New Gig: Train AI Using iPhone-Mounted Cameras for $15/Hour

You may have come across a video from a garment factory in southern India.

Factory workers wear head-mounted cameras to record hand movements for training artificial intelligence systems.

This is because, as companies like Tesla and Figure AI compete to develop humanoid robots, the real-world motion data needed to train them has become extremely scarce.

Micro1 in Palo Alto has therefore recruited approximately 4,000 workers across 71 countries worldwide, receiving over 160,000 hours of video footage each month. Each worker submits at least 10 hours of video per week, alternating between different types of tasks.

Scale AI and Encord are also hiring their own data collection teams, and DoorDash even launched the Tasks app in March 2026, allowing its delivery drivers to record household videos at home as a side task—though it explicitly excluded states with strict data privacy laws.

The specific operations of this job are stranger than they sound.

Candidates must first undergo an interview with an AI agent named Zara. Zara will converse with you to assess your suitability and request a trial recording video.

After approval, you will receive a headband mount, a recording guide, and a task checklist. The instructions state that your hands must remain visible in the frame at all times and that your movements should be at a "natural pace."

The natural speed often appears too fast under the camera, so workers commonly report that they must deliberately slow down during recording, resulting in movements that feel unnatural, as if mimicking sleepwalking.

There’s another requirement: you need an iPhone with a LiDAR sensor, which means at least an iPhone 12 Pro or newer.

After submitting the video, it undergoes dual review by AI and human reviewers, and only about half of the submissions are ultimately approved. Rejection reasons may include insufficient lighting, hands moving out of frame, movements that are too fast, or unauthorized objects appearing in the background.

Workers are paid by the hour, but if a video is rejected, their labor during that time is wasted. Videos that pass review then enter an annotation process, where another group of human annotators label action categories, object names, and motion trajectories frame by frame.

Arjun, a tutor in New Delhi, says it typically takes him an hour to brainstorm enough household tasks to fill a 15-minute recording. Micro1 requires workers to constantly “vary the content,” as diverse scenarios are crucial for training effectiveness—but homes are limited in size, and creativity eventually runs out.

Videos from U.S. households sell for higher prices than those from other regions. Ravi Rajalingam, founder of the data labeling company Objectways, explains that because robot companies assume U.S. consumers will be the first to purchase humanoid robots, operational environment data from American homes is more valuable—resulting in worker hourly wages that can be up to three times higher than those in Vietnam or India. For the same task of folding clothes, hands in Los Angeles can earn three times more than hands in Chennai.

Image source: https://newatlas.com/robotics/figures-humanoid-robots-household-chores-2025-helix-ai-brett-adcock/

Arian Sadeghi, Vice President of Micro1, said that 1.6 million hours of monthly footage is far from sufficient—“We likely need billions of hours. We haven’t even begun collecting data on human-to-human interactions; we’re still at the most basic level of household tasks.”

Billions of hours— at the current collection rate, it would take approximately ten thousand years of continuous operation.

In 2019, anthropologist Mary Gray and computer scientist Siddharth Suri published a book called Ghost Work, which translates literally to "Ghost Work."

They want to describe the human labor—such as labeling images, filtering inappropriate content, and cleaning training data—that makes AI systems seem "intelligent," yet never appears in any product descriptions.

Disappearing Work: The True Future of Digital Labor Author: [US] Mary L. Gray, [US] Siddharth Suri Translator: Zuo Anpu

Gray said that when she first began researching this issue, asking engineers, "Who is doing this work?" yielded responses like, "I'm not sure" and "I don't dare to check."

In the past, ghost work primarily occurred in front of screens, involving actions like clicking, labeling, and reviewing. Now, the body itself—gestures such as folding clothes, the rhythm of cooking, or the motion of opening a refrigerator—is beginning to become raw material that can be collected, priced, and resold.

These raw materials flow from ordinary households in India, Nigeria, the Philippines, and Kenya, converge at companies in Palo Alto and San Francisco, and are transformed into products that enter the market.

