Africa's push for a say in AI safety rules outpaces its capacity to test the models

At Melanin Kapital, a financing company in Nairobi, choosing an AI model meant testing the American and Chinese options against each other internally. Founder and CEO Mélanie Keïta said both came with their own risks ^[12]. "We might become a product of other foreign corporations that ... own all our data as opposed to really being autonomous and free," she told Rest of World ^[13].

Users are racing to put a foreign model into production ^[3]. "Everyone is just trying to catch up [on AI deployment]," Keïta said ^[11]. Teams assume a model tested before release is safe for their own customers. Gathoni Ireri, a research scholar at the Ilina Program, a Nairobi-based AI safety research center, said the companies building the models currently run their own safety tests. She wants those replaced by third-party evaluations of risks specific to Africa ^[21].

Last week the same worry reached the UN ^[1]. "Africa ... must be equal co-architects in determining the standards, ethics, and architectures of this technological era," said Lewis Garseedah Brown II, Liberia's permanent representative, at a Security Council meeting on AI ^[2]. Jonathan Shock, an associate professor of math at the University of Cape Town, described what those standards would start from. "There's essentially no real discussion about AI safety. So, there's a real vacuum," he said ^[4].

The first part of the buyer's decision is control. Jane Munga, a fellow at the Carnegie Endowment for International Peace, takes nationality out of it. "The distinction that matters is not Chinese versus American; it is: Who holds the off switch?" she said ^[16]. ChatGPT and Claude are closed-weight models, centrally controlled by their providers ^[14]. DeepSeek and Kimi, along with a few American models, are open-weight. Users can download and modify them, and Munga said that makes them harder for their developers to shut down remotely ^[15].

The second part is whose harm the tests were built to measure. Ireri said "Africa-centric" evaluations should weigh harm against the size of each country's economy. In her view, an incident that might count as minor in a Western country could do irrevocable harm in an African one ^[20]. The misuse Rest of World documents is local in that sense. Scammers in West Africa have used AI-generated content to blackmail victims, and Boko Haram has used AI to organize attacks and design weapons after getting around safeguards ^[7].

For the vendors, the open question is whether last week's statements become national rules. The reporting does not describe any such rule in force, and Shock expects a long wait. "I think it's going to take a potentially catastrophic incident for governments to really take [heed]," he said, and called that his optimistic view ^[8]. "The pessimistic view is we don't stop it in time," he said ^[9].

For the operator signing off on a pilot on Monday, I'd sort the decision on two axes. The first is who holds the off switch: the vendor, or you. The second is who tested the model against your customers' languages and your scale of harm: the vendor alone, or someone independent. A closed model tested only by its maker leaves both the shutdown and the evidence with the vendor, and I'd expect most fast pilots to sit in that box. An open-weight model you host gives you the switch, and with it the job of knowing when to use it. I'd recommend treating the vendor's own tests as the minimum for anything that touches health care or lending, and paying for an outside evaluation in the languages your customers use. That costs time and money. It also has to be bought in countries where many governments lack the means to test these systems independently, according to experts Rest of World interviewed ^[3].