The Department of Communications and Digital Communications (“DCDT”) has withdrawn the Draft South Africa National Artificial Intelligence Policy (“Draft Policy”). The withdrawal of the Draft Policy is a consequence of the citation of fake research, authoritative sources, and academic studies. These fake citations are likely to have been created by Artificial Intelligence (“AI”) resources itself.
This is unfortunately not the first occasion where we have seen the use of fake citations. Over the last eighteen months there have been two high-profile cases in which lawyers have been found to have cited cases which have been falsely created by AI. In Mavundla v MEC Department of Co-operative Government and Traditional Affairs[1] and Northbound Processing (Pty) Ltd v South African Diamond and Precious Metals Regulator and Others[2] it was found that the legal representatives had referred to cases which do not exist and were created by AI resources. These misrepresentations on the part of AI resources have been dubbed as “AI hallucinations”, and we have now seen possibly the most high-profile example of such hallucinations- in the Draft Policy.
Ironically, the Draft Policy advocates for the ethical use of AI, and yet, references in the Draft Policy have likely been drawn up by AI. These AI hallucinations have brought about many questions on the Draft Policy, with a serious concern being how much of the Draft Policy was actually created by AI.
Despite the AI hallucinations, the Draft Policy is a step in the right direction. South Africa’s AI industry and use have gone largely unregulated up until this point, relying on existing policies and legislation for regulation. The Draft Policy has put forth many useful and innovative ideas which could form a solid foundation for South Africa’s National AI Policy. The Draft Policy, however, is not faultless, and will rely on public comment and input for its effective implementation. The Draft Policy should not be disregarded due to the AI hallucinations which we have observed. The Draft Policy has a set a solid foundation for the creation of a National AI Policy in the future. It has its pros and cons, but this does not mean that the Draft Policy should be torn apart and restarted from the ground up. South Africa requires AI regulation sooner rather than later and must act quickly to not be left behind.
The Pros
The Draft Policy embraces sector-specific approaches and includes many broad goals and aspirations which is set to drive South Africa’s AI governance. These will inform sector-specific policies which will be drafted through public and private sector consultation and collaboration. This will allow each sector to determine and regulate the opportunities, risks, concerns and development of AI as it relates to that sector’s practice area. The Draft Policy therefore makes provision for a National AI Policy that will not have to regulate all sectors which incorporate AI in their daily operations. This allows the subsequent AI Policy to determine more streamlined goals and objectives to be adhered to and not have broad regulations which could leave space for ambiguity.
The Draft Policy places a major emphasis on AI education and ethics, and how these two aspects work together. AI is set to be incorporated in all levels of education, with AI training also being included in mobile training centres and community centres to broaden accessibility. At the centre of this education, on top of the hard skills to be taught concerning AI, is AI ethics and the social impacts which AI can have on society. A broad education on the social aspects of AI is an important step towards the effective use of AI in South Africa and in Africa as a whole. The Draft Policy has used the concept of ubuntu to drive its ethical framework. “I am because we are” is a uniquely South African concept which has the potential to set South Africa’s ethical guidance and education apart from the rest of the world’s own interpretation of AI ethics.
This human-centred approach is aligned with a risk mitigation step found in the Draft Policy which requires human input through every step of AI development with the goal of eliminating autonomous decision making and reducing the possibility of AI models “going rogue” as we have seen globally in recent times. The Draft Policy focuses on incorporated human control into key decision-making process of AI, using the method of Human-in-the-Loop (“HITL”). The strategy employs reinforcement learning with human feedback (“RLHF”) so that AI systems are aligned with ethical requirements and societal norms. Early and continued education on AI can lead to South Africa becoming a pioneer in the field of AI going forward, and the ethics-based approach will mean that those educated on AI will develop AI applications and procedures with societal needs and goals in mind.
Although a major concern with the use of AI applications is where and how personal information is captured, stored, and processed, the Draft Policy emphasises that it is set to comply with POPIA’s eight conditions for processing of information strictly, alluding to stringent privacy regulations in line with POPIA.
The Cons
Despite these positives, the Draft Policy does raise some questions and concerns. These concerns include the lack of measurable targets, ambiguity as to who will lead sectoral policies, energy infrastructure, digital infrastructure, the number of institutions to be created, lack of incentives, and the use of international entities as system providers.
The Draft Policy refers to many aspirations which it commendably seeks to achieve but sets no timelines and targets to be achieved by different sectors upon implementation of a final National AI Policy. The National AI Policy ought to set a roadmap for the development of AI systems and for the achievement of the goals set out in the Draft Policy. Without such measurable goals and timelines, the Draft Policy will remain rudderless.
