There is a document you have probably never read.
It lives on a website, usually buried under a tab called "Eligibility" or "Where We Operate" or some other phrase designed to sound administrative rather than what it actually is. A list of people the platform has decided are worth doing business with, and a list of people who are not.
The approved countries.
You know the list. You've seen it. You've held your breath scrolling down the page, looking for your country's name the way you used to look for your name on a school notice board. And you've felt that specific, quiet humiliation when you reach the bottom without finding it.
This is Issue #4. And today we talk about that list.
The Setup
Let's establish what we're actually discussing.
The AI companies, OpenAI, Anthropic, Google DeepMind, Meta AI, and the entire roster of frontier labs currently racing to build systems that will reshape human civilization, need something that their engineers, however brilliant, cannot manufacture in a lab.
They need human judgment.
Specifically: they need human beings to read AI-generated responses and decide if they're good. They need people to write the questions the AI will be trained on. They need experts, mathematicians, coders, doctors, lawyers, writers, scientists, to evaluate whether the AI's answer to a hard question is actually correct, or just confidently wrong in ways that will cause real harm when deployed at scale.
This is called RLHF. Reinforcement Learning from Human Feedback. It is not a footnote in AI development. It is the process. Without it, the models don't become intelligent. They become sophisticated autocomplete with no guardrails and no judgment, which is terrifying in ways I won't elaborate on because this newsletter is long enough already.
So the companies built platforms to source this human judgment at scale. Platforms like Outlier (owned by Scale AI), Remotasks (also Scale AI, same infrastructure, different door), Micro1, and a constellation of middlemen and subcontractors and outsourcing firms that form what researchers have taken to calling the AI labor supply chain.
They marketed these platforms as the future of flexible, global, democratic work. Work from anywhere. Be part of building AI. Use your expertise. Set your own hours. Join the revolution.
And then they published the list.
The List
Outlier, the largest of these platforms and a direct subsidiary of Scale AI, a company valued at $13.8 billion whose clients include Meta, OpenAI, Anthropic, Microsoft, the US government, and the US Military, operates in 61 countries.
61 countries, out of 195 on the planet.
The countries that make the list include most of Western Europe, the United States, Canada, Australia, and a selection of Latin American nations. The countries that do not make the list include, with impressive consistency, most of sub-Saharan Africa.
Nigeria, 220 million people, one of the largest English-speaking populations on earth, a tech ecosystem that has produced more unicorn companies than any other African nation, does not make Outlier's list. Workers there report being blocked at registration.
Kenya, the continent's most active AI labor market, a country so thoroughly embedded in the global AI training pipeline that its workers personally shaped the safety systems of tools used by hundreds of millions of people worldwide, was served by Remotasks for years. Then, on March 7th, 2024, Remotasks sent an email.
"We are reaching out with an important announcement regarding Remotasks operations in your location. We are discontinuing operations in your current location effective March 8, 2024."
Twenty-four hours notice. No explanation. No severance. No acknowledgment that thousands of people had structured their economic lives around this platform, including Grace Mumo, a single mother of three who lost her job in 2020 and had come to depend on Remotasks entirely. "As I speak," she told journalists the day after the cutoff, "I am wondering what the children will have for dinner because I have no money."
By the time reporters started calling, the block had extended beyond Kenya. Nigeria. Rwanda. South Africa. The exit was not a country-by-country administrative adjustment. It was, as one journalist investigating the story put it, a continent.
Scale AI cited "enhanced security protocols." Which is the corporate equivalent of "it's not you, it's us," deployed at the moment a relationship ends and the more powerful party has nothing left to extract.
The Black Market That Filled The Gap
Here is where it gets both darkly funny and structurally revealing.
The workers didn't leave. The skills didn't disappear. The need for trained, experienced, multilingual, technically sophisticated African workers didn't evaporate because a platform updated its eligibility list. What evaporated was the legal, sanctioned, aboveboard way of accessing those workers.
