Get Paid to Train the Robot That Takes Your Job

In May 2026, a staffing company called Instawork went through a market in Bengaluru, offering vendors money to wear a camera on their heads while they worked.

The pay started at Rs 300 an hour. Within days, it began to drop, and settled at Rs 200 for most workers. Nothing was put in writing.

Several workers told Newslaundry they’d been told the robots learning from their footage “might eventually do the kind of work they currently do.”

So they’re being paid a little, and then less, to train their replacements.

That was the subject of the first post I ever wrote here.

On November 14, 2025, I published Building the Economic System That Takes Us Past AGI. It proposed Data Foundries: places where people learn whatever they want, record themselves doing it, and own the data that trains the machines.

A reader left a five-part critique in the comments that was sharper than most published rebuttals I’ve read. I answered every point, at length.

Eleven months later, I think the original was right about the problem and wrong about where the answer fits.

Data Foundries aren’t the destination. The destination is still the one line this whole series fits in: universal basic goods and services, powered by universal machine labor on demand, distributed by universal basic capital.

Data Foundries are a bridge. They belong in the window I described in Part 19: Plan for the Bad Transition: after machines start taking work, before the floor and the community nodes exist.

And they only work if they’re built ethically. Because, honestly, parts of my own idea make me uneasy now.

The paradox

I wrote the original the day after reading Android Dreams, a scenario about humanoid robots published November 13, 2025. It’s a story, not a forecast, but its point about data stuck with me.

Language models learned from the internet. Robots have nothing like it. In the report’s words, “there’s no ‘internet of robotics,'” and the real world is too slow for robots to learn by trial and error alone.

So where does the data come from? Us. The report predicts workers doing their jobs while “being passively recorded by multiple cameras,” with robots trained to imitate them.

Here’s how I put it last November:

“Workers perform their jobs as usual. Cameras capture everything. The robots learn.”

”And then those workers lose their jobs.”

The machines that could replace human workers need human workers to become useful in the first place. So I asked: what if we closed the loop?

My answer then is still my answer now: “Not robots offsetting our labor with value going to robot owners, but our labor training the robots with the value coming back to us.”

It’s already happening

When I wrote that, learning from human video was a prediction. It isn’t anymore.

Tesla. In 2025, Tesla moved Optimus training away from motion-capture suits and teleoperation toward camera video, Business Insider reported in August. By November, Business Insider was describing workers who repeat each action “hundreds of times during eight-hour shifts,” filmed by five cameras on a helmet. One former worker called it being a “lab rat under a microscope.”
Figure. In September 2025, Figure launched Project Go-Big, filming people in homes through a partnership with Brookfield, which owns over 100,000 residential units. Figure said its robot learned to navigate homes from “100% egocentric human video data.”
DoorDash. On March 19, 2026, DoorDash launched Tasks, an app that pays couriers to film themselves loading dishwashers, washing dishes and folding clothes, to train its own AI and its partners’.
Micro1 and Scale AI. In April 2026, MIT Technology Review reported that Micro1 has hired thousands of contract workers in more than 50 countries to film their chores with iPhones strapped to their heads. One worker it profiled, in Nigeria, earns $15 an hour. Micro1’s CEO estimates robotics companies now spend more than $100 million a year on this kind of data. Scale AI says it has gathered more than 100,000 hours.
Instawork. Before Bengaluru, it opened a robotics lab in the US that pays workers to film household chores.
1X. For tasks its NEO home robot doesn’t know, the order page says “an Expert from 1X can remotely supervise its actions at scheduled times.” A stranger, looking through a robot’s eyes, in your house.

So the loop exists. It just runs one way.

The workers MIT Technology Review interviewed knew their footage trains robots, but “none of them know how exactly their data will be used, stored, and shared.”

In Bengaluru, Instawork said it offers a way to withdraw and delete your data. The workers Newslaundry spoke with didn’t seem to know it existed.

