Workers worldwide are using head-mounted cameras to sell recordings of their physical movements—such as folding clothes, making beds, and cooking—to Silicon Valley robotics firms for $15 per hour. These individuals, based in countries like India, Nigeria, and the Philippines, are contributing unspoken "tacit knowledge" to humanoid robots under development by companies like Tesla, often without clear awareness of how their data will ultimately be used. This form of "ghost work," once confined to digital screens, is now extending into the physical realm.
You may have come across footage from a garment factory in southern India, where employees wear head cams to capture hand movements for AI system training. The rising demand stems from companies such as Tesla and Figure AI racing to build humanoid robots, creating a severe shortage of real-world movement data.
To address this, Palo Alto-based Micro1 has enlisted roughly 4,000 workers across 71 countries, collecting over 160,000 hours of video each month. Participants are required to submit at least 10 hours of footage weekly, performing a variety of tasks in rotation. Other firms like Scale AI and Encord are also building data-gathering teams, while DoorDash introduced an app in March 2026 allowing delivery drivers to record household chore videos at home—though it excluded states with strict data privacy laws.
The job itself is stranger than it sounds. Applicants first undergo an AI-powered interview with an agent named Zara, which evaluates suitability and requests a sample video. Those accepted receive a forehead mount, recording guidelines, and a task list. Instructions emphasize keeping hands visible and moving at a "natural speed," though many workers report having to slow down deliberately, making actions feel unnatural, almost like sleepwalking.
A significant barrier is the requirement for an iPhone equipped with a LiDAR sensor, meaning at least an iPhone 12 Pro or newer model. Submitted videos undergo both AI and human review, with only about half ultimately accepted. Rejections may stem from poor lighting, hands leaving the frame, overly quick motions, or inappropriate background elements. Workers are paid hourly, but rejected footage means unpaid labor. Approved videos then enter an annotation phase where frame-by-frame labeling of actions, objects, and motion trajectories is performed by another set of workers.
Arjun, a tutor from New Delhi, shared that he often spends an hour brainstorming to generate 15 minutes of recordable household content. Micro1 encourages varied tasks since diverse scenarios are crucial for training, but home environments offer limited possibilities, and creativity eventually runs dry.
Notably, videos from U.S. households command higher value. According to Ravi Rajalingam, founder of data annotation firm Objectways, robot developers assume American consumers will be early adopters, making U.S. home environment data more valuable—sometimes tripling the hourly wage compared to workers in Vietnam or India.
Micro1’s Vice President Arian Sadeghi noted that 160,000 hours of monthly footage is far from sufficient, estimating a need for billions of hours. "We haven’t even started collecting human-to-human interactions; this is just basic chores," he said. At current rates, gathering billions of hours would take roughly 10,000 years of continuous work.
The concept of "ghost work" was explored in a 2019 book by anthropologist Mary Gray and computer scientist Siddharth Suri. They described often-invisible human labor—such as image labeling, content moderation, and data cleaning—that makes AI systems appear intelligent. Gray recalled engineers being uncertain or unwilling to examine who performed such tasks.
Previously, ghost work involved screen-based actions like clicking and filtering. Now, physical movements—folding gestures, cooking rhythms, refrigerator door openings—are becoming raw materials to be harvested, priced, and sold. These materials flow from ordinary homes in India, Nigeria, the Philippines, and Kenya to firms in Palo Alto and San Francisco, eventually transforming into commercial products.
Researchers Nick Couldry and Ulises Mejias frame this dynamic as "data colonialism," where tech companies’ data appropriation mirrors historical resource extraction, turning daily life into a capital resource. In Micro1’s case, $15 per hour is competitive in Nairobi or Manila, yet negligible compared to billions invested in robotics. Information asymmetry is another concern: Micro1 withholds client details citing confidentiality, leaving workers unaware of data storage or resale practices. Though paid and under contract, laborers remain at the information tail end of the supply chain.
Gray observed that ghost workers often form informal support networks, as the job itself offers little backing, leaving isolation as a default condition. By 2026, the global humanoid robot market is projected to reach $4.23 billion, with Tesla and others planning mass production to exceed 100,000 units by 2027. These robots may eventually perform manual labor in factories and homes, trained on data provided by those currently relying on physical work for livelihood.
Philosopher Michael Polanyi wrote in his 1958 book "Personal Knowledge" that "we can know more than we can tell," terming this "tacit knowledge"—understanding embodied in actions, perceptions, and intuition rather than explicit rules. Riding a bicycle, for example, involves balance know-how that can’t be fully written down but is learned through practice.
Though Polanyi wrote before AI’s rise, his ideas gain new relevance today. Efforts are underway to extract such tacit knowledge from human bodies into machine-processable data. What Micro1’s forehead cameras capture isn’t just a clothes-folding action, but also how fingers sense fabric weight, wrists turn at just the right moment, and eyes track cloth edges.
Scale AI has announced collecting over 100,000 hours of such material. This marks a first-in-history attempt to externalize bodily knowledge on a massive scale. Polanyi argued that tacit knowledge can’t be fully articulated, but that doesn’t make it immune to extraction. Couldry and Mejias note that data colonialism treats daily life as a ready-to-harvest resource—now including even bed-making.
While AI’s impact is often framed as machines replacing knowledge workers, the most mundane physical actions are now being mined for data. If such acts can become training material, the question of "what constitutes human labor" shifts from philosophical debate to urgent political reality.
Zeus, a medical student in a central Nigerian city, straps his phone to his forehead after work to record himself making his bed. He sees it as a "chance to leave a mark," feeling part of something significant. That may be true, yet it doesn’t change the fact that his "mark" will take the form of motion data, purchased by a company he can’t name, to train a machine he may never afford.
Polanyi held that all knowledge is personal, rooted in specific people, contexts, and practices. To strip such knowledge from individuals and sustain its operation independently raises a pressing question: What do people, as knowledge bearers, truly own? This question may lack an answer for now, but it is being quietly posed—for $15 an hour—in apartments in Nigeria, kitchens in India, and yards in the Philippines.
Comments