What replaced it was a black market.
Researchers investigating Outlier's country restrictions found something extraordinary: a thriving underground economy in which Kenyan workers, experienced, competent, in many cases highly educated, were purchasing or renting accounts registered to approved countries. Canadian accounts. American accounts. UK accounts. Silicon Valley accounts.
The workers used residential proxies to mask their locations. They trained large language models on English-language tasks, evaluated AI responses in mathematics and chemistry and biology, did precisely the same cognitive work they had always done. Except now they did it as ghosts, legally nonexistent, inhabiting identities from approved countries, earning wages that were then skimmed by whoever sold them the account access.
One worker, interviewed under a pseudonym, showed a researcher his dashboard: an account registered in Canada, from which he had been training LLMs the previous week. "I can do all work in Outlier regardless of where the account is based," he said.
He was not wrong. The platform cannot tell the difference. The AI being trained cannot tell the difference. The quality of the output does not change because the hands doing the work are in Nairobi rather than Toronto.
Only the terms change. Only the protections. Only the wages. Only the legal status of the worker, who is now one policy enforcement action away from losing everything again, with even less recourse than before.
The platforms banned African workers from the front door. The workers came through the window. And the AI still got trained.
The Math, Which Is Genuinely Obscene
Let's put some numbers on this, because abstractions are comfortable and numbers are not.
Outlier advertises rates of $15 to $40 per hour globally. Workers in the United States, operating from approved accounts, report earnings of $25 to $30 per hour, with documented peak weeks reaching $1,500 to $3,000.
African workers operating through intermediary platforms, the officially sanctioned route for those in countries that make a partial version of the list, report rates of KSh 2,000 to 4,500 per hour for Kenyan workers. That is roughly $15 to $35 at current exchange rates, which sounds comparable until you notice two things.
First: that range represents the ceiling, not the floor. The entry-level rate is significantly lower.
Second: the workers doing the same task, for the same platform, training the same model, from an "approved" account registered in a Western country, earn more. Structurally, systematically, by design.
And third, the detail that should make you set down whatever you're drinking, there are workers doing this labor who, after being kicked off platforms without notice and locked out of direct access, pay intermediaries a cut just for the privilege of accessing work they are fully qualified to do. The wage gets skimmed before it arrives. Not because the worker lacks skill. Because the worker lacks a zip code.
Meanwhile, Scale AI, the company that owns Outlier and Remotasks and this entire ecosystem, was seeking a valuation of $25 billion as of early 2025. Its founder became a billionaire. Its clients include every major AI laboratory on earth.
The workers who trained those models don't have health insurance. Some of them don't know what they'll feed their children.
This is not a footnote. This is the business model.
The Recruitment Scam Hiding In Plain Sight
There is one more piece of this that needs to be said aloud.
Researchers who spent a year studying the mass recruitment strategies of platforms like Mindrift, a subsidiary of Toloka, itself formerly owned by Yandex, discovered something that makes the country restriction story even darker.
These platforms recruit workers en masse. Tens of thousands of applications, massive LinkedIn campaigns, urgent job postings. The language of abundant opportunity, of getting in on the ground floor of the AI revolution.
Except there isn't enough work.
The platforms know there isn't enough work when they recruit. The recruitment is not designed to find workers for available tasks. It is designed to create the appearance of scale, a large pool of registered contributors, because scale is a signal to investors. Scale means you can handle big contracts. Scale means you're a serious player. Scale means the valuation goes up.
On a single day in May 2025, Outlier had over 42,000 jobs advertised on LinkedIn. Researchers describe this as a strategy of hiring en masse despite knowing there is not enough work, purely to create the illusion of scale.
