Africa Is Adopting AI Faster Than It Can Check If It Is Safe

African businesses and governments are adopting artificial intelligence built elsewhere, but the systems are entering markets where the capacity to independently test them is still developing.

That gap is becoming a business problem.

At a United Nations Security Council meeting on artificial intelligence last week, Liberia’s permanent representative Lewis Garseedah Brown II argued that Africa should be an “equal co-architect” in determining the standards and ethics governing AI. Somalia’s state minister for foreign affairs, Ali Mohamed Omar, made a similar case, saying African countries need the capacity to evaluate the technology they are adopting.

The diplomatic language can sound abstract.

The underlying issue is not.

African companies are already putting AI into products and operations. Regulators are beginning to encourage or require automated systems in sensitive parts of the economy. Governments are experimenting with AI in public services. But the continent still has limited capacity to independently determine whether these systems perform reliably in its own languages, markets and institutional environments.

That creates a dependency that goes beyond buying technology.

If a company imports the AI, and the AI company defines the tests, who decides whether the system is safe enough to use?

The adoption is already happening

Africa is not sitting on the sidelines of the AI economy.

PwC’s 2026 research found that 82% of organisations surveyed across Africa were running AI pilots, although relatively few had scaled those systems across their businesses. The average investment was also only about 2% of revenue, compared with 5% among global leaders.

That means the continent is still early in adoption, but the direction is clear.

AI is moving from demonstrations into business processes.

In Nigeria, for example, the Central Bank has issued baseline standards for automated anti-money-laundering systems that use technologies including artificial intelligence and machine learning to detect, analyse and report suspicious activity. The standards apply across financial institutions and are intended to support real-time monitoring and compliance.

The significance is easy to miss.

An AI system used to write marketing copy can produce an embarrassing answer.

An AI system used to monitor financial transactions can affect whether a customer’s activity is flagged as suspicious.

The second system needs a different level of scrutiny.

The same is true as AI moves into credit decisions, insurance, healthcare, education, recruitment and public services.

Nigeria is building the rules while deploying the technology

Nigeria is not without an AI governance framework.

Its National Artificial Intelligence Strategy calls for a national AI policy framework covering responsible AI design, development, deployment and use. It also proposes a National AI Risk Management Framework for identifying, assessing and mitigating AI safety and security risks.

The Nigeria Data Protection Commission is also treating AI as a data-governance issue.

In December 2025, the Commission highlighted risks from automated decision-making without adequate human oversight or explanation, including decisions affecting jobs, loans, health and profiling. In February 2026, it joined international privacy regulators in warning organisations deploying generative AI systems about privacy, transparency, safeguards and mechanisms for removing harmful content.

That is an important foundation.

But having rules is different from having the technical capacity to independently test every system that enters the market.

And that distinction is becoming more important as AI becomes more capable.

The problem is not only whether AI works in English

One of the clearest reasons Africa needs its own evaluation capacity is language.

A 2026 research study examining AI safety in Hausa and Yoruba found that several models performed differently when instructions and conversations were presented in those languages. The researchers found inconsistencies in safety classification and lower performance in some evaluations compared with English.

Another 2026 study, UbuntuGuard, tested general-purpose and safety models against culturally grounded African-language scenarios. Its researchers concluded that English-centric benchmarks can overestimate multilingual safety and that models still struggle to fully localise African-language contexts.

This is not a minor translation problem.

Suppose an AI system is used to screen customer complaints, moderate content, provide health information or assist with financial services.

If it understands English well but interprets a request incorrectly in Hausa or Yoruba, the system may still appear technically impressive in a global benchmark.

The user experiences something else.

A model can pass a safety test and still fail the people using it.

Africa has started building its own tests

The encouraging part is that the evaluation infrastructure is beginning to emerge.

In July, the GSMA-backed African Trust & Safety LLM Challenge produced a benchmark containing 4,216 verified and reproducible safety tests covering African languages, multilingual prompts and code-switched contexts. More than 42,000 adversarial attempts were submitted by participants through the Zindi community before validation and deduplication reduced them to the final benchmark.

That matters because it turns a broad complaint—“AI does not understand Africa”—into something that can be tested.

Researchers can ask a model the same kind of harmful or sensitive question in different languages and contexts, measure the results and compare models.

Other research is going further.

