A Bengaluru startup, Nexxus AI, has recruited around 400 Indian workers since 2024 to film themselves washing dishes, folding clothes and tidying rooms using smartphones and head-mounted cameras. The footage trains home-assistant robots for clients in the United States, Europe and Japan. Pay for comparable data work in India runs roughly INR 150–300 per hour, below the formal-sector median wage for educated workers.
India hosts about a fifth of the world’s online gig workforce, yet no law specifically protects data-annotation labour. The robots these workers train are built for households they will likely never afford.
The work looks simple. A worker straps a camera to their head, walks into a kitchen, and washes a stack of plates the way they would on any ordinary day. Then they do it again from a different angle. Then with different plates, in a different light, in a different home.
What they are building is a map of human motion precise enough for a machine to copy. Every wipe, grasp and sort becomes a labelled example feeding the next generation of household embodied AI. The pay for this work in India sits around INR 250 (about US$3) an hour by reported rates, a figure consistent with what platform studies record across the country’s data sector.
Here is the part the demos leave out. These workers are not training a tool that helps them do their jobs faster. They are training a replacement, frame by frame, for the kind of domestic and service labour millions of them depend on. The arbitrage is the story: the Global South is being paid subsistence wages to automate away its own economic future.
The dataset is the product, and it is being mined cheaply
Nexxus AI’s model is direct. Workers record household chores; the company packages that footage and sells it to robotics developers abroad. The technical reason this work exists is that a robot cannot learn to clear a messy table from a rulebook. It needs thousands of examples of humans doing it, in conditions varied enough that the system generalises rather than memorises.
That requirement is what makes the labour repetitive and the volume high. Camera angles, backgrounds and objects all have to change so the model does not overfit to one kitchen. A worker may film the same task dozens of times in a shift.
The scale sits well beyond one startup. Scale AI, the US data-labeling firm, reported working with more than 500,000 remote contractors globally by mid-2024, many of them in India annotating and collecting data for AI systems. India’s broader online gig workforce reached roughly 3 million in 2023, about a fifth of the global total, according to the International Labour Organization’s World Employment and Social Outlook.
Mary L. Gray, senior principal researcher at Microsoft Research and co-author of Ghost Work, argues that this invisible labour is foundational to AI and routinely runs under precarious conditions without recognition or protection. Her point lands harder here than in earlier annotation work. Earlier datasets used boxes drawn around objects in static images. This is full-motion video of people’s homes, faces and routines.
The twelve-to-eighteen-month implication is the one most coverage misses. Whoever controls these data pipelines shapes which robot models become global standards — and right now that control is being assembled cheaply, in homes that will never see the product.
India supplies the data while the value sits elsewhere
The hardware race is happening somewhere else. In the US, Figure AI and Tesla are building general-purpose humanoids; Figure has tied itself to OpenAI and to factory pilots with carmakers. Japan’s Toyota Research Institute targets assistive home robots for an ageing population. Chinese firms like UBTech and Fourier Intelligence iterate low-cost models on vast domestic datasets.
India leads none of this on hardware. It is positioning itself as the data and software hub feeding all of it. That division of labour is the structural force behind the wage gap.
Kate Crawford, research professor at USC Annenberg, describes data workers in the Global South as part of an extractive AI supply chain where labour is cheap and value capture concentrates in Western firms. The footage itself complicates the picture further.
Sundar Rajan, general secretary of the All India Gig Workers Union, warns this is a race to the bottom: tasks get subdivided and handed to a large pool of unorganised workers, pushing wages down. The honest limit on the story is that the INR 250 figure is a reported rate, not an audited one, and surveyed pay ranges sit wide. Either way, the worker filming the chore earns once. The model trained on that footage earns indefinitely, for someone else, in a home the worker helped build but cannot enter.
Beyond the headline
The human cost
For the workers filming their chores, this is not a story about futuristic robots but about stitching together enough micro-tasks to cover rent and food in cities where formal jobs are scarce. Their days are split into quotas and quality checks, with little clarity on how long the contracts last or what happens when the robots get good enough that the gigs vanish.
The money trail
Most of the upside from these projects will land where the intellectual property sits: venture funds, chip suppliers and robotics firms in the US, Europe, Japan and China. The chain turns Indian homes into low-cost capture sites while margins and brand power concentrate with the firms that own the models, hardware and cloud, hardening an unequal split of AI profits.
What isn’t being said
Public talk about home robots fixes on consumer convenience and slick demos, not the cross-border flows of intimate household data or the bargaining power of the people producing it. Nearly absent from corporate roadmaps is any binding promise to share productivity gains with data workers or to let them decide how recordings of their lives get reused.
Where the legal exposure actually sits
With India’s detailed gig-worker rules under the Code on Social Security expected late in 2026, the people commissioning, regulating and studying this work face concrete decisions now.
- AI and robotics investors
If you fund projects that outsource data collection to India, read the Code on Social Security, 2020 at labour.gov.in to assess whether your contractors classify as gig workers and what social-security duties attach. The pending 2026 rules could add compliance costs to any India-based capture pipeline within the year.
- Compliance and data-protection officers
Footage of Indian interiors and faces moving to Western servers triggers consent and cross-border transfer obligations under India’s DPDP Act 2023. Review the Ministry’s official summary at meity.gov.in before signing data deals that include children, biometric identifiers, or background audio.
- Labour researchers and policy advocates
Track the IT for Change findings on India’s data-annotation pay and account-deactivation practices to benchmark conditions in this specific sub-sector. The late-2026 social-security rules are the first real test of whether AI-training labour gets pulled inside India’s gig-worker framework.
Explainer
- Embodied AI
- AI built to perceive and act in the physical world rather than only process text or images. It learns from large volumes of annotated video showing humans performing physical tasks, then maps its own sensor data onto those patterns. For household robots, the bottleneck is not computing power but the supply of varied real-world footage, which is why data collection has moved to lower-wage labour markets.
- DPDP Act 2023
- India’s Digital Personal Data Protection Act, the country’s first comprehensive personal-data law, enforced by the Data Protection Board of India. It sets consent and purpose-limitation rules for processing digital personal data, including cross-border transfers. It does not address labour standards or collective rights for the platform workers who generate AI-training data, leaving that gap to separate gig-worker legislation.
- Scale AI
- A US data-labeling firm that supplies annotated training data to AI developers, including robotics teams. By mid-2024 it reported working with more than 500,000 remote contractors worldwide, a large share of them in India. Its scale shows how data work has become an industrialised supply chain feeding Western AI models, with the labour distributed far from where the resulting products are sold.
RELATED STORIES
