South Africa Restarts Its AI Policy: Can It Rebuild Trust After the Hallucination Scandal?

South Africa wants artificial intelligence to support economic growth, create jobs and expand access to digital skills. But its attempt to establish a national framework for AI governance encountered an uncomfortable obstacle: the government’s own draft policy contained fictitious references attributed to research that could not be verified.

The document was withdrawn. Two officials were suspended over their undisclosed use of AI during its preparation, and one has since resigned, according to reporting by TechCabal. The disciplinary process involving the remaining official was still underway in the publication’s 15 October account. Communications Minister Solly Malatsi now says the government is working towards finalising a replacement policy through Cabinet by the end of March 2027, with public consultation expected to follow in April.

The episode has turned a policy-development exercise into a test of public-sector AI governance. The question is no longer only what South Africa should regulate. It is how the country can establish a credible framework after the process intended to define responsible AI use failed to verify the evidence supporting its own recommendations.

The restart offers an opportunity to address both questions, but a new timetable alone will not restore confidence.

The original draft had appeared to make substantial progress. Cabinet approved its publication for public comment on 25 March 2026, and the government gazetted it in April, inviting submissions by 10 June. The proposed framework sought to establish national principles for the development, adoption and governance of AI.

Then the references became the problem.

The draft included fictitious academic sources in its bibliography. After the issue was exposed, Malatsi announced its withdrawal on 26 April, saying an internal review had confirmed that the document contained fabricated references. Cabinet subsequently approved the withdrawal in June, allowing the department to rework the policy and establish national standards for ethical AI use.

The failure was not simply that an AI tool had been used. AI can assist with translation, editing, summarisation and research support. The critical issue was that generated or AI-assisted material entered a national policy document without adequate verification, while the use of AI had not been properly disclosed.

That distinction matters for public institutions. An AI system can generate a plausible-looking citation, but plausibility is not evidence. A source must exist, support the claim attributed to it and be relevant to the policy decision being made. Human reviewers must verify those conditions rather than assume that polished language or an extensive bibliography signals reliability.

The incident also raised a broader governance question: if internal review did not identify fictitious references before publication, what other quality-control procedures failed, and what will change before the next draft reaches the public?

South Africa’s Parliament treated the matter as a substantive accountability issue. In May, the Portfolio Committee on Communications and Digital Technologies called the department to account for the withdrawal, the implications for policy development and the handling of those responsible. Parliament emphasised the importance of credible, transparent and accountable policymaking.

That scrutiny is important because a national AI policy is not an ordinary administrative document. It may influence future regulatory institutions, public-sector procurement, investment incentives, research priorities and the rights of people affected by automated systems.

The government’s timetable has shifted as the process has evolved.

Earlier reporting in May described a target for a revised draft to reach Cabinet in November 2026, followed by public consultation expected around January 2027. In October, Malatsi gave TechCabal a later timetable: Cabinet finalisation by the end of March 2027, with public consultation expected to begin in April.

The October update is the latest of these reported timelines. It should therefore guide the current account, rather than the earlier dates being repeated as if they remain the operative schedule.

The revision is being led by an expert panel chaired by Professor Benjamin Rosman of the University of the Witwatersrand’s Machine Intelligence and Neural Discovery Institute. The panel was appointed after the withdrawal to help produce a replacement grounded in stronger evidence and South African priorities. Its membership includes expertise in AI research, law, governance and related fields.

The later timetable creates room for substantive review, but it also extends the period before South Africa has a completed national AI policy. That is not the same as saying the country has no laws relevant to AI: existing legal and regulatory frameworks can still apply to matters such as privacy, consumer protection, discrimination and sector-specific conduct. The gap is the absence of a completed national AI policy providing a consolidated direction for responsible development and adoption.

The process now has two distinct milestones to deliver. First, government must complete a draft that can withstand internal scrutiny and Cabinet consideration. Second, it must open a credible public process through which affected communities, researchers, businesses and civil society can challenge assumptions and propose changes.

Those stages should not be conflated. A draft reaching Cabinet is not the same as a policy being adopted, and a policy being adopted is not the same as its implementation being funded or completed.

The withdrawn draft was designed to do more than define ethical principles. It proposed a national approach to AI development and deployment, combining economic ambition with governance and public-interest safeguards.

The government’s stated framework was organised around six pillars:

• Capacity and talent development: expanding the skills required to develop, evaluate and use AI.

• Inclusive growth and job creation: encouraging AI adoption in ways that support economic participation.

• Responsible governance: establishing arrangements for oversight and accountable deployment.

• Ethical and inclusive AI: addressing risks to people and ensuring that benefits are not confined to a narrow group.

• Cultural preservation and international integration: reflecting South African priorities while engaging with global AI developments.

• Human-centred deployment: recognising that AI systems should serve people and that deployment risks vary across sectors.

These priorities remain useful as a framework for evaluating the rewrite, but they should not automatically be treated as confirmed provisions of the future policy. The replacement may retain, revise or remove elements of the withdrawn document.

The original draft also proposed institutional and economic measures, including new AI governance bodies, investment incentives and support for infrastructure. Those proposals were reported in the draft’s coverage, but their final form and status must be checked against the replacement text when it becomes available.

This is where the rewrite becomes economically consequential. A policy that focuses exclusively on risk could leave businesses uncertain about how government intends to support adoption. A policy that prioritises investment without sufficiently addressing accountability could leave people exposed to decisions they cannot understand or challenge.

South Africa’s challenge is to make the relationship between these objectives explicit: where innovation should be encouraged, where safeguards are necessary, which institutions will be responsible and how those institutions will be held accountable.

The fabricated-citation episode has a direct lesson for AI governance: the use of AI in government needs rules, but those rules must cover the entire process, not just the final document.

