South Africa’s draft national policy on AI is expected to be published later this month. The draft policy is an opportunity to articulate a coherent national position on the governance of AI as adoption accelerates across sectors and typical usage of AI increasingly shifts from discrete tools that assist with narrow tasks to end‑to‑end systems in which AI is embedded throughout workflows and decision‑making.
As a starting point, it is important to be clear about what the draft policy is intended to do and how that differs from the National AI Policy Framework that was published in 2024. The policy framework was, by design, a high-level statement of principles. The draft policy should go further in setting out an overarching policy position and providing an intelligible basis for more targeted instruments, including sectoral strategies, ethical guidelines and other measures.
That distinction matters because it throws into sharp relief both the opportunity and the risks associated with the draft policy. If it remains too high-level it risks being treated as aspirational rather than practical (and indeed duplicative of the policy framework).
Conversely, if it is overly prescriptive it may constrain innovation and investment before South Africa has developed the infrastructure, institutional capacity and domestic ecosystem required to deploy AI systems responsibly and at scale. If it is delayed, it risks seeking to influence behaviour after market practice has already solidified.
The critical question is whether the draft policy can timeously shape how AI is adopted and governed as it becomes ubiquitous across different sectors.
Inclusive growth and job creation
In a presentation to the portfolio committee on communications & digital technologies, the department of communications & digital technologies described operational priorities likely to inform the draft policy, including building skills and talent, using AI to support inclusive growth and job creation (including through start-ups and small, medium and micro enterprises), introducing safeguards for privacy and accountability, promoting ethical and inclusive AI (including bias mitigation) and emphasising human-centred deployment through oversight, transparency and explainability.
Whether those priorities manifest as more than aspirational will depend on how clearly the draft policy allocates responsibility, sequences implementation and outlines the need for supporting instruments. It is one thing to state that AI should be trustworthy, fair and human-centred; it is another to identify who is responsible for oversight, what standards or guidelines will be developed, and how the policy will be translated into enforceable measures in due course.
It appears unlikely that South Africa will simply seek to replicate regulatory models emerging elsewhere. The EU has adopted a horizontal, risk-based approach that classifies AI systems by risk levels and imposes obligations on developers and deployers accordingly. The UK, by contrast, has favoured a “pro-innovation”, principles-based approach, permitting existing regulators to implement suitable measures within their remits.
A hybrid approach may be the most realistic fit for South Africa, with high-level principles coupled with targeted, sector-specific measures where risks and impacts are concentrated.
Early signals point to a deliberate “middle path” approach to AI regulation in South Africa, anchored in constitutional values and development objectives. In a briefing with the committee, the department rejected the “false choice” between not regulating and over-regulating, indicating an intention to balance innovation and growth with safeguards grounded in public interest- and rights-based concerns.
A hybrid approach may be the most realistic fit for South Africa, with high-level principles coupled with targeted, sector-specific measures where risks and impacts are concentrated. A highly prescriptive, one-size-fits-all regime would be difficult to implement in a system marked by fragmented regulatory responsibilities and uneven capacity.
If the draft policy addresses opportunities relating to developing skills, supporting start-ups, investing in inclusive datasets and implementing risk-proportionate governance, this could deepen participation in the digital economy, support innovation and create clearer expectations for responsible development and deployment.
The draft policy will need to speak directly to high-impact use cases, particularly in employment and law-enforcement contexts, where bias, opacity in decisionmaking and poor oversight can translate into concrete harms. In those domains, fairness requires representative data, meaningful transparency, trained human oversight and clear mechanisms for contesting outcomes.
Indigenous language capability
Inclusion may need to extend to meaningful support for indigenous languages and locally relevant datasets. Many widely used AI systems are optimised for English speakers and without deliberate investment in indigenous language capability AI adoption risks excluding large parts of the population from the benefits of AI-enabled services and undermining objectives of transformation, inclusiveness and cultural sovereignty.
Employment effects will perhaps be the most politically sensitive issue to be addressed in the draft policy, particularly in the context of a labour market already under strain. Concerns about job displacement and whether AI-driven efficiencies will translate into broader employment opportunities or simply accelerate labour shedding have surfaced in the department’s engagements with the committee. Its position appears to be that an appropriate policy response lies in skills development and reskilling. That is a defensible position, but it places a clear burden on the draft policy to articulate a pathway for a “just transition”.
Implementation capacity is a challenging issue for the draft policy’s prospects of realising its objectives. This requires co-ordination across regulators and adequate resourcing of oversight functions. By way of example, regulatory sandboxes can be valuable test environments in which developers trial systems under supervision before wider deployment, but they only work where there is a responsible custodian with capacity, agreed entry criteria and workable safeguards.
Most importantly, the policy must be clear about what it expects now from regulators, developers, deployers and end-users, and what is to be deliberately left for sector-specific regulation. If the draft policy can provide clear guidance on the objectives it wishes to be pursued, the mechanisms it anticipates being employed to achieve such objectives and how such mechanisms are to operate in the South African context, it may well set the stage for practical, predictable and proportionate regulation of AI in the future.
• Rachwal is an associate at Herbert Smith Freehills Kramer South Africa.
Would you like to comment on this article?
Sign up (it's quick and free) or sign in now.
Please read our Comment Policy before commenting.