When Zeus, a medical student in central Nigeria, comes home from the hospital, he straps his iPhone to his forehead and records himself making his bed. He moves slowly, keeping his hands in frame. The footage goes to Micro1, a Palo Alto company that sells real-world human movement data to robotics firms. He earns $15 an hour.
In Chicago, an 18-year-old welding apprentice sold his phone call recordings to a platform called Neon Mobile for $0.50 per minute. In Los Angeles, a man filmed himself washing dishes and wiping a stove for two hours and earned $80, as reported by MIT Technology Review.
None of them work for a tech company. All of them are training AI. And what they are part of is no longer a side hustle. It is the visible edge of a new labor market forming around one of the AI industry’s most pressing constraints: the need for real human data, at scale, in the physical world.
Data Famine Driving the Market
The demand comes from a structural problem the industry has been building toward for years. Epoch AI projects that language models will exhaust the stock of high-quality human-generated text available for training before 2026 at current consumption rates.
For physical AI, the gap is sharper still. Virtual simulations can teach robots to perform acrobatics, but not how to grasp and move objects, because simulations struggle to model physics accurately, as reported by MIT Technology Review. Real-world human movement data, captured one interaction at a time, has no shortcut replacement.
Investors poured more than $6 billion into humanoid robots in 2025, according to MIT Technology Review. Robotics companies are now spending more than $100 million per year on real-world data from collection firms, according to Micro1 CEO Ali Ansari, as reported by MIT Technology Review.
Scale AI has collected more than 100,000 hours of robotic demonstration data from its San Francisco lab and global contributors, as the company stated on its own platform.
Labor Market With Two Very Different Tiers
The market forming around this demand is splitting into two distinct levels. At the entry level, gig platforms are redeploying their existing workforces. DoorDash launched a standalone app in March that pays its 8 million U.S. couriers to film household chores, record multilingual conversations, and scan commercial locations for AI and robotics training, as reported by PYMNTS.
Uber introduced a comparable program the previous October through its AI Solutions division, now operating in 30 countries, as reported by Entrepreneur. Instacart has made similar moves, as reported by Bloomberg.
At the top of the market, the rates look entirely different. OpenAI hired more than 100 former investment bankers from Goldman Sachs, JPMorgan and Morgan Stanley under a project code-named Mercury, paying them $150 an hour to build financial models for IPOs, restructurings and leveraged buyouts, according to Bloomberg.
A separate initiative run through data-labeling startup Handshake AI, known internally as Project Stagecraft, has deployed between 3,000 and 4,000 contractors to build professional task simulations across more than 400 job titles, paying at least $50 an hour for general contributors and up to $500 hourly for domain experts, which reviewed internal documents. Contributors developing a professional persona as a financial manager, a soil scientist, or a commercial pilot write prompts that mirror the actual decision-making those jobs demand.
The spread between a dish-washing clip and a Goldman Sachs financial model is enormous, but both now live inside the same emerging labor category.
Terms of the Exchange
The market is formalizing, but the terms remain uneven. Contributors who sign data collection contracts often grant irrevocable licenses, meaning their voice or movement data can be reused in AI systems indefinitely, as reported by MIT Technology Review.
Workers interviewed by the publication said none of them knew how their data would be stored, shared or passed to third parties, including the robotics companies ultimately buying it. Some asked on internal channels whether their data could be deleted. Micro1 declined to comment on whether such requests are honored, according to MIT Technology Review.
For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.
