The invisible work: the humans teaching robots how to hold a glass

Before humanoid robots step into our factories and kitchens, thousands of people spend their workdays filming their own hands. It is mundane, repetitive, and it sits at the heart of one of the deepest economic transformations of our era.

---

A camera on your forehead, eight hours a day

Picture this: you arrive at work, strap a camera to your forehead, and spend the day folding napkins. Stacking boxes. Plugging in cables. Pouring water from a pitcher into a glass. Again. Again. Again.

That is daily life for thousands of workers in India today. The camera captures their hands from a first-person view — every micro-movement, every grip adjustment, every instinctive wrist correction. Those clips then travel to data centers in the United States, where neural networks dissect them to teach humanoid robots how to interact with the physical world.

The industry calls this handforms.

The logic is simple, almost disarming. For a robot to learn to grasp an object without breaking or dropping it, it needs millions of human examples. Millions of annotated, classified, ranked gestures. And today, filming hands in India costs less than deploying robots to generate the same data. As long as that stays true, people will wear cameras on their foreheads.

That is not a bug in the system. It is the system.

---

White-collar workers are in it too

It would be convenient to believe this story only concerns manual jobs — that engineers, lawyers, and financial analysts are safe. That would be a mistake.

Look at what platforms like Mercor or Scale AI have been doing for the past two or three years. They are not hiring low-skilled workers to click on images. They are hiring former Goldman Sachs staff, doctors, lawyers, senior engineers — and paying them up to $200 an hour for one very specific thing: transferring their expertise into AI models.

Correcting a line of reasoning. Evaluating a response. Annotating a contract. Producing the “ideal” example so a model learns to do what they have done for years.

Mercor was founded in 2023 by three people in their twenties. Less than three years later, the startup is valued at $10 billion. It pays contractors more than $1.5 million per day. Its clients are named OpenAI, Google DeepMind, Anthropic, Meta. The growth curve is so steep it has made its founders among the youngest self-made billionaires in history.

The model rests on a tension few people see. The big labs need sharp sector expertise — finance, law, medicine — to refine their models. But banks and law firms are obviously not going to hand over their data: that would mean signing their own death warrant. So Mercor sidesteps the problem by hiring former employees of those same institutions. Founder Brendan Foody puts it bluntly: “Goldman Sachs does not like the idea of models that can automate their value chain. That is why the labs need us.”

Expertise walks out the side door.

---

What they do not tell you when they hire you

One sentence sums up what is happening better than any analyst report:

You are no longer paid to practice your craft. You are paid to transfer it.

For centuries, a person’s value in the labor market rested on something hard to copy: intelligence, judgment, instincts forged in the field. That scarcity was the implicit contract between people and the economy. It was the price of being human.

That contract is being rewritten — without anyone handing you an updated version at signing.

The handforms worker is not paid to fold napkins. They are paid so the machine learns to fold napkins without them. The analyst who evaluates an LLM’s reasoning at Mercor is not paid for the analysis itself. They are paid to make that analysis obsolete in the long run. At bottom, both do the same job: they gradually make themselves replaceable, and they are compensated precisely for that.

This is not cynicism. It is the most honest description of what is being built in plain sight.

---

A value that collapses

What makes the situation hard to grasp is that the movement is gradual. But the direction is clear.

Appen was the global reference for data labeling in 2020 — valuation: $4.3 billion. Today, the same company is worth less than $130 million. Not because the sector collapsed — on the contrary, it is exploding — but because the simple tasks Appen handled have been absorbed by the models themselves. AI ate the bottom of the market. Only rare, contextual expertise still holds.

For now.

In some sectors, an hour of GPU time already costs less than a month’s wages in emerging economies. The marginal cost of artificial intelligence is falling faster than economists modeled. Our institutions have not internalized that pace. Neither have our training systems.

---

The real question

Public debate still circles the same lazy framing: will AI replace jobs?

That is no longer the urgent question.

The urgent one is this: what happens to an entire economy when intelligence stops being a scarce asset?

When the ability to diagnose, analyze, draft, decide, and anticipate becomes abundant and cheap, the whole value system that has organized labor markets for two centuries wobbles. This is not science fiction. It is not absolute certainty either. It is a probability that strengthens with every thousand hours of human data poured into a model.

What we are living through now is the encoding phase: the moment when human intelligence is systematically captured, compressed, and transferred. What comes next — the autonomy of those machines — is already underway, less spectacularly than people imagine, but with a regularity sector numbers confirm every quarter.

Almost no one is prepared for the moment when the cost of intelligence approaches zero.

---

Sources: Bain & Company Technology Report 2025, TechCrunch Disrupt 2025, sector analysis Mercor / Scale AI / Surge AI / Appen.