That’s an open loop: a one-time wage that falls as the robot improves, for an asset that pays someone else for years.

What a Data Foundry is

Here’s the bridge version, keeping what was strong in the original.

A Data Foundry is a place where people practice skills, get coached, and get paid for recordings they collectively own. A workshop, a kitchen, a garden, a gym, a clinic.

Some are big. Some are a room in a library. You don’t have to live there.

Six ideas carry over from November:

Opt in, and leave anytime. As I wrote then: “You can leave anytime (obviously), no questions asked.”
Three recording zones. Green: no recording. Yellow: recording for your own feedback only. Red: recording for AI training, which you opt into. Your home and your devices record nothing unless you turn them on.
You own your data. Members hold it through a data cooperative. You can review it, delete it, or share it selectively.
Dividends, not just wages. When the cooperative licenses a dataset to a robotics company, it keeps getting paid for as long as the model earns.
Graduation. When machines can do a task reliably, it “graduates” from human labor to machine labor and comes off the list of things anyone has to do. I called it the 99.9% rule. In the original’s example: “Welding graduates.” Not the welders. The category.
Many Foundries, many cultures. Some high-tech, some contemplative, some traditional, some radically experimental. “It would be a system of many models, not one.”

Graduation is still the part I like best. It turns the paradox into a ratchet. As I wrote: “You’re never stuck washing dishes for 40 hours a week because you taught the robot how to do it.”

This feels like surveillance capitalism to me

I need to be honest about something.

Reread now, parts of the original make me uncomfortable. I wrote that online, “your activity is recorded (and anonymized) as RL gyms to teach the AI.”

I wrote that anonymization would be “cryptographically guaranteed.” And I wrote this, about recording zones: “Everyone is okay with this because everyone benefits.”

I don’t believe that last sentence anymore.

Shoshana Zuboff named the pattern in The Age of Surveillance Capitalism (2019): a business model that, in her words, “unilaterally claims human experience as free raw material.”

Her target was the ad business. You get a free service. They get your behavior, and sell predictions about you.

Head-mounted robot data is the same deal, with your hands instead of your clicks.

And “anonymized” means less than I thought. In a 2023 study, researchers identified more than 50,000 VR users with 94% accuracy from 100 seconds of head and hand motion.

Those are the same kinds of signals a head camera captures while you fold laundry. Your movements are a fingerprint.

It even cuts against my own work. In Keep Money, Add a Second Ledger, the first privacy rule is “Measure flows, not people.”

A Data Foundry is the one place where the thing being measured is people. So it needs stricter rules than anywhere else, not looser ones.

Here’s what separates a Foundry from surveillance capitalism. If any of these five is missing, it’s the bad version:

Consent. Specific, informed and revocable, per zone and per use. Never a condition of housing, food or care.
Ownership. The cooperative owns the recordings. Companies license them. They don’t buy them outright.
Payment. A fair rate for the session, plus a dividend for the life of the model.
Deletion. You can pull your data out of future training, and the cooperative audits buyers to check.
A two-way exchange. I share my data willingly, and I get to see what it’s used for. The cooperative sees exactly who licensed what and what it trained. And when my demonstrations teach a robot to fold laundry, my town gets that skill back as machine labor on demand.

None of this is new. Jaron Lanier and E. Glen Weyl called it data dignity in A Blueprint for a Better Digital Society (Harvard Business Review, 2018): people should be paid for the data they create, with collective groups, which they called mediators of individual data, bargaining on their behalf.

Think of a union for your data.

In Switzerland, the MIDATA cooperative already does a version of this for health data. Members keep their data in personal accounts, decide which research projects can use it, and elect an ethics board that reviews requests. Profits go back into the platform.

And in 2019, the legal scholars Sylvie Delacroix and Neil Lawrence proposed bottom-up data trusts: trustees bound by “a fiduciary obligation of undivided loyalty” to the people whose data they hold.

