AI Companies Pay $25 an Hour to Film Your Chores: How Ordinary Bodies Became the Scarcest Training Data
A staffing manager walks into a café in downtown San Francisco and hands out small packages. Inside each one is an elastic headband with a phone mount. The instructions are simple: strap it to your forehead, go home, and film yourself doing the dishes. Roughly two hours of usable footage pays about $80. One longtime gig worker, who has delivered food and hung Christmas lights for a living, summed up the appeal without much ceremony — he has to do the chores anyway, so he may as well be paid for them.
This is the strangest labor market to appear in a decade, and almost nobody has a name for it yet. For twenty years, the internet trained machines for free. Every caption, forum post, review and uploaded photo became someone’s corpus without anyone signing anything. That era is over, not because scraping stopped working, but because the frontier moved somewhere the web cannot reach.
The physical world was never uploaded
Language models ate text because text had already been digitized. Image models ate photographs for the same reason. But a robot that has to fold a towel needs something the internet never contained: the force required to pinch a fabric edge, the way a sleeve resists when it catches, the correction your wrist makes when the first grab fails, the half-second your eyes spend finding the second corner.
Simulation was supposed to solve this. It didn’t. Virtual environments can teach a machine to backflip, because rigid-body dynamics are tractable. They fall apart at contact — the friction, deformation and slippage that define every task a household robot would actually be bought to do. Which leaves exactly one source of the missing data: human beings, doing ordinary things, recorded.
Investors put over $6 billion into humanoid robots in 2025. The hardware mostly works. The bottleneck is that nobody has enough recordings of people living their lives.
What the jobs actually look like
The market has already split into several distinct occupations, and none of them existed in recognizable form three years ago.
The largest is egocentric video capture — filming from your own point of view. Micro1 has recorders in Nigeria and India strapping phones to their foreheads and moving through their apartments with their hands held deliberately in frame. DoorDash pays delivery drivers up to $25 an hour to film themselves doing chores between shifts. Instawork, which used to staff kitchens and stadiums, now recruits people to build motion datasets. Startups called Shift, Pronto and Human Archive are running the same play with different wrappers — one offers free house cleaning in New York and takes the footage of the cleaning as payment. Encord raised $60 million on the back of a tenfold revenue jump in its physical-AI business.
The second is teleoperation, and it is where China has built at scale what the West is buying piecemeal. More than forty state-backed robot data centers were established across the country, with roughly half operational. A 6,000-square-meter facility in Zigong, Sichuan, opened in January and is designed to generate three million high-quality data entries a year — comparable on its own to the largest open robotics dataset ever assembled from dozens of academic labs. The Beijing center runs past 10,000 square meters with hundreds of workers. One instructor there is a former art teacher who spends eight-hour shifts guiding robot arms through sorting motions in a VR headset, repeating the same movement until the model absorbs it. A single robotic hand may be walked through a new skill ten thousand times.
The third is voice. Platforms pay for accents, for underrepresented languages, and increasingly for unscripted paired conversations — two people talking naturally so models can learn interruption, overlap and turn-taking rather than clean dictation. Speakers of less common languages face less competition and better rates.
The fourth is likeness. People sell their faces outright for avatar and video generation systems, sometimes for a four-figure flat fee.
And the fifth is passive, which is the one most likely to become normal. A $20,000 home humanoid ships with a contract stipulating that remote operators in VR headsets can see through its cameras into your house and drive it through tasks it hasn’t learned. The company’s founder has said the quiet part in public: without your data, the product doesn’t improve. Buying the robot and being the dataset are the same transaction.
The qualification is having a body
Here is what makes this genuinely novel rather than just another gig platform. Almost every labor market of the past forty years has rewarded specialization. This one inverts that. There is no test, no degree, no portfolio. There is a supported phone and a sink full of dishes.
More than that — the ordinariness is the product. A professional demonstrator working in a clean studio produces footage of limited value, because a model trained on tidy examples fails the moment it meets a real kitchen. What these systems are starving for is variance: bad lighting, mismatched crockery, a counter with too much on it, the clumsy second attempt, the drawer that sticks. Your mediocrity at folding laundry is not a defect in the data. It is the data.
That inversion produces the first labor market in living memory where being unremarkable is the credential.
Paid as labor, sold as capital
The economics deserve more scrutiny than they are getting. The recorder is paid by the hour for work that is consumed the moment it is performed. The buyer receives an asset that never wears out, can be copied at zero cost, and will be resold into models that generate revenue for a decade. One side of that trade is a wage. The other is a permanent claim.
There is also a depletion curve that almost no one is pricing in. Data collection is extraction, and extraction has a terminal value. Once a task has been covered by enough demonstrations, the ten-thousandth recording of someone folding a towel is worth close to nothing. The rate you can command for filming chores is highest early and falls as coverage fills in. Anyone treating this as a career ladder is misreading a depletion schedule as an income stream. It is paid chores, not a profession — a bridge, not a destination.
Watch the hourly rate, because it is the single cleanest signal in the whole sector. If pay per hour keeps rising, real-world data remains the binding constraint and the humanoid timelines are longer than the marketing suggests. If it falls sharply, either coverage has been achieved or synthetic generation finally works — and either way the jobs evaporate first.
What you cannot take back
Payment terms are the visible part of the trade. The invisible part is that you are selling biometric material that resists anonymization by construction. Stripping names and locations from a dataset does nothing about the fact that a gait, a voiceprint and a face are identifying by their nature. Voice contributions are commonly licensed for training that includes cloning. One actor who sold his likeness for $1,000, with contract terms restricting political and adult use and limiting the license to a year, later had friends sending him videos of his own face and voice circulating with millions of views.
Egocentric home footage adds a category most contributors never think about: the interior layout of where you live, the contents of your cupboards, who else walks through the frame and whether they consented to anything at all.
The recursion, stated plainly
The obvious critique is that these workers are being paid to automate themselves. It is partly true and worth stating carefully. The person filming laundry is usually not employed folding laundry. But the tasks being harvested — sorting, cleaning, stocking shelves, moving parcels, domestic care — map almost exactly onto the sectors that gig workers cycle through between recording jobs. The displacement is real; it just arrives at a different address than the one on the paycheck.
There is a policy vacuum around all of this. Serious proposals have surfaced for a distinct occupational classification covering demonstration work, with mandatory disclosure of how collected data will be used and what it is expected to be worth commercially. Nothing like that exists today. In its absence, the terms are set entirely by the buyer, and the buyer is the only party in the transaction who knows what the asset is worth.
Somebody is going to teach the machines how to live in a house. The only open question is what the people doing it get in return, and whether anyone thought to ask for a royalty instead of a rate.