africa's AI plan, priorities and key enablers and discusses AI governance and institutional mechanisms.

The Draft AI Plan is 53 pages long and starts with a disclaimer that it is "for the general discussion purpose and not for publication or academic use." (sic) I think it is fair to say that the disclaimer is correct and it should not have been published in its current form.

This may sound harsh, but if you read the document, you will understand why. Here's a copy I marked up: yellow highlights are good points, green underlining means that I didn't really understand what was intended, or that there were errors or unfinished points.

The Draft AI Plan is disappointing because:

• it is lengthy, does not really flow and contains way too much jargon;
• it is riddled with errors and, at times, seems to have unfinished thoughts;
• it centres South Africa's plans around the Artificial Intelligence Institute of South Africa, an organisation set up in November 2022, but which, from its website, has not given any updates since March 2023;
• it sets out a plan and timeline for AI adoption but its focus areas are very high-level; it does not attribute responsibility for achieving deliverables to any specific person or government department; it discusses undefined concepts such as "Centralised Computing Power" and "AI Hyper-scale Data and Large Language Model Centre" and sets unrealistic deadlines for their achievement; and discusses AI Regulation but sets unrealistic targets for achieving these regulations (I have yet to see a new and complex piece of South African legislation complete the legislative drafting process in 12 months);
• while it does seem to hint at some sort of data sovereignty, to cater for the huge amounts of data South Africa will need to train its AI, it makes no mention of South Africa's Draft National Data and Cloud Policy, a policy that is critical to facilitating the meaningful use of AI in South Africa; and
• does not discuss the challenges that may be faced in procuring the computing power South Africa will require, to effectively implement much of what it seeks to achieve in the Draft AI Plan.

The document is convoluted, complicated, lacks clear deliverables and fails to allocate responsibility for their delivery. It needs to be more practical with crystal clear deliverables and no doubt about who is responsible for delivering what.

For an excellent example of a similar document that leaves very little room for misinterpreting what deliverables are required and who is responsible for delivering them, consider Rwanda's National AI Policy. It is a forthright, practical document, uncluttered by highfalutin and confusing jargon.

My view is that, rather than stakeholders commenting on this highly flawed Draft AI Plan, it should be completely reworked and released when it is in a more practical and improved form. The Draft AI Plan must align with the practical and clear objectives the Minister of Communications and Digital Technologies outlined in his keynote address at yesterday's National Artificial Intelligence Summit. The clock is ticking and South Africa will be left behind if it does not act speedily.

Lucien Pierce,

6 April 2024