Public institutions increasingly face pressure to use AI to reduce administrative burdens, process information and improve service delivery. These applications can be valuable. But the reliability of the outcome depends on how the technology is used, the quality of the underlying information, the review process and the consequences of errors.

In policy development, an incorrect reference can distort the apparent weight of evidence behind a recommendation. In other public-sector settings, an unverified AI output could misstate a legal requirement, misdirect a service request or introduce an unsupported assumption into a decision affecting an individual.

The appropriate response is not necessarily to prohibit AI assistance. It is to distinguish between assistance and authority.

An AI system may help officials organise evidence or produce an initial draft. It should not become the final authority on whether a source exists, whether a legal claim is accurate or whether a policy recommendation is justified.

A credible government process should make clear:

• when and for what purposes AI tools may be used;

• who remains accountable for the final output;

• how citations, factual claims and legal references are verified;

• what records of AI-assisted work must be retained;

• how errors are identified, corrected and reported; and

• what level of independent review is required for high-impact decisions.

These safeguards should be proportionate to the consequences of the work. A routine formatting task does not carry the same risk as a legal analysis or a national policy that may shape future regulation.

The lesson from the withdrawn draft is therefore institutional. Responsible AI use requires procedures that can detect failure before it reaches the public, not merely explanations after it does.

South Africa is seeking to position itself as a major participant in Africa’s AI economy. Its policy choices will influence how public institutions adopt AI, how businesses assess regulatory risk and how researchers and entrepreneurs engage with the national innovation system.

For businesses, uncertainty around the eventual framework can complicate long-term planning. Organisations developing AI products need to understand which obligations may apply, how high-risk uses will be treated and which authorities will oversee compliance. Investors also need to distinguish between announced policy ambitions and concrete measures supported by institutions, budgets and implementation plans.

For start-ups and smaller businesses, the design of compliance obligations matters particularly. Requirements that are unclear, fragmented or costly to interpret can create barriers to entry. Conversely, safeguards that are clear and proportionate can help establish trust among customers, partners and public-sector buyers.

The original draft’s reported proposals for incentives and infrastructure support also raise a practical question: how will the government translate national ambition into delivery? A statement of intent does not itself provide computing capacity, finance research, train workers or make public services ready to deploy AI.

The replacement should therefore be judged not only on the principles it declares but also on the mechanisms it proposes. These include responsibilities across government, implementation sequencing, support for local research and innovation, public-sector capability, and a realistic account of resources.

The same applies to the distribution of benefits. A national strategy should explain how South African businesses and workers can participate in the AI economy, rather than treating adoption as an objective in itself. Skills, infrastructure, language, affordability and access to markets will shape who can benefit.

A policy can encourage innovation while setting boundaries around unacceptable risks. The difficult work is specifying those boundaries in a way that is understandable to institutions and businesses—and enforceable in practice.

The replacement process provides an opportunity to involve the people and organisations likely to be affected by AI deployment.

Businesses and technology developers can explain technical constraints and the costs of compliance. Researchers can assess the strength of the evidence behind proposed interventions. Labour organisations can raise questions about skills, work and changes to employment. Civil-society groups can identify risks involving privacy, discrimination, access to services and redress.

The public sector itself needs a voice in the process. A policy may assign responsibilities to regulators and departments, but those bodies need the staff, expertise, systems and funding to fulfil them.

Consultation is especially important when decisions concern which AI applications should be encouraged, which require additional safeguards and how people can challenge harmful outcomes. These are not merely technical questions. They involve choices about risk, rights, economic opportunity and institutional responsibility.

The government should publish a clear account of how consultation will work, which parts of the draft remain open to change and how submissions will influence the final document. A summary of the issues raised and the reasons for accepting or rejecting major recommendations, would help make the process more transparent.

Given the history of the first draft, the evidence supporting the replacement also deserves special attention. Major claims should be traceable to genuine, relevant sources, while assumptions and uncertainties should be identified rather than concealed behind confident prose.

The replacement policy should be assessed against several practical tests.

Evidence and verification. Can readers trace major factual claims to real sources? Are legal references accurate, and are research findings represented in context? Is there a documented quality-assurance process?

Clear responsibilities. Does the policy identify which institutions will lead implementation, how their roles relate to existing authorities and where accountability will sit?

Proportionate safeguards. Does it explain how risks will be handled across different uses of AI without assuming that every application requires the same response?

Economic delivery. Are skills development, research, infrastructure and support for local businesses connected to specific actions, realistic timelines and identifiable resources?

Public participation and redress. Can people affected by AI systems understand their rights, raise concerns and seek appropriate remedies?

Transparent implementation. Will government report progress against measurable objectives, explain delays and update the public when the framework changes?

These tests would allow South Africans to assess the policy on its substance, rather than judging it solely by the controversy surrounding its predecessor.

South Africa’s AI-policy process has suffered a setback, but the withdrawal also created an opportunity to revisit the framework before it becomes the basis for future decisions.

The next draft must address two related responsibilities. It must set out how AI should be developed and used in the country, and it must demonstrate the standards of evidence, transparency and accountability that responsible AI governance requires.

The government’s latest timetable points to Cabinet finalisation by the end of March 2027, followed by public consultation expected in April. Those are targets, not completed outcomes. The substance of the replacement—and the process through which it is reviewed—will determine whether the restart produces a durable national framework.

South Africa does not need a policy that merely sounds sophisticated about artificial intelligence. It needs one that can be verified, implemented and challenged; one that supports innovation without treating economic ambition as a substitute for safeguards; and one that recognises that AI governance depends as much on the institutions using the technology as on the technology itself.

The most important measure of success will not be how quickly the government publishes another document. It will be whether the next one gives citizens, businesses and public institutions a credible basis for making decisions about AI.