AI training data: At a garment factory in Karur, 60 of its 200 workers wear GoPro cameras while they work. One of them spends six to eight hours ironing clothes with a camera fixed to his forehead. The recording earns him about ₹10,000 a month in addition to his wages.
The firm can keep using the video long after the worker has been paid. It can copy the recording, sell it to other companies and use it to train machines the worker may never see. His movements may even help machines learn work now performed by people like him.
The worker believes robots will assist humans rather than replace them. The contract offers no assurance about how his recording will be used.
READ | AI in banking carries huge financial stability risks
From India’s BPO workforce to AI training data
This is a new use for India’s old outsourcing workforce. Since the 1990s, multinational firms have sent software support, accounting, call-centre and data-entry work to Bengaluru, Chennai and other Indian cities.
Part of that work later moved into data annotation. Workers label images, audio and video so algorithms can identify what they contain. AI training has now taken annotation beyond computer screens. Workers generate data by recording themselves ironing garments, handling tools or performing other physical tasks.
The recording remains usable after the shift. A worker can perform only one task at a time, but a video of that task can be copied at negligible cost and used by several firms. One physical action becomes a reproducible digital asset.
Workers may receive ₹250-₹450 an hour for recording, in addition to their wages. The payment covers the time spent performing the task. If the buyer owns the resulting video, it can license the file, combine it with other datasets and use it in later training projects. The worker receives nothing from those transactions.
The financial gain from AI does not accrue to Silicon Valley companies simply because they are greedy. The nature of the asset favours its owner. Human time is limited. Digital files can be reproduced and reused.
READ | AI needs a knowledge dividend for the human work it uses
AI data jobs bring income but little ownership
Describing this arrangement as exploitation ignores the choices facing workers. Young women and first-generation workers often have few formal job opportunities. Data collection can supplement low wages and give workers an income of their own.
Those benefits are substantial. They do not determine who should own the recording.
A worker may earn ₹10,000 more each month while giving a company an asset with a commercial life far longer than the labour used to create it. Current contracts value the hours spent recording. They rarely give workers a claim over future use.
The inequity lies in the division of returns. Workers may be paid reasonably for the day’s work and remain excluded from the revenue generated after that day. The owner of the recording receives whatever later value it produces.
READ | AI command centre can plug cyber governance gaps
Can AI data workers receive royalties?
Karya, an Indian non-profit, has tried a different model. It pays workers more than standard data-work rates and promises royalties when a dataset is resold. Workers can receive another payment when the material they created earns money again.
The model remains small and royalty payments are thin. It also faces resistance from commercial firms, which prefer to buy data outright rather than track continuing claims.
Tracking becomes harder once videos are combined with large datasets and converted into model weights. A one-time payment removes the administrative burden of identifying which worker’s data contributed to a model or product.
That burden helps explain the current model. It also leaves all later value with the owner of the file.
Data-collection jobs can continue under contracts that distinguish payment for recorded work from rights over later reuse. Karya shows that some form of continuing payment is possible, although its model has yet to operate at the scale of commercial data firms.
India’s AI economy is turning ordinary labour into reusable digital property. Workers are now paid once to create assets that firms may use many times. The policy question is whether they should retain any claim on those assets.
Gopikrishnan Annamalai is a research student and Dr Kavya Sanjaya is Assistant Professor at the Department of Economics, CHRIST University, Yeshwanthpur Campus, Bengaluru.