Whilst the sectoral approach, and co-operation between the public and private sectors is favourable for the development of AI policies, the methods by which each sector’s policies will be created is not yet clear. It would be logical for the relevant government departments to create the various policies, however, reference to institutions such as ICASA, the SARB, and the Information Regulator create questions as to who will lead the development and implementation of the various sectoral policies. If or example, government departments are required to create and implement policies, the National Treasury may be required to create and implement several sectoral AI policies, which may lead to slow turnarounds and overburdening the department.
Another issue is that South Africa has gone through an energy crisis over the last several years. Rolling-blackouts have been common in the country. Recently, we have seen a great improvement in the reliability of our energy infrastructure, however, this remains a concern. To implement the goals and objectives set out in the Draft Policy, South Africa requires a stable energy grid. AI relies on electricity to survive. Without a stable, constant energy supply, the goals set out in the Draft Policy cannot be achieved. Therefore, South Africa must invest heavily in its energy infrastructure. Should the country fall back into regular loadshedding schedules, the development of AI in the country will be impaired. Consequently, the AI Policy needs to make further provision for energy infrastructure to reliably supply electricity for the development of AI in South Africa to progress.
The Draft Policy explicitly states that there must be investment in South Africa’s digital infrastructure to prepare us for 5G usage around the country, as well as the roll-out of 6G in the future. However, before we can look at expanding our infrastructure, there must be an effort made to upgrade and improve our current digital infrastructure. Digital infrastructure does not stretch country-wide, and connectivity issues in rural areas remain an issue. This does not align with the Draft Policy’s goal of accessibility for all.
The Draft Policy envisages the creation of seven separate institutions to govern AI in South Africa. That means seven entities that require funding, training and guidelines from government. Furthermore, this requires budget allocations for each new institution. Many of the institutions also appear to have overlaps in what they will oversee and regulate. They could easily be incorporated into fewer institutions. The greatest issue around having this number of institutions regulating AI is that there will not be a central authority outside the Department of Telecommunications and Digital Technology. This creates the potential for a lot of in-fighting and showmanship in the battle for which institution is of more significance.
The Draft Policy makes provision for incentives such as tax breaks and grants for those who incorporate AI into their operations, but it does not specify mechanisms on how to go about this. There is no clear mechanism which would encourage a company to invest in South Africa’s AI future. The Draft Policy encourages industry’s input in education, training and ethics however, it does not specify a mechanism to encourage such input. The National AI Policy ought to incorporate a mechanism which will pave the way for the ideas put forth in the Draft Policy.
Lastly, the Draft Policy relies heavily on international companies and organisations for AI applications, systems, and technologies. The Draft Policy acknowledges the reliance placed on America and China’s AI technology and infrastructure but does little to rectify this reliance. Instead, it continues to rely on, and explicitly mentions, the use of infrastructure, AI applications, and AI hardware from other countries and international organisations. This reliance on foreign infrastructure poses a risk to South Africa’s data security and protection of personal information. By continuously using technology from outside of South Africa, we run the risk of losing people’s personal information and data to foreign governments and organisations.
South Africa’s Draft Policy is a step in the right direction. It is paving the way for a comprehensive AI Policy and possibly AI legislation in the future. A draft policy is never going to be perfect, which is why it is made available for public comment. The Draft Policy holds its own positives and negatives which can hopefully be ironed out and improved on to give South Africa a National AI Policy which is groundbreaking and can hold its own on a global scale. Although there is plenty of room for improvement, appreciation must be shown for what the Draft Policy has done correctly. We will have to wait and see whether the public’s feedback is taken on board for the final AI Policy to provide comprehensive guidance on a matter of the utmost importance not just in South Africa, but globally.
The creation of the Draft Policy and its false references is a serious warning regarding AI use. Within the legal field, we have already seen serious consequences to those who have used AI in their drafting processes. DCDT’s use of AI is a reminder that AI is not perfect, and human oversight is an absolute necessity. The use of false references, lying in wait in the footnotes and unobserved by many, has the potential to create lasting negative impacts through policy and judgments. However, the Draft Policy itself should not be relegated to the footnotes of attempts made to regulate AI. The Draft Policy ought to be briefly revised and published once more for public comments. It is the government’s, and legal practitioners’, responsibility to use AI ethically, effectively and in a manner which does not create a future which is governed by AI itself.
[1] Mavundla v MEC Department of Co-operative Government and Traditional Affairs and Others (7940 2024P) [2025] ZAKZPHC 2 (8 January 2025).
[2] Northbound Processing (Pty) Ltd v South African Diamond and Precious Metals Regulator and Others [2025] ZAGPJHC 528 (30 June 2025).
