Workers Train the Robots That May Replace Them

The men walked the floor of a textile factory in Karur, in southern India, handing out GoPros and smart glasses. Shoba Kirubakaran, 40, put one on. It was the third time she had been told to wear the camera, and as it recorded her every hand gesture while she packed aprons, it was capturing data to train AI robots to take over her job — something she says the company never told her. “Nobody told me anything about it,” she told NPR, which informed her of the footage’s purpose during the interview. She earns $90 a month, working six days a week, and has done this job for a decade. “I only know how to do this job,” she said.

The arrangement sits at the center of a paradox in the global AI and automation boom: the workers whose livelihoods are most exposed to automation are the ones being enlisted — sometimes unknowingly, and often for low pay — to supply the training data that could one day displace them.

A New Kind of Training Data

India is emerging as a key collection site for what the industry calls egocentric data — first-person footage gathered from a human’s perspective — because of its large population, relatively low wages, and openness to new technology. Developers say they need millions of hours of it to teach machines physical dexterity that text cannot convey.

“Think of today’s AIs as brains without a body,” said Prathosh A.P., a professor at the Indian Institute of Science in Bengaluru. “Now you have to give them all motor skills … to do all kinds of intricate motions that we do using a physical body.”

The need is not easily met by simulation. “You can easily scale in simulation, but nothing is as good as real data,” the roboticist Jigar Kumar Patel told openDemocracy, describing a “simulation gap” in which real-world properties — weight, malleability, unpredictability — do not translate cleanly into models. One estimate cited in that reporting projects that robotics labs may spend more than $1.5 billion over three years to acquire between 100 million and one billion hours of footage of human hand-work.

The practice burst into public view in April 2026, after a video showed rows of Indian textile workers wearing camera-fitted headbands. One news anchor called it something out of science fiction. “Indian workers are training AI robots to take their jobs,” another outlet warned. What followed was a wave of investigations tracing the phenomenon from factory floors to living rooms.

The picture that emerged is uneven. Some workers are paid by the hour and told why they are recording themselves. Kannagi Periyasaamy, 42, a textile worker, is paid about $2.60 an hour to wear a GoPro and understands its purpose: “We are training them on how to package the aprons, how to do a quality check, how to fold the aprons.” She is convinced a robot will eventually take her job — “I’m thinking about taking up farming” — yet she describes herself as proud to teach the machine well.

Many others are not told at all. At two factories in Delhi’s industrial belt, workers wore camera headsets supplied by the startup Egolab.AI for roughly two weeks in early 2026 without being told the purpose. Their compensation, according to openDemocracy’s reporting, was a warm tetrapak of Mango Frooti, a fruit-flavored drink. “If I knew it was for training a robot, I would prefer not to wear it,” said one young tailor. “If they want a robot to do this work, then why don’t they get one to do stitch clothes already?”

Some workers pushed back. A tailor named Ankush Yadav and his colleagues pried open the cameras, pulled out the memory cards, and reviewed the footage on their phones — discovering it recorded their voices, their conversations, and their hands in three-minute clips. When complaints were ignored, workers began pulling the devices off or removing and replacing them so often that the footage became useless. Within two weeks, the experiment was quietly shelved at both factories.

A Regulatory Vacuum

India’s Digital Personal Data Protection Act, passed in 2023, requires free, specific, informed, and unambiguous consent. But the law “assumes a direct relationship between the data fiduciary and the individual,” notes the Observer Research Foundation’s Purushraj Patnaik in a policy analysis — consent that, in practice, is routed through factory management. Where employment is insecure, declining to wear a camera is not a real option. Much of the Act’s consent provisions do not take full effect until mid-2027.

Patnaik’s analysis also flags harms that reach beyond the wearer. Pervasive first-person capture records co-workers, private conversations, and proprietary processes — and off-the-shelf AI models, benchmark research found, could infer a wearer’s identity, gender, and race from first-person footage with 70 to 80 percent accuracy. In homes, cameras record family members and the worker in frame alike. The impact falls hardest on cleaners, cooks, and factory hands with the least bargaining power.

“Call it what it is, which is exploitation of labor,” said Astha Kapoor, co-founder of the Aapti Institute, which studies technology’s effect on people.

The Economics of Replacement

Yet the vision of a near-term robot takeover in India is contested — even by some of those selling the data. AI robots remain more expensive to run than human labor in India, where power is costly and, in places, unreliable. Objectways, an Indian company that gathers egocentric data in Karur, runs loud generators to keep its operations humming. “We may be able to make the robots, but they’re bloody expensive to run,” Kapoor said. “Humans are cheaper to run than robots.”

Ravi Shankar, Objectways’ chief executive and a Karur native now based in Arizona, largely agrees about deployment in India: “I think it will be more expensive. I don’t think it’s going to be cheaper.” He said clients mainly want robots for jobs they cannot find people to fill — tapping rubber trees, dusting solar panels in remote areas, or destemming chiles — rather than directly replacing workers in every role.

Inside the company’s offices, employees collect exocentric data, or footage from a robot’s point of view, using grippy mechanical claws to screw and unscrew bottle lids roughly 100 times so a machine can learn what a water bottle looks like. On a simulated three-bedroom house set, workers record themselves folding towels and unpacking bags for domestic robots — including one model that retails for $20,000. “The robot is like a child,” said Objectways’ Rajeshkumar Ramaswamy. “We are the humans. We are teaching the robots what needs to be done.”

Workers as the Data Source — and the Cost

For some participants, the trade is straightforward. Sneha Jegadish, 24, records herself doing housework for about $2.60 an hour, earning roughly $120 a month. “The work we are doing now is going to be done by robots in the future,” she said. In Chennai, a homemaker told reporters she records her chores because no other work pays as well, even while acknowledging a robot trained on her footage may one day do her job.

The economics cut both ways. The work is self-terminating: once a model performs a task well enough, demand for footage of that task ends — leaving neither the original job nor the data-documentation work behind. Critics argue the harm is collective rather than individual, extracted from workers as a class, which makes individual consent a weak unit of protection.

India’s government think tank, NITI Aayog, has warned that most discussion of AI and labor focuses on white-collar professionals while paying “little attention, if any” to how AI can serve the country’s roughly 490 million informal workers. Amit Basole, an economics professor at Azim Premji University, put the challenge bluntly: retraining and absorbing displaced workers requires “policy agility,” and “I doubt it” — India has it.

What to Watch

India’s Ministry of Electronics and Information Technology has taken note of in-home recording pilots and is examining the sector, but no penalties have been levied. Patnaik proposes recognizing egocentric collection as a distinct activity under the country’s data rules, requiring consent directly from the person recorded, separating consent for AI training from consent for a service, and mandating device-level safeguards such as default anonymization and geofencing that disables recording in hospitals and schools.

The stakes extend beyond India. Morgan Stanley projects more than a billion humanoid robots could be in use by 2050, most of them industrial or commercial. How the world’s second-largest workforce is treated as it supplies the raw material for that future — informed, compensated, or neither — will shape not just India’s labor market, but the template for a global data economy built on human movement.