The Soul of AI

Africa has already supplied some of the human infrastructure behind AI. The question is whether it will remain invisible labour, or become the source of a much more valuable capability.

For years, a sign hung outside a Nairobi office complex: Samasource. The Soul of AI.

Inside, young Kenyans sat at laptops for long shifts, reviewing and labelling some of the internet’s most disturbing material so that an AI system built thousands of miles away could learn what to recognise, reject and refuse.

According to a 2023 investigation by TIME, some of the workers on OpenAI’s data-labelling project took home as little as $1.32 an hour, while Sama was contracted at $12.50 an hour for the work.

The phrase on the building was unintentionally revealing.

Africa really was part of the soul of AI.

It just wasn’t being paid, recognised or positioned that way.

Now put that image next to another headline from this week.

The United States added 162,000 jobs in August, substantially more than economists expected, while unemployment held at 4.1%. At first glance, it looks like a reassuring answer to the growing anxiety about AI and employment.

I don’t think it is.

The labour market can remain resilient while the nature of work underneath it is changing profoundly.

A single good month is not a verdict

The August number is real and worth taking seriously.

But look underneath it. The information sector lost 23,000 jobs in August, including losses in computing infrastructure, data processing, web hosting and related services. The sector had already been losing jobs over the preceding year.

That does not prove AI caused those losses. There are too many forces moving through the labour market at once to make a claim that simple.

But it should make us cautious about treating one strong month of aggregate employment data as evidence that the AI jobs question has been settled.

I find it more useful to watch how the people closest to the technology are thinking about the future.

Dario Amodei, the CEO of Anthropic, spent 2025 warning that AI could eliminate roughly half of entry-level white-collar jobs within five years. In January 2026, he published a nearly 20,000-word essay warning of much broader economic and societal disruption.

Then, in May, he was making a different argument alongside JPMorgan CEO Jamie Dimon: if AI automates 90 percent of a job, perhaps the remaining 10 percent becomes more valuable, allowing people to do much more with the same amount of time.

I don’t think the important question is which Dario Amodei is right.

I think the important point is that nobody knows yet.

When even the people building the technology are revising their expectations as the technology develops, the responsible response is not to choose the forecast that confirms what we already believe.

It is to prepare for more than one possible future.

What survives either future

Here is what I think survives almost every version of the argument, whether AI ultimately creates more jobs than it destroys or causes a much more painful displacement of human labour.

As AI systems become more capable, the consequences of their mistakes become larger too.

A system that can do very little can cause limited damage when it gets something wrong.

A system that can make decisions, execute workflows, influence people, generate content, write code, approve transactions or operate with considerable autonomy is different.

Its mistakes scale.

Its misuse scales.

Its confidence can scale faster than our ability to notice that it is wrong.

That does not mean humans become less important.

It means that human judgment becomes more consequential.

The more capable the machine, the more important it becomes to have people who understand when to trust it, when to challenge it, when to override it, and when the answer that looks correct is wrong because the system does not understand the context in which it is operating.

This is the argument I have been making in different forms all year:

As execution gets cheaper, judgment becomes more valuable.

And nowhere is that becoming more visible than in the work of making AI systems reliable.

Africa is already doing this work

Here is the part of the AI story that many conversations about Africa still miss.

African workers are not waiting for the AI economy to arrive.

They are already inside it.

Since at least the early wave of large-scale data-labelling and content-moderation work, workers in countries such as Kenya have been performing the human tasks that make AI systems usable: classifying data, evaluating outputs, moderating harmful content, verifying information and helping machines distinguish between things they should and should not do.

The International Labour Organization’s research on digital labour in Kenya documents workers performing AI and machine-learning tasks, including data collection, verification, validation and content moderation.

This work is not trivial.

Someone has to decide whether a piece of language is hateful.

Whether an image is dangerous.

Whether a response is culturally offensive.

Whether an answer is technically correct but contextually wrong.

Whether a model has misunderstood what a person actually meant.

Those are not purely computational questions.

They are judgment questions.

And Africa has already been supplying that judgment.

The problem is the position in the value chain at which it has been supplied.

The workers behind these systems have often been treated as invisible, interchangeable labour: paid by the task or by the hour, given little visibility into the systems they are helping construct, and exposed to some of the most disturbing material produced by the internet.

