6 hours ago
12
The first time the factory supervisors handed garment worker Lalita* a head-mounted camera, she burst out laughing. “The way people mount a CCTV camera on a wall, they mounted one on us,” she says.
The 32-year-old had been working at the garment factory on the outskirts of Delhi for nearly a year when management asked workers on her line to strap small cameras to their foreheads before starting their shifts. Nobody explained why.
As Lalita sat stitching shirts and trousers, the camera recorded everything: the rhythm of her hands guiding cloth through the sewing machine; the precision with which she aligned collars and seams; the speed at which her fingers corrected folds and imperfections; even interactions with colleagues. “We found it funny at first, because of how we all looked with that headgear,” she says.
But the atmosphere on the factory floor soon started to change. Worried that their productivity was being monitored, workers became more conscious of their movements. Conversations that would ordinarily unfold across sewing lines grew quieter. Some paid greater attention to their work, wary that every mistake, pause or distraction could be captured on camera.
What Lalita and her colleagues did not know was that their daily routines were being captured as part of a growing effort by companies in India to collect first-hand data from factory floors, information increasingly valuable in the race to automate industrial work.
First-person recordings of human movements and interactions are called egocentric data and are vital for training robots that might one day replace humans on the production line.
Humanoid robots have emerged as the latest frontier in the rapid evolution of artificial intelligence. Industry experts increasingly describe data as the biggest bottleneck in robotics and automation. Unlike large language models such as ChatGPT or Gemini, which were trained on vast quantities of text available online, robots require first-person recordings of physical work.
Companies collecting egocentric footage say the future may require hundreds of millions – and potentially billions – of hours of human activity filmed across factories, warehouses, shops and homes before robots can reliably navigate real-world environments.
EgoLab, an Indian data aggregation company extracting this information from Lalita’s factory in Gurugram, a city in the state of Haryana, counts Tesla among its biggest clients. The company’s CEO, Elon Musk, has predicted that roughly 80% of Tesla’s future value will come not from electric vehicles, but from its humanoid robots.
India is fast becoming a crucial hub in the global race to collect egocentric data. Sensing the opportunity, a growing ecosystem of firms, including Humyn AI, FPV Labs, Micro1, Egodata, Neocambrian, XP Robotics, Objectways, Scale AI and CynLr, has emerged to build data pipelines for robotics companies.
“South Asia remains the workshop of the world for many labour-intensive industries. If you’re trying to teach a robot how humans work, there are few places that offer the same combination of scale, diversity and density of human labour as India. On any given day, millions of workers are sewing garments, assembling products, sorting goods and performing tasks that robotics companies want machines to learn,” says Puneet Jindal, the founder of Labellerr AI, a technology company that collects egocentric data in India.
Capturing the footage is only the first step before recordings are cleaned and annotated for clients, ensuring that hands remain visible, movements are accurately tracked and actions are separated from background activity. India already dominates this business of data annotation: according to industry estimates, the country accounts for about 35% of the global data annotation market, with roughly 60% of its revenues coming from US clients.
Cost is also a major factor. “A company paying $30 an hour for data collection in the US can often get similar work done in India for less than a sixth of that cost,” says the founder of another technology company, who requested anonymity. “In many cases, firms can simply strike deals with factories to collect footage at scale, without directly compensating individual workers.”
The Guardian’s examination of data-collection practices across six factories in five states found that workers wearing devices ranging from meta smart glasses to head-mounted cameras received no compensation for generating footage that would later be sold to technology companies.
“Sometimes they give us a soft drink,” says Lalita, who earns about $200 a month at the factory. “I’m still not sure whether that’s because we’re collecting footage or because Delhi’s heat is unbearable.”
When asked why workers were not being paid separately for generating valuable datasets, several companies argued that factories were already being compensated for facilitating the recordings and that no additional payments to workers were necessary. Critics say such reasoning obscures who is actually producing the data.
