The Workers Training the Robots That Could Replace Them

In New Delhi, women sort plastic packaging with iPhones strapped to their foreheads, cameras angled down to capture every movement of their hands.

The footage is not for supervision. It is training data. And the machines being trained are designed to eventually do the work those women perform.

What the Investigation Found

A Bloomberg investigation published this week documented thousands of workers across India being paid to film their own labour from a first-person perspective for robotics companies.

The task is as literal as it sounds. Workers wear cameras—often ordinary iPhones strapped to their heads—to record the precise movements of their hands: flicking a blade to remove labels from plastic bags, stacking sacks for recycling, stitching shoes, welding steel. Robotics companies are competing to collect this footage.

The pay: roughly $2.62 an hour.

Why the Data Matters

The technical term is egocentric data: video captured from the perspective of the person performing a task rather than from an observing camera.

For robotics, that perspective is valuable because the hardest problem is not simply making a machine move. It is teaching the machine what a task actually requires: how much pressure to apply, at what angle, in what sequence, and how to react when a material behaves unexpectedly.

Much of that knowledge is tacit. Workers know how to do these things without necessarily being able to explain every adjustment they make.

First-person video captures those movements directly. With enough examples, a model can begin learning to reproduce them.

That is why robotics companies are willing to pay for footage of tasks that may look mundane from the outside.

The Consent Question

This is where the story becomes more than an economic curiosity.

According to the reporting, many workers were not clearly told that their footage would be used to train automation systems targeting the same categories of jobs they perform. India's IT Ministry has reportedly raised questions about whether consent in these arrangements is adequate.

The distinction matters.

India's established data-labelling industry involves workers tagging images, text and other information for AI systems. Here, the workers themselves are the data.

Their movements, bodies and workplaces become the raw material. And the stated purpose of collecting that raw material is to teach machines to reproduce the labour those workers currently perform.

The Bigger Picture

India's place in the global AI economy has long depended on supplying labour to it—from data annotation and content moderation to IT services.

That model is already under pressure. The combined market capitalisation of TCS, Infosys, Wipro, HCL Technologies and Tech Mahindra has fallen more than 46% since August 2024, as generative AI increasingly automates coding and support work. Nearly 50,000 job cuts announced globally in 2026 have been explicitly attributed to AI.

The Bloomberg reporting documents the same dynamic moving into physical labour.

Only now, the mechanism is strikingly literal: the workers themselves are recording the instructions for their replacements.

None of this makes the robotics work inherently illegitimate. Training data has to come from somewhere, and workers should be paid for their time.

But at $2.62 an hour, particularly when the ultimate purpose of the footage may not have been clearly disclosed, the terms of that exchange deserve scrutiny.

A worker is being paid a few dollars an hour to surrender the one thing that makes their labour valuable: the knowledge of how to do the job.