The legal battles in Kenya make the human cost impossible to ignore.

In proceedings involving Facebook content moderators, Kenyan courts ordered interim measures requiring proper medical, psychiatric and psychological care and recognised the hazardous nature of the work.

This is not simply a story about bad employment conditions.

It is a story about where value sits in the AI economy.

The people doing the work closest to the human edge of these systems have often captured the least value from it.

The opportunity is one level higher

This is where I think the conversation about Africa needs to change.

The answer is not to reject AI annotation, moderation or evaluation work.

Those are real capabilities.

The answer is to move up the stack.

The work currently happening in Nairobi is often the bottom rung: repetitive, anonymous, paid by the hour and separated from the decisions being made about how AI systems actually behave.

But there is another layer above it.

AI evaluators who understand not just whether an answer is wrong, but why it is wrong in a particular cultural or linguistic context.

AI safety specialists who can test systems against the realities of African markets.

Red-teamers who can find failure modes that engineers working thousands of miles away might never encounter.

Auditors who can assess whether an AI system is behaving responsibly once it leaves the laboratory and enters an actual institution.

Researchers who can build evaluation frameworks for languages, cultures and contexts that remain poorly represented in the datasets and benchmarks shaping frontier systems.

This is a different kind of work.

It requires technical fluency.

But it also requires judgment, cultural intelligence, domain knowledge, communication and the ability to recognise when something is wrong even when the system itself is confident that it is right.

That combination is valuable.

And it is precisely the combination that Africa should be building deliberately.

From labour arbitrage to capability

There is a temptation to describe this as another outsourcing opportunity.

I think that would be a mistake.

Africa does not need a better version of the same bargain in which companies in Silicon Valley discover that a highly educated Kenyan, Nigerian, Ghanaian or South African can perform the same task for less money.

That is still arbitrage.

The opportunity is fundamentally different.

It is to build a professional class whose expertise is the interface between AI systems and the human contexts in which they operate.

That means training people to understand models, evaluation, safety and governance.

It means creating recognised career paths.

It means building independent assessment and certification.

It means establishing professional standards.

It means creating relationships with the companies building and deploying these systems so that African expertise is bought for its value, not simply for its lower cost.

And it means governments treating digital workers not as an inexhaustible pool of cheap labour, but as part of a strategic capability.

Kenya is already moving toward a broader regulatory architecture for AI. Its draft AI and Emerging Technologies Policy places responsible governance, human capital, public trust and strategic autonomy among its objectives.

That direction matters.

Because the question is no longer simply whether Africa participates in the AI economy.

It is on what terms.

Why this matters for young Africans

Every essay I have written this year about education, from early childhood through university, has ultimately returned to the same question:

What capabilities remain valuable when machines become extraordinarily good at execution?

My answer keeps coming back to the same things.

Judgment.

Context.

Curiosity.

The ability to recognise ambiguity.

The ability to understand people.

The ability to know when the obvious answer is wrong.

AI will make many forms of execution dramatically cheaper.

That does not make human capability irrelevant.

It changes where human capability matters.

And AI evaluation and safety may be one of the clearest places where this transition is already visible.

The field is still young. Its professional structures are still being built. Its standards are still evolving. Its assumptions about who gets to perform this work are not yet deeply entrenched.

That creates an unusual opening.

Africa does not have to wait for the rest of the world to define these professions and then ask for permission to participate.

It can help define them.

The soul of AI

For five years, Africa has been part of the soul of AI in the most literal and least dignified sense.

Human beings have been sitting behind machines, teaching them what is harmful, what is useful, what is accurate and what should be refused.

Much of that work has been invisible.

Much of it has been underpaid.

Much of it has been treated as a temporary bridge to a future in which the machines would no longer need us.

I think we should imagine a different future.

Not one in which Africa remains the cheapest place to find humans willing to do the work machines cannot yet do.

A future in which Africa becomes one of the places the world turns to because it has developed people who understand how intelligent machines behave when they meet real human beings, real institutions and real cultures.

The opportunity is not to become the human layer underneath AI.

It is to become the human intelligence around it.

That is a much more valuable position.

And the time to build it is now.