AI's Next Shortage Is Other People's Work

In Andhra Pradesh, workers strap cameras to their heads and film themselves cooking, folding clothes, packing boxes.

Little attention, if any, is paid to how AI can serve India's 490 million informal workers, the very people who form the backbone of our economy. — NITI Aayog

The footage is sold to Fortune 500 companies and to Amazon to teach robots how to move ^[1]. In Munich, Agile Robots has opened a Robot Academy that films people's hands and fingers so industrial machines can copy them ^[2]. OpenAI has bought tens of thousands of Mac minis — practice rooms where its agents learn to use software by trial and error on the real thing ^[3]. The same purchase is happening in all these places, and the thing being bought has quietly become the scarcest input in the AI buildout. Firms that once scraped the web now pay six- and seven-figure sums for anonymized internal records — emails, Slack messages, slide decks — and pay individuals to record themselves cooking and packing ^[4]. The internet is where the last generation of AI fed: a near-infinite text of books, posts and code. Work was never in it. The robots-and-agents generation needs something else — recordings of how humans actually do things. One robotics startup founded by former DeepMind engineers puts a number on the shortfall.

If they have to call the whiz team robotics guys from wherever they are to fly in and fix it, they just won't be able to do it. It will become an expensive paperweight. — Jonathan Scholz

The gap it names is the distance between the trillions of words a language model can read and the near-total absence of footage showing a person doing a task ^[5]. Even the largest trove of recorded reality doesn't close it. DeepMind feeds 280 billion Street View images across 110 countries into its Genie world model ^[6] — but a photograph of a street supplies the surfaces where work happens, not the work. You cannot teach a machine to cook by showing it a kitchen. Nor can the labs fake it. When DeepMind's researchers set out to stretch scarce training data, they did not generate new text. They built a method that purifies and recycles real text that had been discarded as useless, because data a model generates itself pushes it toward collapse ^[7]. The industry's own specialists treat synthetic data as a supplement at best and a source of brittle, biased models at worst ^[8]. The one input that cannot be fabricated is the one this product now runs on. And the free-scraping commons that fed the last decade is closing on every front at once. Five publishers and the novelist Scott Turow are suing Meta, alleging it pirated millions of books to train its Llama model and asking a court to order the data destroyed — after Meta walked away from a plan to pay up to $200 million for licenses and instead gambled on a fair-use defense ^[9]. The UK has scrapped a proposal that would have let AI scrape copyrighted work unless creators opted out, turning instead toward paid licensing ^[10]. Cloudflare is building the toll road: it bought a data marketplace to sell "Pay Per Crawl" access.

The Human Native team will help us accelerate the next phase, where we build the tools that allow content to be discovered, priced, and purchased through transparent channels. — Matthew Prince

In Qingdao, the state runs a training center where component assembly, kitchen work and supermarket restocking are repeated thousands of times to generate motion data for humanoid robots ^[11]. And the demand multiplies by region: one robotics chief warns that a robot trained on a single country's data becomes a regional tool rather than a general-purpose one, which means every country that wants useful machines will need its own supply of filmed, demonstrated work ^[12]. The same weeks brought the money question to a head. Firmus Technologies, an AI data-center builder, tried to raise $7 billion in what would have been Australia's second-largest float, and pulled it after analysts called its earnings forecast a "fairytale" ^[13]. CoreWeave's junk bonds are reported reaching 15 percent, the kind of yield that means bond buyers have stopped trusting the story ^[14]. These two turns did not cause each other. They are the same shift arriving from two directions. Lenders stopped waving AI borrowers through in exactly the weeks the industry's declared answer to the revenue question became the agent — software that does work — and by its builders' own account, it cannot exist without exactly the purchased experience above.

If human beings can learn from school, why can’t robots? — Zhaopeng Chen

Google is in talks to put more than $1.5 billion into Mechanize, a startup whose founder's stated aim is full automation of the economy; Meta is recording its own employees' keystrokes and mouse movements; and nearly half of Scale AI's new training projects now run in simulated or observed work environments rather than static datasets ^[15]. The pattern has already built a market. Surge AI booked $1.2 billion in revenue connecting specialists — lawyers, doctors, people fluent in internet culture — to labs at rates reaching $200 an hour ^[16]. Meta paid $14.3 billion for a 49 percent stake in Scale AI, a company that remains unprofitable ^[16]. That is real money. It is also a rounding error next to data-center capex. What has changed is not the size of the bill but the character of the input: work that has to be bought from the people who have it, in cash, one demonstration at a time. The people being bought from are, in most cases, the people the product is built to replace. The entry-level digital jobs that once absorbed them are going. The International Trade Centre estimates that basic digital work for refugees at Kenya's Kakuma camp has fallen by half since 2022 ^[17]. Scale AI's Remotasks operation left Kenya in 2024 ^[17].

AI has taken people by surprise with its speed and impact, and we need to ask what labor organizing looks like now. — Pamela Coke-Hamilton

In Andhra Pradesh, the worker with the head-mounted camera is still filming the folding of clothes — for a robot being built to fold clothes. The purchase contains its own endgame.