Amazon Just Killed Mechanical Turk. The Real Robots Did It.

In 1770, a Hungarian inventor named Wolfgang von Kempelen unveiled a chess-playing robot to the Habsburg court. The Mechanical Turk — a brass-headed, turban-wearing automaton — toured Europe for decades, reportedly beating Napoleon and Benjamin Franklin. It was also completely fake. A human chess master was folded inside the cabinet, working the levers.

In 2005, Jeff Bezos resurrected the scam as a business model. Last Tuesday, Amazon killed it. And this time, the robots did it for real.

The thing Bezos built — Amazon Mechanical Turk, shuttering September 30 — was an API for human brains. You uploaded a task too hard for software but too boring for a salary, and a queue of people around the world did it for a few cents each. Label these images. Transcribe this receipt. Is this sentence positive or negative? Amazon called them HITs — Human Intelligence Tasks — and Bezos, with a straight face, described the service as "artificial artificial intelligence." The joke was honest in both directions. The original Turk was a human pretending to be a machine. Amazon's version was a machine that was secretly a crowd of humans. Two hundred thirty-five years apart, same product.

For twenty-one years, it quietly powered the internet you use. And its death notice — posted without ceremony to the MTurk website last week — reads like a eulogy for an entire era of computing.

The invisible workforce behind everything

Here's what most people never knew: MTurk was load-bearing infrastructure for the modern world.

Amazon built it originally to solve its own problem — cleaning up the endless messy product data on its store — but it escaped into the wild fast. Academic psychologists ran experiments on it. Insurance companies used it. The CAPTCHA-adjacent odd jobs of the early web flowed through it. At its peak, more than 500,000 workers — "turkers" — were on the platform.

The detail that should stop you cold: ImageNet was verified by turkers. The dataset that launched the deep learning revolution — the one AlexNet devoured in 2012, the one that made Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton famous, the one that put us on the path to GPT everything — was human-checked label by label by people earning pennies per task. The neural network era was bootstrapped by a crowdsourcing site named after a fraud.

It even had a walk-on role in the Facebook–Cambridge Analytica scandal: the personality quiz app that harvested the data got its seed population partly by paying survey-takers through Mechanical Turk. History's most infamous data grab started, in part, with HITs.

And the startup world loved the original scam's playbook: "AI" products that were secretly humans in a box, doing the work manually until investors funded the real thing. The tech industry even has a verb for it — "mechanical turking" your product. The name was a confession the whole time.

The snake eats itself

So how does a 500,000-worker platform with a twenty-year head start just die?

Slowly, then all at once. The July canary came first: Amazon quietly stopped accepting new customers on July 30, which workers correctly read as a death sentence. Then last week, the sentence: effective September 30, done.

But the real cause of death is a 2023 study that should be taught in every AI course. Researchers in Switzerland examined MTurk output and found that up to 46% of workers were using large language models to complete their tasks.

Sit with that. The platform existed because these tasks were easy for humans and impossible for machines. Then the machines became able to do them. The workers — rational people, paid cents per task — started feeding the tasks to ChatGPT. The "human intelligence" supply chain was quietly, invisibly, backfilled with machine intelligence. Data labeled by turkers was now labeled by a model pretending to be a turker. The 1770 cabinet, one more time, with a different ghost inside.

For AI labs buying training data, this was catastrophic in the most boring way possible. You cannot train a model on its own echo. For Amazon, the platform's entire premise had inverted: the HITs that were worth doing were being done by AI anyway, and the remaining human work demanded judgment that a few cents wouldn't buy. Reddit's epitaph was blunt — workers and researchers had abandoned it "years ago" over bots and fraud. Tuesday's announcement was just the paperwork.

What replaced it is the interesting part

Here's where the story flips from obituary to market analysis. The human data-labeling industry didn't die — it bifurcated.

At the top: companies like Scale AI, Mercor, and Prolific now sell expert human judgment to AI labs, and the money is serious. Mercor — founded in 2023, three years old — hit a $2 billion gross annualized run rate in June, double its pace from earlier in the year. Its workers aren't labeling cats and dogs. They're physicists and finance specialists producing training data and evaluating model outputs, paid by the hour at rates that would make a 2008 turker weep. (The gross number comes with an asterisk — Mercor pays 60–70% of it straight through to contractors, so net revenue is roughly $600–800M — but the growth curve is the point.)

