Amazon hid a human inside a machine for 21 years. It's closing the machine because the humans found one to hide inside

The trick was always the cabinet. When Wolfgang von Kempelen wheeled his chess-playing automaton into the Viennese court in 1770 — a life-sized figure in a turban, seated at a wooden box crowded with brass gears — he would open the doors one at a time before each game, letting the audience see the machinery, the candle, the clockwork, the impossibility of a person fitting inside. Then he would close the doors, and the figure would play, and beat almost everyone it faced. The doors were the whole act. They were how you proved there was no one in there. There was, of course, always someone in there: a chess master folded into the false compartment on a sliding seat, reading the board through magnets, moving the Turk's arm with a pantograph. The genius of the thing was never the automaton. It was the choreography that made you forget to look for the human.

Two hundred and thirty-five years later, in 2005, Amazon built a website and gave it the automaton's name. Jeff Bezos supplied the tagline himself, and it was almost too honest: he called Mechanical Turk 'artificial artificial intelligence.' Not artificial intelligence — the artificial version of that, the kind you get by hiding people inside the machine and closing the doors. On September 30, after twenty-one years, Amazon is switching it off.

The notice

The end arrived the way these things do now, as an administrative update. Amazon said it had reviewed its programs and would retire Mechanical Turk on September 30, 2026, along with two adjacent services, SageMaker Ground Truth and Augmented AI, that did neighboring work. New customers had already been turned away on July 30. The people who noticed first were not the customers. They were the workers, reading the tea leaves on a Reddit forum where they have compared notes for two decades, and they understood immediately what the closed signup meant. They had spent their working lives inferring the state of a machine from the outside. They were right.

It is worth being precise about what is closing, because the phrase 'crowdsourcing platform' does a lot of laundering. Mechanical Turk was a marketplace for what Amazon called Human Intelligence Tasks — HITs, in the interface, a word that told you exactly how the work was measured. A HIT was a fragment: label whether this photo contains a car, transcribe thirty seconds of audio, decide if two product listings are the same shoe, read this paragraph and rate its tone. Amazon built the thing in the first place to clean up duplicate listings on its own store, then realized other companies had the same kind of work — the small, judgment-shaped tasks that software still could not do — and opened the side door to the public. At its peak the platform had more than half a million registered workers. They were paid, mostly, in cents.

You could see the whole psychology of the place in the vocabulary. The workers were 'Turkers.' The companies posting work were 'requesters.' A requester could reject a completed task and simply not pay for it, and the worker had little recourse; a rejected HIT also dinged an approval rating that gated access to the better-paying work, so the incentive was to accept the rejection quietly and move on. The interface never showed the requester who the worker was. It showed a string of characters. The doors, in other words, stayed closed. That was not a flaw in the design. That was the design.

Amazon built a marketplace where the human was the hidden mechanism and the machine was the brand. For twenty-one years it worked exactly as named.

The people who were the intelligence

I want to resist the thing that usually happens to Turkers in a story like this, which is that they become a statistic about wages and then disappear. So hold the specifics that the research and the reporting have established over the years, because they are the character here. The work was done in kitchens and spare rooms in the American Midwest, in Hyderabad, in Manila, in Nairobi, on laptops kept open beside other lives. Some Turkers ran it as a side hustle for gift-card money. Some depended on it. Many built elaborate private tooling — scripts, browser extensions, shared spreadsheets rating which requesters paid fairly and which ghosted — an entire folk infrastructure of self-defense that the platform itself never provided. They organized their own dignity, because the marketplace was not designed to extend them any.

And their labor went somewhere specific. The modern AI industry has a founding dependency it does not enjoy discussing: the models learned to recognize a cat, or a stop sign, or a slur, because hundreds of thousands of people were paid pennies to look at cats and stop signs and slurs and say what they were. Mechanical Turk was one of the places that happened first and at scale. The reasoning systems now praised for their fluency are standing on a floor of human judgment laid down one HIT at a time, by people whose names never appeared, for money you would not stoop to pick up off the street. The automaton could only play chess because a master was folded inside it. The analogy held all the way down.