They imagined many trusts, so people could pick one that fits their values and “switch Trusts when needed.”

Many Foundries, many cultures, leave anytime. They got there first.

A reader’s five critiques

Now, the comment thread. A reader opened with real doubts and five points. Here’s each one, how I answered, and where the critique still lands.

1. AI will outpace human input

The critique: if AI quickly surpasses human skill, there won’t be much left to teach. Not everyone wants to teach. And you don’t need millions of people repeating the same skill.

I replied: “We will absolutely still need humans to perform most tasks for at least the next 10-20 years, if not longer.” For physical work, I still think that’s about right.

The Berkeley roboticist Ken Goldberg told MIT Technology Review, “It’s going to take longer than people think.” And no single model will learn every task to every person’s taste. I expect many specialized models, each needing its own data.

Where it still lands: the value of any one task’s data falls as the robot learns it. Bengaluru shows it in miniature. Paying by the hour of footage is a melting ice cube.

That’s why the dividend matters more than the wage. A wage for data trends toward zero. A share of the model doesn’t.

2. Real work and real skill

The critique: some work, like massage therapy, is valuable because a human does it. Doing is learning. And if robots get near-perfect, humans become, in the reader’s words, “Sort of like a zoo animal.”

My answer was the massage therapist. It’s physically grueling work.

Picture a robot that takes part of it off your hands: “you focus on the sensitive, deep tissue work while the robot handles the repetitive coverage of the rest of the body.” Healthcare already splits work this way.

And underneath it, the rule: “If a human wants to do it, a human gets to do it. Just because a robot can do it better doesn’t mean the robot has to do it.” That became Part 2: If a Human Wants to Do a Job, a Machine Shouldn’t Take It.

Where it still lands: deskilling. If nobody welds for a living, who notices when the robot welds badly? Graduation can’t mean nobody practices.

That’s the guilds’ job in Part 4: Quests Should Replace Corporations: a charter obligation to keep the craft alive and train anyone who wants in.

3. Struggle and motivation

The critique: removing job pressure ignores what struggle does for meaning, maturity and purpose. And people want different things.

I told the reader, “This is your strongest critique, and I want to address it head-on.” I still think so.

My answer was that struggle doesn’t go away. It changes source. A good Foundry keeps each person in what I called the proximal zone of difficulty: “working on something just hard enough to figure out, but not so easy it feels routine or boring.”

The idea comes from the psychologist Lev Vygotsky, whose zone of proximal development, from the early 1930s, is the gap between what a learner can do alone and what they can do with help.

As for people wanting different things: that’s the point. Not one purpose for everyone, but a place to pursue yours. As I put it then: “Replacing the corporate office with an R&D lab.”

The critique: real communities form around family, place and identity, not a single mission. And you can’t engineer cultural diversity.

I agreed with half of it: “I agree communities don’t organize around one mission. Every university you know today has multiple schools with different missions.”

Where it still lands: the reader was more right than I admitted. Communities do form around place.

That’s why the series moved from campuses you join to trusts that own a town’s infrastructure, in Part 5: Every Town Can Own Its Own Power Plant and Chip Fab. In this version, a Foundry isn’t somewhere you move to. It’s a building in your town.

5. Systemic risk

The critique: engineered societies spiral. The reader’s example was Facebook, which started as a way to help people stay connected and became something far outside its creators’ control. And massive change all at once is brittle.

My reply began, “Unfortunately, you’re right about systemic risk and fragility.” Then: “let’s make one of these foundries, as an experiment to test, and iterate to see if it works.”

That still holds. And Facebook is the surveillance worry in another form: a data business with good intentions can drift. That’s why the five rules have to be written into the charter, and why leaving has to be real.

Where I changed my mind

The roof. The original tied better housing to more hours: “30+ hours in high-demand fields gets you better housing and more resources.” In this series, the floor is unconditional.

