The Consent Gap at the Heart of the Robot Boom

DoorDash is paying gig workers a few dollars to film themselves loading dishwashers, folding laundry, and wiping down counters. The consent form says the footage trains "automation systems." It does not name humanoid robots, venture capital, or the $38 billion market Goldman Sachs expects the industry to reach by 2035.

I found this out by signing up myself. On a Tuesday afternoon, the DoorDash Tasks app asked for my name, a photo of my hands, and then offered me a list of household chores I could film for up to $25 an hour. A fifteen-minute video of me loading a dishwasher earned me $4.17. The consent form used phrases like "improve machine learning models" and "help train automation systems." No one at DoorDash said those words meant teaching a robot to replace someone doing the same task in a warehouse somewhere.

This is the consent gap at the center of the robot training data boom: workers know something is being built with their footage. They do not know what, or how much it is worth to the people buying it.

The business underneath is straightforward. Humanoid robots need to learn how humans move through the physical world, and there is no internet-scale dataset of human motion the way there is text for language models. So robotics companies are paying people, through intermediaries, to record themselves doing ordinary tasks. The footage is annotated, labeled, and sold to companies building robots meant to work in warehouses, factories, and eventually homes.

Scale AI has gathered more than 100,000 hours of robot training data, according to MIT Technology Review. Micro1, a US data company, has thousands of contract workers in more than 50 countries filming chores in their homes. Instawork, a gig work platform, recently renamed its CEO "Chief Robot Officer" and announced it collected roughly 100,000 hours of movement data in 2024, 1 million hours in 2025, and is projecting 20 million hours in 2026. In China, more than 40 state-funded robot training centers have workers in exoskeletons and VR headsets repeating the same motions, hundreds of times a day, to generate data for 150 humanoid robot companies now operating in the country.

Ali Ansari, CEO of Micro1, estimates robotics companies are spending more than $100 million each year to buy this data. The people generating that motion are getting a fraction of it.

Zeus, a medical student in Nigeria who asked to be identified only by pseudonym, earns $15 an hour filming himself ironing clothes, making his bed, and washing dishes. He found the work on LinkedIn. He says it is boring but pays better than most jobs available to him. He understands the data goes to robotics companies. He does not know which ones, how the data is stored, or what happens to it if he stops working for Micro1. The consent forms say the data may be used to train "automation systems." They do not name Tesla, or Figure, or Agility Robotics.

"We want humans being humans," Aaron Bromberg, head of Instawork Robotics and Applied AI, told Business Insider, describing the company's Instacore camera system, a wearable array of four body cameras and a small backpack designed to capture how warehouse and hospitality workers move through real workplaces. Instawork says the data is opt-in and anonymized. The company is selling it to robotics companies and research labs.

There is something genuinely strange about watching an industry try to scale human motion the way the internet scaled text. The parallel is imperfect in ways the industry prefers not to emphasize. Text can be copied at zero cost. Human movement cannot. You cannot scrape a person's muscle memory from a server. You have to pay them, somehow, to generate it.

The question no one in the industry can answer clearly is whether this approach works. Ken Goldberg, a roboticist at UC Berkeley and cofounder of Ambi Robotics, popularized the "100,000-year problem": at current data collection rates, a general-purpose robot trained the way language models are trained would require roughly 100,000 years of human movement footage. Instawork says it collected 20 million hours in 2026. It estimates that is 0.04% of the gap.

"It is going to take longer than people think," Goldberg told MIT Technology Review.

Other roboticists think the data flywheel is the only viable path. Either way, if it works, the economic implications are staggering: $6.1 billion poured into humanoid robots in 2025 alone, according to MIT Technology Review, with Goldman Sachs projecting a $38 billion market by 2035. The labor arbitrage of replacing a warehouse worker with a robot that learned from gig workers paid $15 an hour to fold laundry is not subtle.

Right now the deal for workers is: film your hands doing chores, earn a few dollars, contribute to a product that may one day make your job unnecessary. The deal for robotics companies is: buy the data, train the model, ship the robot. Nobody is lying, exactly. The consent forms are technically accurate. The market projections are plausible. The gap between what workers understand about what they're selling and what buyers understand about what they're acquiring is just a gap in disclosure, not fraud.

But it is a meaningful gap. And the people on the short end of it are the ones filming themselves in their apartments, in Nigeria and India and Argentina, doing repetitive tasks for hours so that someone else can build a machine that does them faster, forever, without breaks.

I finished my DoorDash Tasks shift with a recording of me loading a dishwasher. The app rated my footage acceptable and sent me $4.17. Somewhere, I imagine, a robotics company is using it to train a robot to do the same thing.

Whether that robot ever learns to do it well enough to matter is the question nobody can answer yet. The industry is betting billions that it will. The workers are hoping it takes a long while.