A study by African researchers on Africa-centric AI safety evaluations argues that risks need to be evaluated against African conditions, including weaker infrastructure, limited technical capacity, institutional constraints and different levels of exposure to harm.

This is the beginning of something Africa will need more of: independent evidence about how imported AI behaves locally.

The commercial stakes are bigger than safety

There is a temptation to treat AI safety as a regulatory issue that businesses can leave to governments.

That would be a mistake.

For companies, AI evaluation is increasingly becoming part of operational risk.

A bank using AI for anti-money-laundering checks needs to know whether its system generates too many false positives. A lender needs to understand whether an automated model treats different groups fairly. A healthcare provider needs evidence that a system performs reliably on the patients it serves. A business using an AI agent needs to know what the agent can access and what happens when it makes a mistake.

The IMF’s 2026 work on AI in sub-Saharan Africa makes the broader point: the region’s ability to benefit from AI is constrained by gaps in electricity, digital infrastructure, skills and regulatory and institutional capacity. It also warns that AI can increase dependence on foreign providers.

That dependence matters because the biggest AI systems are still largely developed outside Africa.

African businesses therefore have two separate decisions to make.

The first is which AI system to buy or use.

The second is whether they have enough information to know what they are buying.

The second question is receiving much less attention.

The off-switch matters too

There is another layer to the problem.

Different AI systems give users different degrees of control.

Closed models are controlled centrally by their providers. Open-weight systems can be downloaded and modified, which can provide more flexibility but can also make it harder for a provider to intervene remotely if a deployment goes wrong.

The distinction, as Carnegie Endowment fellow Jane Munga told Rest of World, is less about whether a model is American or Chinese and more about who holds the off-switch.

That is a useful question for African businesses.

If a Nigerian company builds an important service around a foreign AI provider, what happens if the provider changes the model, changes its terms, restricts access, suffers an outage or decides that a particular use case is no longer permitted?

AI dependency is therefore not simply a question of where the model was trained.

It is also about who controls the infrastructure and the rules around its use.

Africa already has a continental strategy

The continent is not starting from zero.

The African Union adopted its Continental Artificial Intelligence Strategy in 2024, describing AI as a strategic asset and calling for an Africa-centric, development-focused approach to ethical and responsible AI.

In 2025, an AU high-level policy dialogue went further, calling for domestic AI capacity, African-led research, high-quality datasets, computing capability and stronger governance. It also called on African countries to regularly assess AI risks and adapt regulatory frameworks to protect privacy and personal data.

In other words, the policy ambition already exists.

The problem is capacity.

The continent needs people who can test models, researchers who can build relevant benchmarks, regulators who understand the technology, companies that can audit AI systems and institutions capable of investigating failures when they occur.

It also needs businesses willing to pay for that capacity.

This could become an African technology market of its own

That last point is important.

The AI safety debate is usually framed as a cost.

For Africa, it could also become an opportunity.

The continent already has researchers building language-specific safety benchmarks. It has data scientists stress-testing models. It has companies working on local-language quality assurance and AI evaluation. The emerging ecosystem suggests that AI safety does not have to mean importing another layer of foreign technology.

It can become a local industry.

Companies could build services around model testing, red-teaming, African-language evaluation, data governance, compliance monitoring and AI incident reporting.

That would give African businesses something they currently lack: a way to assess AI before putting it into sensitive operations.

And it would give the continent a more meaningful role in the AI economy than simply being a large market for products developed elsewhere.

The real test is whether Africa can build the capacity before it needs it

The World Bank’s latest work argues that AI could help developing economies accelerate productivity, but only if countries close gaps in electricity, connectivity, skills and institutional quality. It also warns that AI is spreading faster and is more context-specific than earlier general-purpose technologies.

That creates a narrow window.

African governments do not need to build the next OpenAI or Anthropic before their businesses can benefit from AI.

But they do need enough technical and institutional capacity to ask difficult questions of the companies supplying it.

Does the model work reliably in the languages our customers speak?

What happens when it makes a consequential mistake?

Who can audit it?

Where does the data go?

Can the company explain how the system behaves?

What happens if the provider changes the model?

And who is accountable when the answer is wrong?

The African leaders who raised these issues at the UN were asking for a bigger seat at the table.

For businesses, the more immediate issue is simpler.

Africa needs the ability to test the AI it is already buying.

Without that capacity, the continent risks becoming a sophisticated customer of technology whose most important rules are still written somewhere else.