Criticism: the global AI safety agenda overlooks deployment risks in Africa and the Global South

Liz Orembo of Research ICT Africa, a member of the technical committee drafting AI policy for Kenya, criticizes the way the global AI safety agenda addresses risks in an article for Tech Policy Press. According to her, UN Secretary-General António Guterres called for agreement among countries on testing AI systems, measuring risk and assigning responsibility at the inaugural Global Dialogue on AI Governance in Geneva (early July 2026) – but the author argues that this call assumes countries already agree on what safety means in the first place.

The author distinguishes “model risks” (risks arising from model capabilities – deception, assistance with biological or chemical weapons, cyberattacks, actions beyond human control) from “deployment risks” (risks of deployment – discrimination, exclusion, surveillance, language errors, the inability to challenge an automated decision). According to her, the global discussion focuses mainly on the first category, while the second dominates experiences in Africa and other countries in the Global South – examples she cites include credit scoring systems that exclude informal workers from access to financing, digital identification programs that citizens cannot opt out of, automated welfare systems with no right of appeal, and medical translation models that fail in languages they were not tested for.

According to the author, the safety discussion at the meeting in Geneva was framed around the interoperability of standards. A delegate from Pakistan pointed out that safety standards are written by a small group of countries and companies, a representative from South Africa called for frameworks reflecting the realities of developing countries, and the science minister of Costa Rica pointed out that the concentration of AI infrastructure also creates a concentration of evidence about its risks. According to the author, the Independent International Scientific Panel on AI (IISPAI) stated in a preliminary report presented at the dialogue that AI risks disproportionately affect the Global South because of structural vulnerability, limited domestic AI infrastructure and dependence on foreign technologies, and that AI systems are least predictable in resource-constrained environments. As evidence, IISPAI cited cases in which machine translation confused antibiotics with insecticides and mixed up disease categories.

The text of the source article is incomplete; details can be found in the source article.