Factory workers in India are being asked to wear GoPro cameras while they work, so that the recordings can be used to train robots, according to an NPR report published today.

The arrangement is a clean illustration of something that is usually abstract: what exactly is being bought when a company buys training data.

Why video of hands is valuable

Language models learned from text that already existed on the internet. Robots cannot learn that way, because there is no comparable archive of physical work.

The problem robotics researchers describe is that physical tasks carry enormous quantities of knowledge nobody has written down. How much pressure to apply to a component before it seats. How to orient a part that is nearly symmetrical. What to do when a piece does not fit on the first attempt. Workers know these things through practice and generally cannot articulate them if asked.

A head-mounted camera captures it anyway. Hours of first-person footage of skilled hands doing a repetitive task is among the most useful training material available for machines meant to do the same task, and it does not exist unless somebody records it.

The part that makes it uncomfortable

The stated purpose is automation of the work being recorded. The people generating the data are the people whose jobs the resulting systems are intended to perform.

It is worth being careful about how strongly to put this. Workers participating in a programme that may reduce demand for their own labour is not new — it describes most process documentation and time-and-motion study of the past century. What is different is directness, and the fact that the expertise transferred is tacit knowledge the worker could not have written into a manual.

The questions that follow are concrete rather than philosophical. Is participation genuinely voluntary, in a workplace where a supervisor is asking? Is it paid separately, or treated as part of the job? Who owns footage of a person’s own hands, and can it be sold on? And what happens to a worker who declines?

The case made for it

Those running such programmes make arguments that deserve stating rather than dismissing.

The automation is coming regardless, and a country that hosts the data collection captures some of the economic activity rather than none. The recordings are additional paid work. Robots in practice take over the most repetitive and physically damaging tasks first, which is not purely a loss to the person doing them. And manufacturing employment has been reshaped by machinery continuously for two centuries without producing the permanent mass unemployment repeatedly predicted.

The counter is about distribution rather than totals. Economists who study automation generally agree that aggregate employment adjusts and that the adjustment falls unevenly — on particular workers, in particular places, within a working life too short to wait out a generational transition. The person wearing the camera is squarely in that group.

Why India

Data collection of this kind concentrates where labour is cheaper, as data labelling did before it. Much of the annotation work behind current AI systems has been done in India, Kenya, the Philippines and elsewhere, under conditions that have been the subject of considerable reporting.

The pattern holds here. The expertise being captured is Indian; the systems being trained on it will mostly be owned elsewhere. Whether that is exploitation or ordinary trade is the argument, and it is the same argument that has run through every stage of this industry’s supply chain.

Sources

• NPR, These workers help train robots who could one day steal their jobs, 14 October 2026
• Published robotics literature on imitation learning and demonstration data
• Reporting and research on data annotation labour in India, Kenya and the Philippines
• Economic literature on automation, labour displacement and distributional effects