Orembo: Global AI safety agenda cannot see African harms

A useful piece of framing out of Research ICT Africa names something the international AI safety conversation keeps circling but not quite saying. In Tech Policy Press, Liz Orembo, who leads international and regional engagement at Research ICT Africa and is a UN AI Fellow, splits AI safety into two categories and argues the second is the one African users actually live with.

The first she describes as model risks, the things the frontier labs and AI Safety Institutes largely focus on. In her words, model risks concern what AI systems are capable of: deception, autonomous behavior, cyber capabilities, or assistance with biological weapons. The second she calls deployment risks, which she frames as what happens after those systems leave the laboratory: discrimination, exclusion, surveillance, language failures and the inability of affected communities to challenge automated decisions. The examples she reaches for are grounded and specific: credit-scoring systems that quietly exclude informal workers from finance, a digital identity program that citizens cannot meaningfully refuse, an automated welfare system that offers no avenue for appeal, or a medical translation model that fails in a language it was never tested to understand.

The structural point is that testing capacity is asymmetric. Countries that cannot independently evaluate imported AI systems remain dependent on evidence generated elsewhere, and, as Orembo notes, very few countries from the Global South participate in the AI Safety Institute networks that generate that evidence. Her proposed fix is regional: a safety institute operating through the African Union that pools scarce expertise and builds on existing instruments like the African Union Data Policy Framework and the Continental AI Strategy.

The honest caveat is that this is one analyst's argument, not a survey or an audit. The examples are illustrative rather than a systematic count of harms, and the piece does not put numbers on the participation gap or price out what a regional institute would cost to run. What the reporting doesn't give you is who at the AU would sponsor it, or which frontier labs would agree to submit models for regional testing. The useful thing it does give you is a cleaner vocabulary: if your AI safety strategy stops at frontier model evaluations, it is by definition not measuring the harms that most deployed systems already cause.