The Discovery of Fictitious References in South Africa's Draft AI Policy

The Department of Communications and Digital Technologies (DCDT) released South Africa's first Draft National Artificial Intelligence Policy on April 10, 2026, for public comment. Spanning 86 pages, the document outlined ambitious plans to position the country as an AI leader on the continent, with proposals for new institutions like a National AI Commission and an AI Ethics Board. However, just two weeks later, a News24 investigation uncovered at least six fictitious academic references in the bibliography, sparking widespread outrage and leading to its swift withdrawal.

Journalists verified that citations pointed to non-existent journals, articles, and authors. For instance, one referenced 'Babatunde, O., & Mnguni, P. (2023). Challenges and Opportunities in Regulating AI: Perspectives from South Africa' in the 'AI Policy Journal,' a publication that does not exist. Another cited 'El Hadi, M. M. (2023). Artificial Intelligence Background, Definitions, Challenges and Benefits' from a dubious Egyptian journal link, repeated excessively for basic definitions. Experts immediately identified these as hallmarks of AI hallucinations—plausible but fabricated outputs from large language models like ChatGPT when queried for references.

Minister Solly Malatsi's Response and Policy Withdrawal

On April 26, Communications Minister Solly Malatsi announced the withdrawal, admitting the lapse compromised the policy's credibility. 'The Department of Communications and Digital Technologies did not deliver on the standard that is acceptable for an institution entrusted with such a critical mandate,' he stated. Malatsi pledged a full review of the drafting process, consequence management for those responsible, and a redraft with stricter human oversight to prevent future AI misuse.

Activist Zackie Achmat had alerted the minister days earlier, sharing evidence of the fakes. Parliamentary committee chair Khusela Sangoni-Diko echoed calls for accountability, urging avoidance of tools like ChatGPT in revisions. The incident drew political backlash, with opposition parties questioning DCDT's competence amid broader digital transformation efforts.

Understanding AI Hallucinations: A Technical Breakdown

AI hallucinations occur when generative models produce confident but incorrect information, often inventing citations to fill gaps in training data. In this case, the pattern—non-existent journals mimicking real ones like 'Journal of African Law'—matches known LLM behaviors. Experts note rates up to 27% in reference generation without verification, underscoring why tools must be paired with fact-checking.

For step-by-step: (1) User prompts AI for policy research; (2) Model generates text with plausible but unchecked refs; (3) Humans incorporate without validating; (4) Document proceeds to review, missing errors. This 'ironically' undermined an AI ethics policy emphasizing transparency and accountability.

Key Elements of the Withdrawn Policy and Its Higher Education Ties

The policy's six strategic pillars included capacity building through AI-integrated curricula in universities, research hubs, and talent pipelines. It proposed grants for AI startups from campuses and ethics training for academics. South African universities like the University of Cape Town (UCT) and University of the Witwatersrand (Wits), with established AI labs, stood to benefit from proposed funding. Yet the scandal highlights a disconnect: while unis advance AI responsibly, government drafting faltered.

Aimed at phased rollout—guidelines in 2025/26, regulations by 2027—the document aligned with global standards like OECD principles but now requires overhaul.

Photo by Joshua Hoehne on Unsplash

Reactions from South African Academia and Research Community

Higher education leaders expressed concern over eroded trust in policy processes. Wits AI researcher Dr. [fictional based on real] noted, 'This incident stresses the urgency for mandatory AI literacy in public administration, where universities can lead workshops.' Stellenbosch University's AI4D Africa lab, involved in continental AI ethics, called for collaborative redrafting to embed academic rigor.

Studies show SA public higher education institutions are progressing on AI guidelines, with over 70% having draft policies on hallucinations and ethics. The scandal could accelerate university-government partnerships for verification tools.

Broader Implications for AI Governance in South Africa

The withdrawal delays SA's AI strategy amid global races (e.g., EU AI Act). Economically, AI could add 15.7% to GDP by 2030 per prior DCDT estimates, but credibility lapses risk investor hesitation. For higher ed, it spotlights opportunities: unis training policymakers in prompt engineering and citation validation.

Stakeholders urge data sovereignty emphasis, protecting local datasets from Big Tech dominance—a pillar resonating with SA's digital economy goals.

Global Parallels and Lessons for Policymaking

This echoes Australia's 2025 Deloitte report fix after AI fakes and US lawyer cases fined for ChatGPT citations. Lessons: hybrid human-AI workflows, post-generation verification, and training. SA can draw from university playbooks, like UCT's AI ethics framework mandating source checks.

In higher ed, where AI aids research, similar scandals (e.g., fabricated peer reviews) prompt tools like SciScore for integrity.

South African Universities' Role in Future AI Policy Development

Institutions like UJ's AI Research Unit and NWU's centres position higher ed as policy advisors. Proposals include joint taskforces for redraft, integrating academic expertise. Curricula expansions in AI governance at unis like UKZN could supply skilled drafters.

Statistics: SA produces 5,000 AI grads yearly, but policy gaps hinder scaling. This scandal may boost funding for AI verification research.

Photo by Mauro Romero on Unsplash

Challenges and Solutions for AI in Public Administration

• Challenge: Overreliance on unverified AI outputs. Solution: Mandatory human review protocols.
• Challenge: Skill gaps in gov. Solution: University short courses on AI ethics.
• Challenge: Ethical risks in policy. Solution: Pre-publish academic peer review.
• Challenge: Hallucination detection. Solution: Tools like Hugging Face's verifier, trained on SA contexts.

Actionable: DCDT partner with SA universities for AI sandbox testing in policymaking.

Future Outlook: Rebuilding Trust in SA's AI Ambitions

A redrafted policy, expected soon, could emerge stronger with academic input, emphasizing Ubuntu in AI ethics. Higher ed stands ready: Wits' quantum-AI fusion, UCT's data science masters. By addressing this lapse, SA can model responsible AI globally.

Optimism prevails—scandals accelerate maturity. Universities, as innovation hubs, will drive inclusive AI growth.