Why the Global AI Safety Agenda Cannot See African Harms

Liz Orembo is Lead, International and Regional Engagement at Research ICT Africa and a UN AI Fellow. She serves on the technical committee developing Kenya's AI policy and chairs the Africa Tech Policy Summit.

When UN Secretary-General António Guterres opened the first Global Dialogue on AI Governance in Geneva earlier this month, he argued that when countries align on how they test AI systems, measure risk and assign responsibility, safety travels with the technology. This was an important call for international coordination, which also rested on an assumption that countries already agree on what safety is.

From an African perspective, the global AI safety agenda—for as much as it has concerns labeled as the Global South—is increasingly measuring one category of risk while overlooking others. Most international efforts focus on model risks that arise from the capabilities of increasingly powerful AI systems themselves. Yet many of the harms experienced across Africa are deployment risks that emerge from interactions among AI systems, the institutions deploying them, and the societies receiving them. In other words, risk is relational and does not reside in code alone.

The global AI safety conversation has largely centered on frontier models and their potential to deceive humans, assist biological or chemical attacks, enable sophisticated cyber operations or eventually operate beyond meaningful human control. These are legitimate concerns, and they deserve serious attention. Not to assume perfect institutions in the global north, but comparatively, for millions of people, especially in countries where public institutions have fewer resources to prevent or remedy harm, the technologies already threatening livelihoods are far more ordinary.

Such deployment risks might appear ordinary in other contexts, but are existential especially because of the nature of their slow violence. Take examples such as 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.

What the safety conversations actually cover

At the Dialogue, the safety session was framed around interoperability, so the discussion focused on how national frameworks could be made compatible rather than on defining what safety is. Within that frame, scientists warned that frontier models have shown the ability to deceive humans, while a middle bloc argued for compatible systems built on common standards. Representatives of most governments in the Global South argued over power and capacity. A delegate from Pakistan observed that safety standards are being written by a small group of countries and companies, while someone from South Africa demanded frameworks that reflect developing-country realities, and Costa Rica's science minister noted that the concentration of AI infrastructure produces a concentration of evidence about its risks.

In a nutshell, these interventions exposed the question of who gets to define what counts as a safety problem in the first place, and the distinction becomes even clearer when comparing model risks and deployment risks.

Model risks concern what AI systems are capable of: deception, autonomous behavior, cyber capabilities, or assistance with biological weapons. Deployment risks concern what happens after those systems leave the laboratory: discrimination, exclusion, surveillance, language failures and the inability of affected communities to challenge automated decisions.

A model can therefore pass every frontier safety evaluation and still produce unsafe outcomes when deployed elsewhere because even though the model didn't change, the relationships around it did.

The AI Safety Institutes' agendas center on matters such as whether models can assist with biological or chemical weapons, enable cyberattacks, deceive their evaluators, or act autonomously in unsafe ways. Their access to those models rests on voluntary arrangements with the companies that build them, and some of these arrangements were established with commitments to access frontier models for testing and risk evaluation. Global South countries not part of that network of safety institutes not only deny themselves access to shape the safety agenda, but also miss the opportunity to test and evaluate foreign AI systems, which can often be safe in jurisdictions of deployment and unsafe when exported to their jurisdictions.

Safety frameworks assume a safe background

With little participation from the Global South in such policy processes, conversations on AI safety may assume ‘safer contexts’ with proposals that assume functioning courts, empowered data protection authorities, reliable electricity and connectivity, formal labor markets that generate records, and a press and civil society robust enough to surface failures. Against that background, a risk tier, a transparency report, or a voluntary commitment can be a reasonable instrument because the surrounding institutions convert disclosure into consequences. Transplant the same instrument to a context without that background and it stops working, or it will work with other new harms.

The Independent International Scientific Panel on AI (IISPAI) reached a similar conclusion in its preliminary report presented during the Dialogue. It found that AI risks fall disproportionately on the Global South because of structural vulnerabilities, limited domestic AI infrastructure and dependence on foreign technologies. It also warned that AI systems are least predictable precisely in the low-resource environments where much of the Global South lives.

The IISPAI highlighted how machines have mistranslated antibiotics as insecticides and how these translations have confused disease categories. Such medical instructions, when inaccurately translated into languages that receive little attention during model development, can become matters of life and death, yet these failures rarely appear in frontier safety evaluations because the relevant languages and contexts are seldom included in testing.

The same problem extends to AI Safety Institutes, whose work has become central to evaluating frontier models, often through voluntary arrangements that allow access to proprietary systems before public release. Yet very few countries from the Global South participate in these networks.

Countries that cannot independently evaluate imported AI systems remain dependent on evidence generated elsewhere. They have limited opportunities to test how supposedly safe systems behave within their own languages, institutions and social realities, and as a result, deployment risks remain systematically under-measured.

Africa therefore does not simply need a seat at existing safety tables. It needs institutions capable of producing safety evidence from African contexts.

Kenya has taken encouraging first steps through participation in the international AI Safety Institutes Network and by establishing the Communications Authority's regulatory sandbox to supervise emerging AI applications. These efforts demonstrate growing institutional commitment; however, they require technical capacity to evaluate imported frontier systems, contribute benchmarks and conduct independent testing.

Regional cooperation may offer the most realistic path forward because it would draw on national AI strategies and enable interoperability across Africa’s diverse contexts. Africa's fifty-four countries differ enormously in institutional capacity, digital infrastructure and regulatory maturity. Replicating sophisticated evaluation institutions in every country is unlikely to be feasible. A regional safety institute operating through the African Union could instead pool scarce expertise and build on existing governance instruments, such as the African Union Data Policy Framework and the Continental AI Strategy.

Civil society groups should be integral to this architecture, as they are the organizations working directly with communities and often encounter deployment harms long before they appear in formal reporting systems. They are best placed to experience these relational harms.

Before New York

The first iteration of the UN dialogue, though successful, demonstrated limited capacity amid broad competing agendas on Global AI Governance and coordination. This means that between iterations the secretariat will lean on the background work done by the stakeholders.

The region must invest in regional evaluation capacity, ensure that African languages and deployment contexts are included in international safety benchmarks, and ensure that evidence gathered by African researchers and civil society informs future scientific assessments.

Global coordination on AI safety is necessary. But coordination around an incomplete definition of safety risks scaling the world's blind spots rather than reducing them.

Until AI safety learns to evaluate not only what models can do but also what institutions allow them to become, systems certified as safe in one part of the world will continue to produce preventable harms in another.