Are Your Suppliers Forcing Workers to Train Their AI Replacements?

Factories across the world are witnessing a fundamental shift in the relationship between workers and their employers. Since the first decades of industrial production, owners have struggled to extract more value from the time and labour a worker provides, in exchange for a wage. But the worker also generates a quieter kind of value for themselves: an innate knowledge of, and instinct for, their work.

Consider a seasoned pilot at the Panama Canal, who brings vast cargo ships through its narrow locks as though it were second nature. That skill is highly marketable, and the pilot can trade on it for better pay, better conditions or new work.

The long appropriation

It is in this ephemeral category of knowledge and skill that a long game of cat and mouse has been played. Charles Babbage recognised in 1832 that a task requiring expensive skill could be broken into its constituent parts and staffed by cheaper labour. Frederick Winslow Taylor's time studies showed that the craftsman's skill could be articulated into a process sheet. Harry Braverman, in Labor and Monopoly Capital, called this the separation of conception from execution: the worker was present on the factory floor not to conceive, but only to execute.

Yet every one of these projects was limited by its own method, because skill could only be captured once it had been articulated. No planning sheet could capture the instinct of a worker who, after decades with a machine, knows when it is about to break down. Michael Polanyi called this tacit knowledge, the practical understanding a skilled person holds without being able to state its rules. The mule spinner reading the behaviour of the yarn and the moulder judging when the sand was right each kept something the manager could not write down. That residue was the substance of the worker's bargaining position. It made them expensive to replace and able to insist that their pay reflect their years on the job.

The economist David Autor named this limit Polanyi's paradox: we cannot automate what we cannot articulate. Yet even as he named it, he was careful to note that a system trained on examples, rather than instructed by rules, need never be given the articulation at all.

The portrait upstairs

Consider, then, what becomes possible if that residual value could be captured without ever being articulated. Imagine a worker whose skill was installed rather than learned, competent on the first day at a job that once took a decade to master, and reproducible without limit. Robotics firms now promise exactly this, wrapped in a picture of a post-work world. Why work when machines can work for you? Why not live as the aristocracy once did, concerned only with sex, play, thought and art? A utopia indeed, and a very old one.

Oscar Wilde's Dorian Gray kept his flawless face in public while his portrait rotted in a locked room upstairs, his secret safe only while no one looked. The post-work utopia has the same structure: a surface of abundance and liberated pleasure, with the extraction that drives it kept firmly out of view. That extraction needs raw material, and the raw material is data. The perfect robotic worker cannot be coded into existence; it must be trained, and the training has turned the worker's tacit knowledge into the instrument of their own obsolescence.

Wearing the Panopticon

"Surveillance is permanent in its effects, even if it is discontinuous in its action." Michel Foucault, Discipline and Punish

In May 2026, Scroll's Ayush Tiwari and Raghav Kakkar reported that workers at the Gurugram unit of Pearl Global Industries, an export garment manufacturer, had been given head-mounted cameras to wear through their shifts in April. The cameras, angled to film the workers' hands, belonged to Egolab.AI, a startup that sells "labour-sourced egocentric video footage" packaged for companies building robotics and autonomous systems. Workers told Scroll that nobody had asked for their consent, either verbally or in writing. Egolab's head of business operations told the reporters: "We took consent to use the device and we did it in our own way."

Egocentric video bypasses the process sheet entirely, recording the output of the worker's instinct at the source, and seeks to breach at last the defences of tacit knowledge.

Just enough of its shadow

Whether it succeeds is still a genuinely open question. A skilled machinist or tailor relies on touch, on the heat of a machine, on shifts in sound and on the resistance of the material, and none of this reaches a lens. A network trained on millions of hours of footage may reproduce a skilled worker's hand movements with considerable fidelity while remaining blind to the haptic reality of the craft. Statistical regularity is not the same thing as embodied intuition.

For the purveyors of algorithmic capital, however, the distinction hardly matters. They have no need of the worker's instinct itself, only just enough of its visual shadow to convince the market that the worker has become replaceable. Polanyi's paradox survives intact while losing every practical consequence it once carried. What has changed is not the epistemics of skill but the economics of plausibility, because the investment decisions that follow will reshape the factory floor whether or not the instinct was ever captured.

Is data really labour?

The "data as labour" camp gives this appropriation a name, and on the factory floor its case is stronger than on any social media platform. The worker is already doing a paid task, and the employer adds a layer of capture that turns practical knowledge into a reusable asset. Yet data as labour must not become a licence for unlimited surveillance. If workers can earn more only by allowing ever more intimate measurement, a right to payment could legitimise the very capture they should be able to refuse. In a hub like the National Capital Region, where unions are thin and a worker who refuses the camera can be replaced by the end of the shift, a royalty risks becoming a compulsory piece rate for being watched.

Individual ownership is also a poor fit for footage like this. A movement caught on camera reflects one pair of hands, but it also depends on coworkers, supervisors, tools and the organisation of the whole line, and the recording captures their conversations and bodies too. That is why the researcher Astha Kapoor told Scroll that the answer lies in "collective community muscle". "Otherwise," she said, "we are doing the work that is supposed to make us redundant."

What due diligence law already asks

Workers need not wait for property law to decide whether a gesture belongs to them. The UN Guiding Principles on Business and Human Rights already expect companies to identify, prevent and mitigate the harms linked to their operations and supply chains, and that expectation is now written into binding law. The EU's Corporate Sustainability Due Diligence Directive lists the right to privacy, just and favourable working conditions and freedom of association among the rights companies must respect across their chains of activity. From July 2029, a European brand above the directive's thresholds that buys from a factory filming its tailors will have to weigh that practice as a potential adverse impact, not dismiss it as a supplier's private business.

Some of these tools are already sharper. In March 2026 a Paris court ordered the parent company of the Yves Rocher group to pay damages under France's vigilance law for union-related dismissals at its Turkish subsidiary, proof that collective labour rights abroad can be enforced in a European court. Norway's Transparency Act lets any person ask a covered company how it handles adverse impacts in its supply chain, so a union or a journalist can simply ask a Norwegian retailer whether its suppliers film their workers.

There is an irony here too: the CSRD value chain cap adopted this summer shields small suppliers from their buyers' data demands, while nothing shields workers from their employers' demands. Supply chain codes of conduct already ask whether fire exits are unlocked and child labour is absent, and they should now ask the same question of the camera. A factory that makes its workers train their robotic replacements, without collective consent, transparency about purpose or transition guarantees, should fail an audit as surely as one with unsafe machinery.

These frameworks drift towards corporate self-reporting, and this year's omnibus stripped the CSDDD of its EU-wide civil liability. A rights standard must therefore be wedded to collective bargaining, with worker representatives entitled to know what is captured, why, and for whom, and with the power to refuse it altogether.

From consent to power

The workers in Gurugram have already shown what that power looks like. The Wire reported in July that tailors pulled the memory cards from the headsets, and kept taking the devices off until the footage was useless. After two weeks the experiment was quietly shelved. One worker put it plainly: "If I knew it was for training a robot, I would prefer not to wear it."

Treating data generation as labour is a necessary beginning, because it makes a hidden contribution visible, but it is not sufficient if it turns surveillance into another compulsory job. The fairer principle is collective: workers should have information, bargaining power, the ability to refuse unjustified capture, and a durable share in the gains built from their work. Models such as the worker-owned cooperatives proposed by Sriraman, Bragg and Kulkarni in 2017 show that such a share can outlast the moment of collection. If the workforce supplies the blueprint for an automated future, it should not be asked to finance that future with its own dispossession.