The short version
- Companies are hiring thousands of workers to film their hands and movements while performing routine tasks like hair cutting or manufacturing.
- The recorded video data is sold to technology firms developing physical AI that can control robots in complex, real-world environments.
- Experts warn that while the work provides income, it may eventually contribute to the automation of the very jobs being filmed.
A quiet but expanding sector of the global technology economy is emerging from everyday labor. Across India and other regions, workers are being paid to strap cameras to their heads or wear sensors on their hands while they perform routine tasks. This footage is not for entertainment or social media; it is raw material for training artificial intelligence systems designed to control physical robots. The industry relies on the premise that machines cannot learn to navigate the messy complexity of the real world solely through computer simulations.
Mohammed Ashim, a twenty-two-year-old barber in Bengaluru, illustrates this new dynamic. For four years, he has cut hair and trimmed beards, but recently his routine has changed. He now wears a head-mounted camera for five or six hours daily, ensuring his hands remain within the frame while keeping client faces blurred. The work requires discipline regarding lighting and movement, but Ashim notes that the compensation is good and the futuristic nature of the task keeps him engaged. He never imagined his razor techniques would become valuable assets for global tech companies.
Ashim is one of hundreds of thousands of individuals worldwide participating in this data collection effort. Companies like Rashdan Technologies aggregate this human-generated video, processing it before selling it to research organizations and technology firms. The demand spans a wide variety of sectors, including industrial tasks, agriculture, household chores, and small business crafts such as pottery or garment making. Co-founder Shahil Baig explains that the company goes wherever specific tasks are performed, paying between five and ten dollars for eight to ten hours of recorded work.
The technical requirements for this data are becoming increasingly sophisticated. Early efforts focused primarily on capturing hand movements, but clients now demand more comprehensive coverage. Some requests involve rigs with four to six cameras positioned around a person’s head to capture the entire scene rather than just the hands. This shift makes consent and comfort more challenging to secure. Baig notes that convincing someone to wear a single small camera is easier than asking them to don a multi-camera headset, though higher payments are offered for this more intrusive tier of data collection.
Manish Agarwal, founder of Humyn Labs, argues that high-quality signals involving limb movements and dynamic force application are essential for physical AI. Unlike digital tasks, physical work requires understanding how much force to apply to different objects, such as gripping a plastic bottle without crushing it versus holding a steel glass. Until recently, no internet archive existed for embodied human activity, meaning data on cleaning kitchens or assembling cars had to be captured from scratch in diverse real-world environments.
Humyn Labs currently employs ninety-six thousand data collectors across Latin America and Southeast Asia. The company filters thousands of hours of supplied video, selecting the most useful content and labeling it so AI systems can interpret what they are seeing. Agarwal emphasizes the importance of geographic diversity in sourcing. If the goal is to deploy robots in US factories or homes, the training data must reflect environments that serve as proxies for those settings, rather than relying solely on residential contexts from India.
The industry is becoming more crowded, with expectations of dozens of new competitors emerging in India alone this year. As the market expands, the tension between economic opportunity and long-term labor implications grows. Bino Paul, a senior professor at the Tata Institute of Social Sciences, suggests it is too early to determine whether AI will serve as a useful tool or lead to widespread automation for these workers.
Paul warns that if machines eventually take over these roles, workers may have little recourse by the time they realize what they have participated in. The concern is not just about job displacement but about the lack of regulatory frameworks to protect laborers who are inadvertently training their replacements. Paul advocates for government regulation of AI, arguing that it cannot be viewed merely as a technological advancement without considering its profound impact on the workforce and social structures.
Sources behind this briefing
Go to the original reporting
- BBC Business↗The Indian workers filming their work to train AI
