There is a moment in The Wizard of Oz when the dog Toto pulls back the curtain to reveal a small man desperately maneuvering buttons and levers. "Pay no attention to that man behind the curtain!" he shouts.
Silicon Valley has built the exact same curtain. It is just woven from finer words: neural networks, self-learning systems, and emergent intelligence. So far, no one seems to have wanted to pull it.
I am going to pull it.
The Machine That Does Not Teach Itself
Imagine you are sitting in Nairobi. The year is 2022. You work nine hours a day in front of a screen, and your only job is to consume the content that is too extreme to ever be shown in a glossy keynote from San Francisco. You classify it, sort it, and clean it up, so that the AI model "officially" never has to see it.
For this, you earn just under two dollars an hour. Your work is the absolutely essential lifeblood of a product valued at tens of billions of dollars. But on the balance sheets, you do not exist.
You are a ghost.
This is not a metaphor. Researchers Mary L. Gray and Siddharth Suri coined the phenomenon Ghost Work back in 2019. Today, the ghosts are more numerous, but they are just as invisible. Behind every major AI system, from Google's Gemini to OpenAI's GPT models, staggering amounts of human-labeled data are required. Algorithms do not learn what a cat is until tens of thousands of images have been labeled "cat" by a human. Self-driving cars do not become self-driving until human annotators have drawn precise bounding boxes around pedestrians and cyclists in millions of video frames.
This is not a teething problem that the technology will soon outgrow. It is structural. The more sophisticated the systems become, the more nuanced, context-dependent data they require, data that only humans can produce.
The Three Layers of the Illusion
To understand the scope of the deception, we must break it down into three layers:
Layer 1: The Annotators, The Digital Assembly Line
The World Bank estimates that between 154 and 435 million people globally work in the online gig economy. Data annotators are a subset of this, the precise number is hard to pin down, but researchers at Oxford's Fairwork project and the DSA Coop place the active AI data workforce in the low millions, concentrated almost exclusively in the Global South. Hourly wages sit around $2 in Kenya and $1.70 in Argentina. Their counterparts in the US start at $18.
Take Scale AI as an example, the industry's flagship. In May 2024, the company was valued at $13.8 billion. Their platform, Remotasks, was long the primary tool used to reach this cheap workforce. When Remotasks abruptly shut down its operations in Kenya in March 2024, thousands of young people were left completely without an income. The notice came via a cold mass email, hours before the shutdown.
But the company didn't stop using cheap labor, they engaged in a corporate shell game. Workers were migrated to new, opaque internal platforms like "Outlier" to keep the exact same machine running under a fresh facade. This lack of transparency is not a bug, it is a design feature meant to confuse regulators and keep workers disposable.
Layer 2: The Human Guardrails
Even darker is the layer of workers who do not train the models, but rather protect us from them. A 2025 survey by Equidem of 76 workers from Colombia, Ghana, and Kenya revealed 60 independent cases of severe psychological harm, including PTSD, panic attacks, and substance abuse.
One of these workers, Nathan Nkunzimana, testified before CBS News about how a psychiatrist had diagnosed him with severe PTSD as a direct result of his work as a content moderator. He had no health insurance. He was employed through a subcontractor to a subcontractor. The legal chain of responsibility back to the trillion-dollar company had been completely erased.
Layer 3: The Lie of Independent Scalability
The industry loves to brag that the percentage of pure manual data labeling is finally starting to drop as they increasingly transition to "synthetic data." But this is a statistical smoke screen.
Even if the percentage of humans drawing boxes around cars is decreasing, the sheer scale of the AI boom means the absolute volume of human labor is exploding. Workers aren't being replaced, they are being shifted to RLHF (Reinforcement Learning from Human Feedback), tasked with grading, correcting, and arguing with AI outputs to make them sound human. The more advanced and "autonomous" the AI claims to be, the more complex and psychologically taxing the human effort required to prop it up becomes. This is not temporary training wheels. It is a permanent, load-bearing wall actively hidden in the accounting.
The New Form of the Colonial Economy
There is a pattern here that is much older than AI. The West consumes a cognitive luxury, smooth chatbots, magical image generators, and robotaxis, and offloads the psychological toxicity onto Kenya, the Philippines, Venezuela, and India. The structure is familiar: the raw material is extracted in the South, while the refinement and the profits remain in the North.
Did They Build Russian Surveillance Systems Without Knowing It?
In some cases, it is even darker. In 2024, The Bureau of Investigative Journalism, together with Follow the Money and Paper Trail Media, revealed that gig workers recruited through the Yandex-owned platform Toloka had unknowingly trained facial recognition systems used to identify and detain protesters in Russia.
Workers in India, Turkey, Pakistan, and Bangladesh were paid pennies per task to draw rectangles around bodies in surveillance footage, or submit photos of their own faces for 50 cents per five images. The end clients were NTechLab and Tevian, both sanctioned by the EU in July 2023 for their role in arbitrary arrests and detentions. Some tasks continued after sanctions were imposed.
They did not know what they were building. They had not chosen to do it. The opaque supply chains are not an unfortunate side effect of how the industry has evolved. They are a design principle. Ignorant workers make no demands.
What Happens If We Do the Math Right?
Let us run an economic thought experiment, with actual numbers.
Scale AI reported roughly $870 million in revenue in 2024 on the way to a projected $2 billion in 2025. The company has an estimated tens of thousands of active annotators globally. If just 50,000 of these were raised from $2 to the US-level $18 an hour, at a conservative part-time volume, wage costs would increase by roughly $830 million per year. That is close to the company's entire 2024 revenue, every single year, just to pay a fair wage.
Add psychiatric care, job security, and reasonable notice periods. Suddenly, the business model looks quite different.
The global market for data collection and labeling was valued at $3.77 billion in 2024, and is expected to reach $17.1 billion by 2030. But those are the visible numbers, the ones that are invoiced and accounted for. They do not include psychiatric care, lost productivity, or the fact that 25 percent of data workers in the US are forced to rely on public food assistance to survive.
These costs have been externalized. They are borne by the individuals, by their families, and by the taxpayers. It is not an accidental market failure. It is a calculated equation.
The question is not whether we can afford to pay fairly. The question is who is already paying, and whether we are prepared to continue pretending that the bill does not exist.
Behind the Curtain
When the dog, Toto, pulled back the curtain, he found an ordinary man. The man in Silicon Valley is probably not evil. He genuinely believes in his mission and sprinkles his presentations with words like alignment, safety, and beneficial AI.
But he has also made an active decision about where to draw the line for his responsibility. And that line happens to consistently fall exactly at the point where human suffering starts costing the company money.
This is not magic. It is not autonomy. It is an assembly line, distributed, digital, and just as monotonously brutal as the one Henry Ford built in 1913.
The only difference is that Ford's assembly line was at least visible.