A dinner in San Francisco, paid for by the startup Shift, was no different from dozens of others—except for the camera on the chef’s cap. Alisher Demir cooked for free: in return, the startup took a first-person recording of a person chopping vegetables, flipping an omelet, wiping down a table. “Now I feed AI too,” he told a Business Insider reporter, without a trace of posturing. The deal sounds like a joke, but it is the most ordinary economics of 2026: dinner in exchange for a panoramic point of view. Someone has finally found a way to rent your eyes—for a dinner they paid for themselves.
Germany’s Microagi, which owns Shift, called itself “a social network for food” until recently. Now it hands out cameras and is one of thousands of firms buying human movements wholesale. The industry did not grow out of a laboratory, but out of a kitchen—and to understand where it is headed, it is worth rewinding two and a half centuries, to a Hungarian palace where someone had already hidden a man inside a machine and sold his labor to the public as a technological marvel.
The scale of today’s hunt for movement: Shift pays more than 35,000 people to record physical tasks, operates in at least fifteen countries, pays up to $30 an hour, and disbursed around $10 million in 2026. Its competitor Human Archive, also based in San Francisco, collects footage from more than thirty countries for $1–10 an hour—the price depends on where a person opened a door. The platform Encord, which sells annotation to robotics labs, reports that demand for robot-training data has risen roughly tenfold in a year: clients who once ordered a thousand hours now ask for ten—or even 100,000.
The money swirling around matches the appetite. In the first half of 2026 alone, robotics drew $33.4 billion, against $28.7 billion for all of last year. The market for AI-training datasets is valued at $11.9 billion, with a forecast of nearly $50 billion by 2031. The people in the race admit in black and white that they are feeding a pupil that is still stupid: Encord compares the robot brain to a two-year-old child, while for the least structured environment of all—the home—some estimates call for video on the scale of five YouTubes.
While the robot is stupid, a human thinks for it, and that is not a metaphor. At American warehouses, contractors in Mexico and the Philippines remotely teleoperate packing robots; shifts run eight hours, hands cramp up, and in the company office, as one Texas operator admits, “they even have a massage gun” for cases like that. Tutor Intelligence founder Josh Gruenstein describes his operator-filled warehouse more honestly than any marketing paper could: “This is how the great transfer of knowledge from people to machines happens.”
The name of this transaction is not “knowledge transfer.” It was coined 220 years ago near the Danube, and to see just how precisely, we need to return to Schönbrunn in 1770.
In April 1770, the court official Wolfgang von Kempelen presented Empress Maria Theresa with an automaton: a figure in a Turkish robe and turban sat behind a cabinet with a chessboard, moved the pieces with its left hand, turned its head, said “eh!” when it captured a piece—and beat the Empress’s adviser in half an hour. Kempelen opened the cabinet doors, displaying gears and levers, wound a key, and the machine “came to life.” Then the Turk toured for half a century: Leipzig, Dresden, London, Paris, Boston. Among those it defeated were Benjamin Franklin, Frederick the Great, and Napoleon.
The secret was not technical. A living chess player sat inside the cabinet. Kempelen opened the doors one at a time, and while spectators peered into one compartment, the man on a sliding seat moved into another, while a fake mechanism slid into his place. The chess player saw the board through magnetic telemetry from below: magnets sat in the bases of the pieces, needles beneath the board were drawn to them, and the operator read the position without taking his eyes off an internal panel. His hand moved the wooden arm through a pantograph. A candle lit the master from within, while ventilation tubes carried smoke into the turban, where it blended with the smoke from the audience’s candles—so that no one would notice the machine breathing.
Organizationally, it was flawless. No one asked the Turk whether it was honest—they asked it to play. Kempelen even staged counter-demonstrations: he invited spectators to bring magnets and iron rods, so they could see that the machine did not operate “by attraction.” Exposés appeared during his lifetime—the Dutch scientist Willem ’s Gravesande was looking for a man beneath the stage just two years after the debut—and found none; Kempelen answered his suspicions with a curt “nothing, nothing.” The magician Decremps described the solution in 1789 as “a dwarf chess player hidden in a chest of drawers,” while the Dresden baron Racknitz published his own diagram of the device, complete with illustrations, that same year. But the rational Age of Enlightenment, which had routed alchemists and astrologers, capitulated before the Turk for half a century: the label “technological marvel” proved cheaper than any investigation.
Payment was the whole point. The Turk existed because people liked believing in a machine that thought. The fake worked as a good deal: the spectator got a show from the future, Kempelen got money and fame, and the real chess player got stifling heat, confinement, and candle wax on his shoulders.
