South Africa's Communications Minister Solly Malatsi admitted in an interview published on 21 September 2026 that his office 'didn't interrogate the bibliography' of the draft National AI Policy. The draft was withdrawn after it cited fabricated, AI-generated references. He also moved the target for a Cabinet-ready redraft to March 2027, later than the November 2026 date the department had given Parliament. He argued that regulation should enable sectors rather than respond to anxiety, and he framed digital sovereignty as a developmental issue.
The interview does not mention open-source AI. That omission is the analytical question here, not a claim about what the minister said. The redraft will have to decide how South Africa treats open-source and open-weight models, and a delay makes it a better moment to decide deliberately.
What happened, and what the delay buys
The department withdrew the first draft after reports that it cited fictitious academic journals. TimesLIVE reported that AI had hallucinated at least six of its citations, and that two officials were suspended. Cabinet had approved the draft on 25 March 2026, with public comments due by 10 June, according to Connecting Africa.
The fix is an independent seven-member expert panel chaired by Professor Benjamin Rosman of Wits University, announced by Wits. Its members include Vukosi Marivate, Alison Gillwald and Heather Irvine. Press reports describe the panel as advisory, with drafting responsibility staying with the department.
The strongest case for the original timetable is that South Africa cannot afford to sit out AI governance. Delay leaves developers, public bodies and investors without a framework while foreign rules set expectations for them. That argument has force. But a fast policy built on unchecked evidence is worse than a slower one. The episode shows what happens when a policy's evidence base is never verified.
Why open-weight models are the sovereignty question
Malatsi framed digital sovereignty as developmental. For a country without hyperscale training budgets, that points to a practical question: who can run, adapt and audit models on local infrastructure and in local languages?
Open-weight models are the main answer. A South African university, clinic or start-up can fine-tune a freely licensed model on isoZulu or Sesotho data without paying per-token fees to a foreign provider or sending sensitive data offshore. Closed models offer none of that control.
The pressure runs in both directions. Fears over access to foreign hubs are already pushing others toward domestic stacks. Rest of World reports that worries about Washington restricting Chinese models on Hugging Face are accelerating Chinese efforts to build sovereign open-source platforms. Open ecosystems are strategic infrastructure.
The risk-based draft and its open-source blind spot
The withdrawn draft was reported to take a risk-based approach. As Connecting Africa summarised it, it proposed new bodies including an AI Ethics Board, a National AI Commission or Office and an AI Regulatory Authority. I could not confirm from public summaries how it handled open-weight releases, so the points below are about design choices, not what the draft said.
The core hazard of a developer-and-deployer liability model is that it assumes a commercial provider who controls distribution. Open-weight publishers do not. They release weights once, and thousands of downstream users modify and deploy them. Imposing provider-style duties on the original publisher would push a university lab or community project to stop publishing, while large foreign firms absorb the compliance cost easily.
The case for caution is real. Once weights are public, a developer cannot recall them or enforce usage limits, so a harmful capability is out permanently. A regulator that ignored this would be negligent about the most capable models.
What the EU's version shows
The EU AI Act offers a tested template with a clear flaw. Article 2(12) says the Regulation does not apply to AI systems released under free and open-source licences, unless they are placed on the market as high-risk systems or fall under Article 5 (prohibited practices) or Article 50 (transparency). For general-purpose models, Article 53(2) relieves open-licence providers of some documentation duties, provided the weights, architecture and usage information are public. It states that this exception does not apply to general-purpose models with systemic risk.
That is a sensible shape: the exemption is tied to use and to capability, not to the licence label alone. South Africa can improve on it by writing a precise definition of an open-weight release, which the EU text leaves loose, and by focusing obligations on deployers who put a system in front of the public in high-risk settings.
What the redraft should do
- Regulate deployment, not publication. Duties should attach to whoever places a system in a high-stakes setting such as credit, health or policing, not to whoever posted weights.
- Define the category. Say what counts as an open-weight release, including whether weights, licence terms and usage information must all be public.
- Reserve capability triggers. Apply heightened duties only above a clearly stated, evidence-based capability threshold, and review that threshold as evidence develops.
- Use open tooling for oversight. Open ecosystems can also strengthen security. Silicon Republic reports that Anthropic is launching an opt-in programme giving open-source projects free periodic AI security scans with suggested fixes. Public policy should support that kind of defensive capacity.
- Verify the evidence. The panel's remit covers assessing supporting evidence, so the redraft's open-source positions should rest on named, checked sources.
The bottom line
The fabricated bibliography damaged trust in the process, and the longer timeline is the right trade if it is used to get the substance right. The pro-sovereignty goal the minister described and the pro-open-source outcome are the same policy, but only if the text says so. A redraft that stays silent on open-weight models would leave the rules to be filled in later by regulators or by whichever draft language is easiest to apply to a foreign closed-model provider. Writing the open-source treatment in now, with a defined category and capability-based limits, costs little and protects the cheapest route to local AI capacity.