Indian Workers Record Footage to Train AI

AFP photographed and interviewed workers across Tamil Nadu, including Chennai homemaker Nagireddy Sriramyachandra, who is paid about 250 rupees (roughly $2.6) for an hour of first-person video recorded for data firm Objectways; the wire report was carried in full by Dawn and Al Jazeera (both bylined AFP) on June 11, 2026. Objectways operates studios where workers repeat household tasks, folding towels, making coffee, arranging objects, in front of GoPro cameras and smart glasses, and the firm works with Amazon SageMaker and lists Fortune 500 clients, per Dawn's reporting. Separately, an April viral video of garment workers wearing head-mounted cameras on a factory floor prompted an investigation by Scroll.in (published May 20), which identified the location as the Gurugram unit of Pearl Global Industries and traced the hardware to Egolab.AI, a startup founded in January 2026 by two teenagers, later acquired by Build AI. Workers told Scroll.in they were not asked for written or verbal consent.

Scroll.in reports that Egolab's head of business operations, Arnav Kabra, said the firm "took consent to use the device and we did it in our own way" without elaborating, while Pearl Global did not respond to questions. Astha Kapoor of the Aapti Institute told Scroll.in that egocentric footage falls outside India's Digital Personal Data Protection Act because it lacks a direct one-to-one mapping between a worker's data and an individual harm, arguing the issue instead needs to be addressed as a collective data right. That gap is the crux of the practitioner risk: a dataset can be commercially packaged and resold to "global AI companies building robotics, computer vision, and autonomous systems" (Egolab's own description, per Scroll.in) without the kind of individual consent record that would satisfy a Western data-protection review.

This is not an isolated vendor. Scroll.in cites industry estimates that robotics labs will spend $1.5 billion to $50 billion over the next two to three years to acquire 100 million to 1 billion hours of egocentric data, and reports Build AI's CEO's claim of $22 million in venture funding (confirmed by one investor, HF0; two others had not listed the investment as of publication). Gizmodo separately reported that Human Archive, a competing data-collection startup operating in India and elsewhere, raised an $8.2 million seed round from Y Combinator and investors linked to OpenAI, Nvidia and Google in May 2026, evidence that the same collection model is attracting mainstream venture capital rather than being one company's practice.

• •Whether any of the robotics companies named in vendor marketing (Tesla, Boston Dynamics, Figure AI, per Egolab's own claims) confirm sourcing this specific data.
• •Documentation practices: license terms, consent records, and chain-of-custody metadata that would make downstream datasets auditable.
• •Regulatory movement in India on collective data rights for workplace-collected biometric and video data, an area the Aapti Institute and others say current law does not cover.
• •Whether buyers of egocentric datasets begin requiring documented consent and provenance as a purchase condition, similar to model-card or dataset-card practices elsewhere in ML.