Among the spectators was one man who did not buy it and did not forget it. In 1819, the young Charles Babbage played two games against the Turk and lost both. According to historian Tom Standage, he suspected that the automaton was not “intelligent”—that a person simply sat inside. Babbage did not expose it, but the encounter with the fake automaton stayed in his mind.
Ten years later, he abandoned work on the Difference Engine—a four-ton calculator made of 25,000 parts—in favor of the Analytical Engine: a store, a mill, punch cards, in other words, a prototype computer. Ada Lovelace understood in this machine what the father of computer science did not: programmability made it universal—one day, Lovelace wrote, it would “compose elaborate and scientific pieces of music.” And in 1864, Babbage wrote in his diary that, to test his “mechanical notation,” he had chosen “a machine capable of successfully playing an intellectual game, such… as chess.”
There is no technological link between the Turk and Babbage’s machines—the fake passed not a single gear on to him. It passed on something else: the very framing of the question “can a machine think” had become a public demand before the first real computer existed. Sir David Brewster, Babbage’s colleague, summed up the arc of several decades: “Automaton toys that once amused the crowd are now used to expand human capabilities and spread civilization.” The idea of a thinking machine entered culture through the door of a magic trick—and never left.
By the 1830s, there had been plenty of exposés of the Turk. But almost all of them were written about mechanics—magnets, partitions, false panels. The man who saw it in Richmond, Virginia, during the Turk’s American tour wrote something else. Edgar Allan Poe published “Maelzel’s Chess-Player” in April 1836—a text regarded as the best-known analysis of the Turk, though historians later disproved most of its technical hypotheses. Poe was wrong in the details and right in the scale of it: he was interested not in the device, but in what the person inside the cabinet felt, “tightly compressed” among levers and gears “in exceedingly painful and awkward positions.”
It is an exceedingly rare gesture. The public was entertained by the machine’s mind; Poe wrote about the body that serviced that mind. In an office today, it would be a direct quote from Poe—only now his cabinet is called a warehouse, and his tub of gears is called a teleoperation protocol.
When Amazon launched a service for distributing small pieces of work to living people in November 2005, it was not looking for a pretty legend. It was looking for a solution to a real problem: the site was drowning in tens of thousands of duplicate products, and its engineers considered the task of removing duplicates programmatically insoluble. Manager Venky Harinarayan patented a scheme for “hybrid machine/human computing”: a task is broken into pieces and scattered across a network of people, each tugging at their own piece from home. Bezos decided to sell it as a service. The legend behind the name was not a legend but a résumé: Amazon Mechanical Turk—“a service designed so that human labor is invisible,” as IEEE Spectrum put it in its series on AI history.
The invisibility was architectural: workers have no names, only handles and a task queue; for the customer, it is the interface of a machine that “simply” has access to intellectual labor. The economics proved brutal: according to a 2017 study, the median Turk worker earned around two dollars an hour, and only four percent made more than the US minimum wage ($7.25). Today’s prices for human movement—from $1 to $30 an hour—are not progress on that scale but rather the restoration of a market: a person still costs somewhere between two and thirty dollars, depending on the country where they opened their laptop.
The difference between 1770, 2005, and 2026 is that Kempelen’s person hid in the cabinet so the public would not see the deception; Amazon’s hid there so the customer would not see the employee; Shift’s hides there so the camera sees everything, while you do not see where the dinners you ate are going.
There is one counter-history in the story of miracle machines that makes the current arrangement easier to see. In 1948–1949, the Bristol neurophysiologist William Grey Walter built two machines the size of tortoises—Elmer and Elsie. Their hardware was unimpressive: a body made from military surplus, old alarm-clock relays for a “brain,” two neurons in the literal sense—two circuits, two motors—and one photoreceptor. Their public effect, by contrast, was off the charts: the tortoises crawled toward light, skirted obstacles, trembled before a mirror with a little lamp on their noses as if they were “nervous,” and reporters wrote about “thinking machines.” When a tortoise found in the mirror its only rival, equal in brightness, and began a melodic minuet with it—approach-retreat, approach-retreat—the audience watched it as a dance of deference.
Walter was completely honest with the public: he did not claim that Elmer and Elsie were intelligent—he claimed that behaviour was the minimum that already looked like life. His wheeled “conditioned reflex,” CORA—made, ironically, from Elsie’s parts—and the “tameable” machina docilis, which came at a whistle, worked—modestly and in plain view. He even anticipated debates over self-awareness in the mirror, carefully observing that if an animal behaved that way, it “might be taken as evidence of some degree of self-awareness.”
The tortoises were poorly received. Not because they did not work, but because modesty does not sell. His colleagues in the Ratio Club winced at the newspaper hype; the academic verdict for thirty years was: “amusing trifles.” Meanwhile, the fake Turk and its heirs—robot dogs, singing heads, “thinking” automata—remained the currency of exhibitions. The market lesson Walter learned before the perceptron advocates and the startups of the 2020s: an honest small model almost always loses to a beautiful big promise—a bigger illusion sells better, because the buyer is buying a show, not a mechanism.