At the bottom: the microtasks themselves. CAPTCHA-solving, receipt transcription, sentiment tagging, product matching. Done by models now. Not "$0.02 per task" cheap. Fractions of a cent cheap, instant, no queue, no breaks.

What got deleted is the middle. Commodity human microtask work — not expert enough to command Mercor rates, not mechanical enough to survive automation — that entire category of labor just... stopped existing. If you want a pattern for what AI does to labor markets, MTurk is the cleanest natural experiment we have. The work didn't disappear. It fell to the two extremes: brilliant people training the machines at one end, and the machines themselves at the other. The middle class of crowdwork went first, exactly like the middle rank of software work is going now.

The part that isn't funny

I've been glib so far because the arc practically begs for it — fake robot killed by real robots is a punchline machine. But the CNBC report has a detail worth pausing on.

Krista Pawloski started turking in 2008, on maternity leave, for extra income. In 2012 she lost her job and made MTurk her full-time work — flexible enough to care for her son, who has special needs. She later became an organizer with Turkopticon, the worker advocacy group, because someone had to: platform wages were famously brutal, with effective pay in some studies working out to a few dollars an hour, and Amazon's idea of support was famously absent.

MTurk was, by any fair measure, an exploitative gig — and it was also, for hundreds of thousands of people, income that bent around lives that normal jobs wouldn't bend around. Both things were true simultaneously, which is the most honest thing you can say about the crowdsourcing era. The shutdown leaves insurance and travel companies scrambling for alternatives, and leaves workers who never migrated to the newer platforms with five weeks of runway. "We regularly evaluate our programs," Amazon's notice says, in the corporate equivalent of shrugging at a funeral.

So where does the work go now?

If you were an MTurk customer — and plenty of academic labs and mid-size companies were, some until this month — you have five weeks. The migration path depends on which MTurk you were using:

If you needed research participants: Prolific is the standard answer now. Better vetting, fairer pay, and none of the "is my survey being answered by GPT" ambiguity that plagued MTurk's final years. Panels cost more and are worth it.

If you needed training data: The expert marketplaces (Mercor, Scale, Surge) are overkill for small jobs and the only serious option for frontier ones. The pricing is hourly and professional, because the work now requires people whose judgment a model can't fake.

If you needed microtasks done: Don't replace the platform. Replace the worker with software. Receipt transcription, listing cleanup, deduplication, tagging — this is model territory now, and it's not close. What used to be a $40 HIT batch is a script and an API call, or better, a single instruction to an agent that handles the whole pipeline without you writing the glue code.

That third category is where most MTurk volume actually lived, and it's the one nobody's mourning, because the replacement is strictly better: faster, cheaper, consistent, and available at 3 a.m. without a human anywhere in the loop. The tragedy of MTurk's closure is concentrated entirely in the second-order effects — the workers — and none of them were ever in the microtasks themselves.

The elegant ending

Still — you have to admire the symmetry.

A machine pretending to think, powered by a hidden human. A service named after it, selling human thinking through a machine interface. Killed, finally, by machines that actually think, trained on data that hidden humans verified — some of them, by the end, secretly using the machines to pretend they were thinking.

Bezos called it "artificial artificial intelligence." Two artificials. In 2005 the second one was the punchline. In 2026 the first one is, and the joke cost a 21-year-old platform its life. The Turk didn't get debunked this time. It got out-evolved.

The microtask API didn't die, though. It moved into your machine. The exact jobs that flowed through MTurk's queue — transcribe this, extract that, check this listing, summarize this page, one thousand times over — are now the daily workload of AI agents running on open-source models. The difference is the queue is gone: no HITs, no waiting for a worker in another timezone, no paying humans cents to be a slow function call. An agent with a browser and a terminal just does the task, end to end, while you get coffee.

That's roughly the thesis behind CopperRiver, the desktop assistant we built for exactly this — an agent on your Mac that browses, reads files, runs commands, and finishes the kind of chore-work that used to be somebody's HIT. From $9/month, which, fittingly, is about 450 Mechanical Turk tasks. The humans are off training frontier models now. The cabinet's yours.