The Workers Teaching AI Robots to Replace Themselves: What Builders Need to Know About India's Data Labeling Pipeline

Robotics and AI companies are quietly building a critical part of their training pipelines on the backs of low-wage workers in India, who wear head-mounted cameras to film their everyday tasks. The footage trains models to recognize and manipulate objects in real environments, but the workers often do not know what the data will be used for. For AI builders, this reveals a hidden layer of the data supply chain that carries real ethical and operational risk.

The hidden data pipeline behind robot training

Companies competing to build general-purpose robots need vast amounts of real-world human activity data. Their solution: recruit workers to perform routine tasks while recording every hand motion with a smartphone strapped to their foreheads. The resulting first-person video is fed into specialized AI models that learn to imitate human behavior Bloomberg.

In one example, a housewife in Tamil Nadu films herself slicing mangoes for just over two dollars per hour Economic Times. The work is mundane, but the data is invaluable for teaching machines how to move like humans in the real world.

A New Delhi recycling worker's day on camera

Sunita Rathore, a 43-year-old plastic-recycling worker in New Delhi's Karan Vihar colony, earns about 20,000 rupees ($211) per month sorting and cleaning discarded plastic. Now she also wears an iPhone strapped to her head, angled to capture every motion of her hands as she strips labels off plastic bags and stacks sacks for reprocessing. For wearing the device, she earns an extra 150 rupees per hour. She has no idea what the footage is for or that her work is critical to the future of AI and robotics Bloomberg.

Why this matters for AI builders

This is not just a labor story. For anyone building AI products that rely on real-world training data, the quality, consent, and governance of that data are becoming core product risks. If your model was trained on footage collected without informed consent, or if the data pipeline relies on exploitative labor practices, you face regulatory, reputational, and legal exposure down the line.

Builders should ask: Who collected the data? Under what terms? Were workers informed of how their recordings would be used? The answers are often murky. The Bloomberg report notes that Rathore shuts off the camera if anyone calls out to her, and capturing faces is forbidden. But the footage of her hands and workspace is still being used to train systems that could eventually automate her job.

The consent and compensation gap

The ethical dimensions are stark. Workers earn a small premium for participation, but the long-term value of the data far exceeds their compensation. More importantly, the lack of transparency about how the footage will be used means workers cannot give meaningful consent. This is a pattern that has played out before in AI data labeling, but the stakes are higher when the data captures physical tasks in real environments.

For AI builders, the practical takeaway is clear: audit your data supply chains. Demand transparency from data vendors about how training data was collected, and consider whether the practices behind your training data align with your company's values and regulatory obligations. The cost of ignoring this pipeline is not just ethical, it is a business risk that compounds as scrutiny increases.