Roboticists have been using the same crypt as Kempelen for decades: a person inside the machine. The difference is that Walter’s person was outside—honestly, in the role of spectator and operator. Today’s industry has put the person back inside, because that is the only configuration in which data about human movement can be extracted at all.
First: consent. Mecka AI and Human Archive did not tell Business Insider how they secure the consent of people filmed in partner cafés, hotels, and plumbing businesses. Researcher Mark Graham of Oxford, co-author of Feeding the Machine, describes the typical participant in this market as “a disposable appendage of a large machine,” who often does not even understand what their movement is becoming. This is not moralizing but business risk: one wave of “we didn’t know where our video would end up,” and the platform loses both workers and its public shield.
Second: the price of the body. Teleoperation is not “the work of the future,” but desk-bound recruitment of hands: eight-hour shifts, cramps, massage guns. The price a person pays for the “mind of the machine” is still physical wear and tear; only now, instead of being hidden behind a cabinet, it appears in HR reports rather than spectators’ diaries. Ïndustry promises that teleoperation is a temporary “knowledge transfer”; no one gives a timeline for this “temporary.”
Third: the extraction model itself. Startup Eidon AI shut down in 2026: founder Sam Padilla called the business “data dropshipping”—a race to the bottom with thin margins and difficult logistics. At the same time, World Labs—the subject of an $8.2 billion AMD deal—and others are building generative simulations from which movement can be synthesized from a handful of recordings. If synthetics work, all 35,000 people with cameras on their caps will become contractors in a market that may already be replacing itself. Ken Goldberg of Berkeley summed up the industry’s central question without consolation: “It could take five years or fifty… even with unlimited data, I’m not sure this approach will work.”
Kempelen’s Turk was not a machine pretending to be a person. It was a person pretending to be a machine—and in that lies the central formula of the entire story that followed: techno-futurism almost always turns out to be a contract for invisible labor. The fake gave Babbage a question; Poe—the measure of the person inside; Amazon—a trademark; Shift and Microagi—a business model.
Robots that can chop vegetables and cannot be trained any faster than a two-year-old child will not be born in simulations—they will be taught by a person fitted with a camera and promised $13 an hour. That is engineering-honest. What is dishonest is when all the rest of the talk about “machines that can do everything on their own” is arranged just like the Turk: the board on top is visible to everyone, the cabinet is locked, and no one asks about the person inside until the moment he calls for help.
Poe wrote about it in 1836, Walter in 1953, Gruenstein in October 2026. Each time, the wording is the same: the one making decisions for the “mind of the machine” is the one no one can see.
Sources: Business Insider (Rya Jetha, October 2026, “AI Data Grab” series)—Shift (35,000+ people, >15 countries, up to $30/hr, ~$10 million in payouts), Human Archive ($1–10/hr, >30 countries), Encord (demand ×10, dataset 5× YouTube), teleoperation (Mexico, Philippines, Texas, Tutor Intelligence), Microagi/chef Demir, Mecka AI and sensor helmets, $33.4 billion in robotics investment (PitchBook, H1 2026), Mordor Intelligence ($11.9→49.8 billion by 2031), World Labs/AMD ($8.2 billion), the shutdown of Eidon AI, quotations from Josh Gruenstein, Mark Graham, Ken Goldberg, Virgil Velmurugan (Encord), Bercan Kilic (Microagi); Wikipedia—Mechanical Turk (cabinet construction, magnetic board, pantograph, sliding seat, turban ventilation, magnet demonstrations, ’sGravesande, Decremps, Racknitz, tours, 1854 fire) and Elmer and Elsie (alarm clocks, phototaxis, mirror experiment); IEEE Spectrum, “Untold History of AI” series (Oscar Schwartz, 2019)—Amazon before MTurk (duplicates, “insurmountable,” Harinarayan’s “hybrid machine/human computing arrangement” patent, 2005 launch), median pay of $2/hr and 4% above $7.25 (2017 data), “designed to make labor invisible,” Poe’s “tightly compressed... painful and awkward positions”; IEEE Spectrum, “Charles Babbage and the Turk” (18.03.2019)—the 1819 games, 1864 diary, David Brewster quotation, Lovelace analysis; Rutherford Journal (Margaret Boden, “Grey Walter’s Anticipatory Tortoises”)—Ratio Club, CORA made from Elsie’s parts, machina docilis and machina reproducatrix, the mechanical minuet of tortoises with pilot lights, newspaper hype and its cost, “two neurons,” colleagues’ reactions; for the Babbage/Lovelace biography—Standage, The Turk (2002).