“The demand for egocentric data is exploding, and new companies are entering the market every month promising to deliver it more cheaply,” says Jindal. “International clients are often willing to pay significantly more for this footage, but the pressure to undercut competitors keeps pushing costs downward. By the time that trickles through the supply chain, the workers generating the data are often left with nothing.”
Jindal says attempts to compensate workers directly are often resisted by factory owners. “Their argument is that labour costs are already rising and margins are under pressure. They say that if costs increase further, factories may shut down or scale back operations, leaving workers without jobs.”
In some factories, the footage is being used for more than training AI. Records reviewed by Scroll.in found that some companies also generated productivity reports from the recordings, ranking workers based on time spent actively working, estimating losses from “idle” periods, and even tracking how much time workers spent talking to colleagues. In some cases, reports singled out specific workers and identified when and where co-workers gathered to socialise.
Beyond concerns of surveillance and managerial control, Geeta Thatra, a researcher at the Bengaluru-based non-profit Work Fair and Free Foundation says the rapid collection of workplace data is raising difficult questions around privacy. “I’ve heard accounts of women garment workers going to the washroom and forgetting they were wearing head-mounted cameras,” says Thatra. “What happens to issues of safety and privacy in such instances? We have no answers to it yet.”
She says the choice to participate in factory settings is also far from straightforward. “In workplaces where employment is insecure or mediated through contractors, the question of consent becomes extremely complicated,” she says. “A worker may appear to agree to wear a camera, but can they realistically refuse without fearing consequences for their job?”
None of the seven technology companies interviewed by the Guardian said they sought consent directly from workers, instead some stated that permissions were obtained through factory management.
The collection of egocentric data is also expanding beyond factory floors. Several technology companies are now recruiting informal workers – particularly construction labourers, delivery workers and street vendors – to record their daily activities. Unlike factory settings, where payments are often routed through employers, these workers are typically compensated directly through local contractors working with technology firms.
Munazir*, who recently began recording his work as a mason at a construction site in Bengaluru, says he earns between $30 (£22) and $40 a week from the assignments, with payments averaging about $3 an hour. The additional income is significant for him: on most days, he earns less than $8 from his regular work. “The phone feels heavy and uncomfortable to wear,” he says. “But I’ve only just started. Maybe I’ll get used to it with time.”
Although he participates voluntarily, Munazir has little idea what happens to the footage he generates. “I only know that it gives me extra income,” he says. “What they do with the data afterwards, I have no idea.” Companies interviewed by the Guardian acknowledge that workers recording the footage are not told exactly how the data will ultimately be used.
“Traditionally, workers sell their labour for a wage. Here, they are also generating a valuable digital asset,” says Madhumita Dutta, an Ohio State University researcher who studies the relationship between AI, technology and labour. “If they are unaware that their movements, skills and routines are being converted into datasets that can be licensed, sold or used to train commercial AI systems, they have little opportunity to negotiate compensation or object to downstream uses.”
For Sarayu Natarajan, founder of the Bengaluru-based Aapti Institute, which researches the intersection of technology and society, the debate goes beyond consent and compensation.
Unlike conventional workplace data, these recordings capture workers’ bodily knowledge, the movements, instincts and skills accumulated through years of experience. Yet once converted into datasets, that knowledge can be circulated through global AI supply chains.
“The data originates in a worker’s body and actions, but once extracted it no longer remains attached to them in the same way,” Natarajan says.
That, she argues, raises difficult questions about ownership and compensation that current labour arrangements are ill-equipped to answer. Workers, if paid, are typically paid for their time, not for the long-term value that may be generated from the data they produce. As companies build increasingly valuable AI systems on top of such datasets, policymakers may need to consider new mechanisms, from royalties to other forms of value-sharing, that recognise workers’ contributions beyond a day’s wage.
Back at the factory, Lalita continues to stitch collars and seams as she always has. The footage she helped to generate now exists elsewhere: cleaned, annotated and transformed into data.
Asked whether workers should receive a share of the value created from the datasets built on their labour, Lalita laughs. “We are not even getting our full worth for the work we do now,” she says. “Who is going to pay us when we are replaced by robots?”
* Names have been changed