Nick Couldry and Ulises Mejias proposed a framework called "data colonialism" in their study of the digital economy, meaning that tech companies' appropriation of data structurally continues the historical logic of colonialism’s extraction of land and resources, transforming human daily life itself into a raw material available for capital extraction.

In the case of Micro1, workers earn $15 per hour—a competitive wage in Nairobi or Manila—but pales in comparison to the billions of dollars invested in robotic companies.

More notably, there is a significant information asymmetry. Micro1, citing confidentiality, does not disclose the client list to workers, and the workers are unclear about how their data will be stored or whether it will be resold to other third parties. Workers sign agreements and receive payment, but they remain at the bottom of the information chain, with little knowledge of the full scope of what they are participating in.

While studying ghost labor, Gray discovered something that deeply impressed her: workers often spontaneously find one another and form informal mutual support networks, because the work itself provides almost no support—people must rely on each other to sustain a sense of meaning. Isolation is the default state of this kind of labor.

In 2026, the global humanoid robot market is projected to reach $4.23 billion, and by 2027, mass production plans by companies such as Tesla will drive the global cumulative installations past 100,000 units.

These robots are likely to enter factories and homes to take over physical labor, and the data used to train them comes from people who currently rely on physical labor to make a living.

Image source: https://developer.nvidia.com/blog/teaching-robots-to-tackle-household-chores/

Philosopher Michael Polanyi wrote a book in 1958 called Personal Knowledge, in which he stated: "We know more than we can tell." He referred to this as "tacit knowledge," meaning that humans possess vast amounts of knowledge that do not exist in propositional form, but are instead embodied in actions, perceptions, and intuition.

Cycling is a common example: you know how to maintain balance, but you cannot write down a set of rules to teach it to someone else. It can only be learned through practice, gradually internalized through observation, imitation, and repetition, not directly transferred.

When Polanyi wrote this book, AI did not exist. But his arguments have acquired a new weight of reality today.

What we are doing is attempting to extract this tacit knowledge from the human body and convert it into data that machines can process.

The camera on the worker’s forehead at Micro1 captures not just the motion of folding clothes, but also how the fingers sense the weight of the fabric, how the wrist flips at just the right moment, and how the gaze tracks the edge of the fabric throughout the folding process.

Scale AI has announced that it has collected over 100,000 hours of data. https://scale.com/blog/physical-ai

This is the first time in human history that an attempt has been made to externalize bodily knowledge on a large scale.

Polanyi said that tacit knowledge cannot be fully articulated, but that does not mean it cannot be appropriated. Couldry and Mejias argue that data colonialism transforms everyday life itself into a resource—a thing “just there, ready to be extracted.” Even making the bed at home is now included.

People often describe the impact of AI as “machines replacing knowledge workers,” but now even the most ordinary actions, not even considered skills, are being collected. If even these can become training data, then the question of “what constitutes human labor” is no longer a philosophical speculation—it has become a very practical political issue.

Zeus is a medical student from a city in the central highlands of Nigeria. After work each day, he mounts his phone on his forehead and begins making his bed.

He said he felt this was an opportunity to leave a mark. He didn’t feel like he was merely being used—he felt he was part of something meaningful.

This may be true, but it doesn’t prevent another possibility: that the imprint he left will ultimately take the shape of the motion轨迹 of his bed-making gesture, purchased by a company whose name he can’t recall, to train a machine he may not be able to afford in the future.

Polanyi said that all knowledge is personal, generated by specific individuals in specific contexts through specific practices. When this knowledge is detached from the person and made to continue operating after the person is gone, what, then, does the person, as the bearer of knowledge, truly possess?

This question currently has no answer. But it is being quietly asked in apartments in Nigeria, kitchens in India, and courtyards in the Philippines—at $15 per hour.

Reference materials:

https://www.technologyreview.com/2026/04/01/1134863/humanoid-data-training-gig-economy-2026-breakthrough-technology/

This article is from the WeChat official account "APPSO," authored by APPSO, discovering tomorrow's products.