AI Ethics in Africa: Who Is Responsible When the Algorithm Gets It Wrong?

The conversation about AI ethics in Africa is changing. It is no longer theoretical. AI systems are making — or materially influencing — decisions about credit access, insurance pricing, healthcare triage, public service allocation, and law enforcement across the continent. When those systems are wrong, people are harmed. And when people are harmed, the question of accountability becomes urgent: who is responsible?

This is not a comfortable question. It implicates the technology vendors who built the systems, the organisations that deployed them, the regulators who permitted their use, and the policymakers who have been slow to establish frameworks governing AI in high-stakes contexts. It also implicates the AI practitioners — people like me — who have spent years advocating for faster AI adoption without always being sufficiently rigorous about the governance frameworks that need to accompany it.

The Accountability Gap

The fundamental governance challenge with AI systems is the diffusion of accountability. When a loan officer denies a credit application, accountability is clear — the officer made the decision, their manager can review it, the customer can appeal it. When an AI credit scoring model denies the same application, accountability becomes murky. The data scientist who built the model is not present. The business analyst who validated it may have moved to another organisation. The vendor who supplied the underlying algorithm may be in a different country. The compliance officer who approved its use may not have fully understood how it worked. And the person whose application was denied has limited recourse and often no explanation.

The absence of clear accountability is not a technical problem. It is a governance design problem. And in Africa, where regulatory frameworks for AI are still nascent and civil society organisations monitoring algorithmic systems are few, the accountability gap is wider than in most other regions.

Where the Risk Is Highest

Credit and Financial Inclusion

Alternative data credit scoring — using mobile money history, utility payments, and behavioural signals to assess creditworthiness — has genuine potential to extend financial access to underserved populations. It also has genuine potential for systematic bias. If the training data reflects historical patterns of financial exclusion, the model will perpetuate those patterns. If the features used as proxies for creditworthiness correlate with protected characteristics — geography, ethnicity, gender — the model will discriminate, regardless of whether those characteristics are explicitly included.

Public Service Allocation

AI systems used to allocate public services — healthcare resources, social grants, housing — operate in contexts where errors are not just commercially costly but potentially life-altering. An algorithm that systematically under-allocates resources to particular communities because of patterns in historical administrative data is causing harm that is invisible in the aggregate statistics but devastating at the individual level.

What Responsible AI Deployment Requires

Explainability as a Non-Negotiable

Any AI system making decisions that affect individuals' access to services, credit, or opportunities must be able to explain its outputs in terms the affected person can understand. This is not a technical nicety. It is a prerequisite for accountability. Without explainability, there is no basis for appeal, no mechanism for identifying systematic errors, and no possibility of meaningful oversight.

Bias Auditing Before Deployment

Models should be tested for differential performance across demographic groups before deployment in high-stakes contexts — and regularly after deployment as conditions change. This requires having the demographic data to conduct the audit, which itself raises privacy questions that must be addressed thoughtfully. But the alternative — deploying systems that may be systematically biased and finding out only when the harm has already been done — is not acceptable.

Human Override at High-Stakes Decision Points

For decisions with significant consequences for individuals, there should always be a meaningful human review mechanism. AI can assist and accelerate decisions. It should not be the final and unappealable arbiter of access to credit, healthcare, housing, or justice. This principle is not anti-AI. It is pro-accountability. The two are not in conflict.

My Own Reckoning

I have spent much of my career advocating for faster, more ambitious AI deployment in African organisations. I believe in that mission. Organisations that use data and AI well serve their customers better, operate more efficiently, and create more value. But advocacy for AI adoption without equal advocacy for the governance frameworks that make AI trustworthy is incomplete. The conversation about who is responsible when AI causes harm is one that practitioners, not just regulators, need to be leading. I am committed to leading it.