So the complete picture is this: African workers are recruited with the promise of participation in the AI economy. Many are then told their country is not approved. Those who find workarounds operate in legal gray zones, vulnerable and unprotected. Those who make it onto the approved platforms discover that task availability is wildly inconsistent, abundant one week, gone the next, with no notice and no compensation for the gaps. And all of this sits inside a recruitment funnel specifically engineered to look larger than it is, to impress investors who will never speak to a single one of these workers.
You are not a contributor. You are a number on a slide deck.
The Igbo Language Project
Let me tell you about a project that trained AI in the Igbo language.
Igbo is spoken by roughly 45 million people, primarily in southeastern Nigeria. It is a language of considerable complexity, rich with tonal distinctions and cultural specificity that makes it genuinely difficult for machines to learn without human guidance from native speakers.
An AI company needed that guidance. They set up a project. They recruited Nigerian Igbo speakers, people with a linguistic competency that is, by definition, rare and irreplaceable. They formed a group. They began working.
Then the project ended. The coordinator was removed. The workers asked what happened. They were told it was an error, the coordinator would be added back. Then, before anyone understood what was happening, every single worker was removed from the Slack group.
One worker had spent over 20 hours on five Igbo language projects. After all of it, he had less than one dollar to show.
The Igbo language knowledge was extracted. The people who possessed it were discarded.
This is not a data point. This is not an anecdote. This is the structure of the transaction, expressed clearly, without ambiguity, in the account of a real person who gave his expertise and received in return the experience of being quietly removed from a group chat.
Why The List Exists
The platforms will tell you it's compliance. Regulatory complexity. Payment infrastructure. Fraud prevention. The difficulty of operating in markets with inconsistent internet access or banking systems.
These are not lies. They are also not the whole truth.
The whole truth is that the "approved countries" list is not primarily a technical or regulatory artifact. It is a labor market design decision. It determines who can access work directly, under what terms, with what protections, at what wage. It concentrates negotiating power with the platform and removes it from the worker. It creates a tiered labor market in which geography, not skill, not expertise, not quality of output, determines compensation and status.
And it ensures that when the work dries up or the project ends or the "enhanced security protocols" kick in, the workers with the fewest protections absorb the impact first.
Africa was useful when the model needed to learn Igbo. Africa was not useful when the model had learned enough Igbo. The door opened for extraction. The door closed for everything else.
The Demand
This newsletter is not a petition. But if it were, here is what it would say.
African governments need to treat AI labor, the cognitive work done by African workers to train AI systems that will generate trillions of dollars in value, as a regulated industry, not a gray-market gig economy. This means labor protections. Minimum wage requirements applied to digital work. Mandatory notice periods before platforms exit markets. Requirements that platforms operating in African countries register locally, pay taxes locally, and be subject to local labor law.
It means African data workers, the engineers of intelligence, the people who taught the machines, being recognized as the skilled professionals they are, not treated as interchangeable units of cognitive output that can be switched on and off depending on the quarterly needs of a company in San Francisco.
It means African AI companies building the alternative. Not waiting for Outlier to expand its approved list. Building Outlier. Building the platform. Owning the pipeline. Training the models on African data, with African workers, under African terms.
Because here is the thing about approved lists: somebody writes them.
Right now, that somebody is not us.
That can change. It has to change. And unlike most of the things that have to change on this continent, this one requires no minerals, no foreign investment, no infrastructure megaproject, no political will from leaders who have other priorities.
It requires a decision.
A decision to build the table instead of waiting to be added to the list.
Next issue: The Language Wall, where your mother tongue is invisible to AI, and the people who profit from that silence know exactly what they are doing.
Africa AI Today publishes when the continent needs to hear something it already knows but hasn't been angry enough about yet.
If this made you want to do something, good. If you're not sure what, start by sharing it with someone who is building.
Let’s Stay Connected ✨
I’d love to keep the conversation going beyond this post. If you found these insights valuable or simply want to exchange ideas, feel free to connect with me on LinkedIn. It’s a great space to share perspectives, build meaningful connections, and grow together.