Nobody records anything to earn a roof. In the transition, a Foundry sits on top of whatever safety net exists. It’s never the price of a home.

The nines. I wrote 99.9% in one place and 99.9999% in another. Both were wrong in the same way.

The bar should match the stakes. Folding laundry and surgery shouldn’t graduate at the same reliability, so each guild should set its own bar, and the safety-critical ones should set it very high.

Careers. I wrote that “learning replaces labor” and that “careers die.” I’d say it more carefully now.

The career doesn’t die so much as turn into something like a skill tree: many skills, practiced deep or wide over a lifetime. I work that out in You Won’t Have a Career. You’ll Have a Skill Tree.

The dividend. It shouldn’t stand alone. In Part 5: Every Town Can Own Its Own Power Plant and Chip Fab, every resident holds an equal share that can’t be sold, through their town’s trust.

Foundry licensing income should flow into those same shares. And a dataset is what Part 17 calls a fund: something that keeps serving for years if it’s maintained. It belongs on a town’s resource accounts, like a power plant.

That’s universal basic capital, with one more source of income.

Where this started

The original post had a bigger name for all of this: the noosphere, an old word for a planet-wide layer of shared human thought. Every Foundry was meant to be one node in it.

One idea from that post deserves a life beyond the Foundries. You have personal goals, and society has communal ones, so a community should rank the learning and skill areas it needs most and put resources behind them.

That idea now lives in You Won’t Have a Career. You’ll Have a Skill Tree., and at the top of the list for machine labor in Keep Money, Add a Second Ledger. This time it runs on the five rules above: consent first, local by default, nothing recorded in personal spaces, and a share of what it earns for whoever contributed.

Why it’s a bridge

In Plan for the Bad Transition, I wrote that the floor breaks the wage spiral once it exists, and that “The most dangerous period is the window before it does.” Data Foundries fit that window for three reasons.

First, they work at today’s prices. Robotics companies are already paying for this data.

A Foundry doesn’t need a new tax to start. It needs a cooperative with bargaining power.

Second, they’re apprenticeships that pay twice. That piece’s first stage includes employer-led apprenticeships, which, in Julian Jacobs and Alex Imas’s reading of the evidence, beat public retraining “by a considerable margin.”

A Foundry is that, plus a dividend. You learn a skill, and the recording of you learning it earns.

Third, they build ownership early. If labor’s share keeps falling, the plan’s Switch C speeds up universal basic capital. People who already hold a stake in the models trained on their work start ahead.

Then, as tasks graduate and the floor arrives, the Foundry stops being an income source. It becomes what I always wanted it to be: a place to learn, with a camera you can switch off.

In April, I called the idea a potential stepping stone. That’s the right word for it.

Objections

Isn’t this just gig work with a nicer name?

Without the five rules, yes. That’s what’s happening right now. The difference is ownership.

A gig worker sells footage once. A cooperative member keeps a share of everything it trains.

Why would robotics companies agree to share?

Some won’t, and they’ll buy elsewhere. But the most useful data comes from skilled people practicing well, coached and consenting, in places built for it.

That’s what a Foundry makes. Organized sellers of something scarce get better terms.

What happens when everything graduates?

Then the bridge did its job.

The dividends keep flowing from the machines, and the Foundries become what the original pictured at its best: places like Alpha School, which says kids can cover their academics in two hours a day, except for everyone, for life, and for the joy of it.

Close the loop

Right now, people around the world are strapping cameras to their heads and teaching robots their work, for a falling wage and no say in what happens next.

The robots are getting trained either way. The only question is whether the people training them own a piece of what they built.

So I’ll end where I ended with that reader. Let’s build one, as an experiment. One town, one cooperative, one Foundry, with the five rules in its charter and a camera you can turn off.

A Foundry runs on a charter, and so does nearly everything else in this series: trusts, guilds, a tax formula. So next, I hold the whole design up against the Constitution, in Part 21: The World We Want Is Mostly Constitutional.