On July 16, 2026, South Africa put its name to the founding charter of the World AI Cooperation Organization (WAICO), a 29-nation bloc launched in Shanghai on the eve of that city's World AI Conference. The agreement, signed alongside Russia, Brazil, Indonesia, Kazakhstan, Pakistan and 24 others, commits members to a 'people-centered approach' to AI governance and casts itself as a Global South alternative to Washington's AI diplomacy — one built, implicitly, around China's growing fleet of open-weight models like DeepSeek, Kimi and Z.ai (gov.cn).
The timing is awkward. Ten weeks earlier, on April 26, 2026, South Africa's own Minister of Communications and Digital Technologies, Solly Malatsi, was forced to withdraw the country's Draft National AI Policy — published in the Government Gazette on April 10 for 60 days of public comment — after journalists found that its reference list cited academic papers and authors that do not exist. Malatsi called it AI-generated citations 'included without proper verification,' said the failure had 'compromised the integrity and credibility of the draft policy,' and suspended two officials pending investigation (TechCentral). South Africa, in other words, joined an international AI governance bloc while its own domestic AI governance document was a fabricated mess sitting in a drawer.
The Case for Caution
Skeptics of WAICO membership have a real argument, and it deserves to be stated plainly before it's dismissed. A Shanghai-headquartered body founded by Beijing, with the US and EU conspicuously absent, is not a neutral standards forum — it is a vehicle for exporting Chinese technical norms, data-flow preferences and possibly surveillance-adjacent tooling to 29 governments that will have limited leverage over how those norms evolve. Open-weight models carry genuine dual-use risk: the same weights that let a Kenyan startup fine-tune a customer-service bot can, in principle, be stripped of safety training and repurposed. Anthropic has argued publicly that unrestricted diffusion of powerful open models could hand adversarial states tools for cyber or biological harm. None of that is paranoia; it is the same logic that underwrites export controls on dual-use hardware, and a country signing a Chinese-run AI compact without reading the fine print on data governance, IP and dispute resolution is taking on real exposure.
Pretoria's Credibility Gap
But the more immediate problem for South Africa isn't WAICO's provenance — it's that Pretoria has no coherent domestic AI policy of its own with which to engage it. The August 2024 National AI Policy Framework was never gazetted. Its April 2026 successor, meant to fix that, was gazetted and then pulled inside three weeks over the fabricated-citations scandal, with no announced relaunch timeline as of this writing (TechCentral). That leaves South Africa's only operative AI-adjacent law as the Protection of Personal Information Act (POPIA), whose Section 71 restricts decisions based 'solely on the automated processing of personal information' where those decisions carry legal or substantial effect — a narrow, privacy-focused backstop, not an AI governance framework (POPIA Section 71). Joining a 29-nation AI cooperation body is a meaningful act of foreign policy. Doing so with no functioning domestic AI policy to anchor your negotiating position is not principled non-alignment; it's improvisation.
Why Open Weights Are Still the Right Bet
Even so, the pro-innovation case for engaging with the open-weight ecosystem WAICO champions is stronger than the bloc's politics suggest. Chinese open-weight models went from under 2% of tracked global token usage in late 2024 to nearly 30% by the end of 2025, according to an OpenRouter/Andreessen Horowitz analysis of roughly 100 trillion tokens — driven by developers who can download, inspect and run models like Qwen and DeepSeek on their own infrastructure rather than renting access to a closed API (South China Morning Post). For a country where compute, dollar-denominated API bills and connectivity costs are real constraints on AI adoption, that shift matters more than the geopolitics of who wrote the weights. Openness is also, on balance, an accountability tool: a model whose weights and training approach can be inspected by South African researchers is easier to audit for bias or security flaws than a black-box API — the opposite of the case Silicon Valley incumbents make when 179 US startups petitioned Washington to preserve open-model access rather than restrict it (Rest of World).
What Should Follow
WAICO membership and open-weight adoption are not, by themselves, the problem. The problem is treating bloc membership as a substitute for the domestic governance work South Africa still hasn't done. Malatsi's department should finish a credible, human-verified AI policy — POPIA's Section 71 is a privacy backstop, not an AI framework, and the country needs one — before its next multilateral commitment outruns its regulatory capacity again. Nothing about WAICO's charter requires South Africa to adopt Chinese content or data standards wholesale; nothing about open weights requires uncritical use. But a government that just got caught letting an AI hallucinate its own AI policy has a specific obligation to show its diplomatic commitments on AI are being read as carefully as its domestic ones evidently weren't.