The reversal

Here is the part that turns the story, and it is the kind of contradiction that, if you sat with it long enough, would tell you the whole platform was already over before Amazon wrote the notice. In 2023, researchers in Switzerland ran an experiment: they hired Mechanical Turk workers to summarize medical abstracts, then checked how the summaries were produced. As many as 46 percent of the workers, by the researchers' estimate, appeared to have used a large language model to do the task.

Read that again slowly, because it is a hall of mirrors. The service existed to supply human answers where machines fell short. Its workers were the ground truth — the reference against which the machines were graded. And a good share of them, being rational people paid by the task, had quietly started routing the work through the very machines they were supposed to be the human alternative to. The people hidden inside the automaton had found a smaller automaton to hide inside. The doors closed on a cabinet that, near the end, had a chatbot folded into the false compartment.

It would be easy to read that as a punchline about the workers, and it is not one. They were doing what the structure rewarded: the task paid for an output, not for a method, and no one had ever paid a Turker for their humanity as such — only for the appearance of it, delivered on time. If the machine could produce the appearance faster, the machine got used. What collapsed was not the workers' integrity. It was the fiction that had always held the marketplace together: that the human inside and the machine outside were different things, cleanly separable, one grading the other. By 2023 you could no longer tell which was in the box.

Where the work went, and what it costs now

The work did not disappear. It moved, and in moving it revealed what Mechanical Turk had actually been — the cheap, anonymous, first draft of an industry that has since learned to dress better. A newer crop of companies now recruits the people who train models: Scale AI, Prolific, and Mercor among them. But the economics have inverted. Where Mechanical Turk paid pennies to an anonymous crowd for fragments of judgment, the frontier labs now pay $50 to $100 an hour, and more, for credentialed experts — doctors, lawyers, PhDs — to supply the kind of specialized reasoning a model cannot yet fake. Mercor, recently valued at around $10 billion, reportedly pays out roughly $1.5 million a day to a bench of some 30,000 contractors.

• The old model: a HIT paid in cents, done by an anonymous worker rated by a string of characters, rejected without recourse, hidden from the company that used the output.
• The new model: an hour of a named specialist's time, paid at professional rates, on a platform that markets its workforce's credentials as the product — the ghost given a LinkedIn and a rate card.
• What stayed the same: the model on the other end still cannot do the hardest part alone. It still needs a person folded into the process. The industry simply decided the person should be a radiologist instead of a stranger with a spare hour.

That is the quiet thing the closure marks. Not that human labor left AI — it did not, it is arguably more essential now than in 2005, and better paid at the top — but that the anonymous, penny-wage version of it stopped being useful. The tasks a stranger could do for cents are precisely the tasks the models absorbed first. What is left for people to do is the work that is still genuinely hard, and that work commands a professional's fee. Mechanical Turk is closing not because the machine finally replaced the human, but because the particular bargain it offered — human judgment, stripped of the human, sold by the fragment — stopped clearing at a price anyone would pay.

The doors

Kempelen's automaton toured Europe and America for decades, beat Napoleon and Benjamin Franklin, and kept its secret longer than most people believed possible, until a fire finally took it in a Philadelphia museum in 1854. The operators — the succession of chess masters who spent their careers folded into a box, winning games no one would ever credit them for — are mostly lost to record. We remember the machine. We forgot to write down the people inside it. That was the point of the machine.

On September 30 the interface will go read-only, then dark. Somewhere a last HIT will be submitted — a photo labeled, a sentence rated, a couple of cents booked against an approval score that will never be spent. It is tempting to call it the moment a machine outgrew the people who built it, but that is the trick again, the doors swinging open to show you an empty cabinet. Look once more before they close. The intelligence was never artificial. It was half a million people, and a turban, and a box, and a name chosen to make sure you never asked who was inside.