In 1770, a wooden figure played chess at the Viennese court and won.
The apparatus looked serious. A life-sized carved figure dressed in Ottoman robes sat behind an ornate cabinet with a chessboard on top. Its inventor, Wolfgang von Kempelen, opened the cabinet doors in front of Empress Maria Theresa and her court, revealing what appeared to be an intricate system of gears and levers. He closed the doors, invited a challenger forward, and the wooden figure lifted its arm, picked up a piece, and made its opening move.
It won. Not just that game. Over the following decades this contraption, known as the Mechanical Turk, toured Europe defeating Benjamin Franklin, playing well enough that French chess master François-André Danican Philidor called his match against it exhausting. Europe was rattled. Had a machine actually learned to think?
Of course it hadn’t. A skilled chess player was hidden inside the cabinet, operating the figure through a system of levers and magnets. What audiences took for artificial intelligence was a stage trick. The illusion was that a machine was thinking. The reality was a person doing the work.
The hoax lasted nearly eighty years before it was fully exposed. Two and a half centuries later, Amazon named an internet service after that same automaton, paying tribute to precisely this trick of hiding humans behind an apparent machine.
This time, though, the story ends differently.
Bezos and “artificial artificial intelligence”
In 2005, Amazon launched Amazon Mechanical Turk, or MTurk. The operation was almost crude in its simplicity. Companies would post small tasks: classify these images, transcribe this audio clip, judge whether these two product photos show the same item. Then real people on the platform would do them, usually for a few cents per task.
Amazon originally built the system to solve an internal problem. Millions of product pages on its e-commerce platform needed to be deduplicated, categorized, and proofread. Computers at the time were bad at this kind of thing. Human eyes were reliable. So Amazon came up with a plan: chop the massive workload into tiny tasks, distribute them across the internet to anyone willing to click for pennies, and pay by the piece.
Jeff Bezos gave the service a memorable nickname: artificial artificial intelligence.
That phrase captured MTurk’s design philosophy. On paper, it was an API. A company fed tasks in one end and pulled results out the other, and in between something that looked like a high-throughput processing engine did the work. The engine wasn’t algorithms. It was thousands of people sitting at their computers. From the perspective of the company calling the API, it didn’t matter who was doing the labor. Only that the answers came back correct.
The logic is identical to the 1770 cabinet. On the outside, a machine. On the inside, a human.
MTurk grew fast. At its peak, more than 500,000 workers, called Turkers, were active on the platform. They came from everywhere. Some were American stay-at-home parents earning pocket money during their kids’ naps. Others were in developing countries where a few cents per task added up to a reasonable side income. A woman named Krista Pawloski started taking MTurk tasks in 2008 and by 2012 had turned it into full-time work.
The platform’s real value to the tech industry turned out to be enormous, but in a way nobody planned.
The people who trained the machines
For the first few years, MTurk mostly served Amazon’s original use cases: data cleaning, content moderation, product classification. Then machine learning and deep learning exploded in the 2010s, and MTurk found itself cast in a much bigger role. It became the invisible labor force training AI.
The reasoning was obvious. If you want an image recognition model to learn the difference between a cat and a dog, you first need millions of images already labeled “cat” or “dog.” If you want a language model to understand sentiment, someone has to read hundreds of thousands of reviews and mark each one “positive,” “negative,” or “neutral.” That work is called data annotation, and it’s the ground floor of AI training.
MTurk was built for exactly this shape of labor. Tasks are simple, easy to break into pieces, priced per unit, and returned fast. Researchers piled in. Countless academic datasets, spanning sentiment analysis, image classification, question answering, and semantic understanding, were built through MTurk annotation. It’s not a stretch to say that the voice assistants in our phones, the smart recommendations in our search engines, and the content filters on our social media all carry traces of MTurk workers’ labor in their training data.
Which produced a deeply ironic loop. Humans, posing as machines, worked to train the actual machines.
MTurk’s name originally paid tribute to the trick of hiding a human inside a machine. In the AI era, those same hidden humans fed the machines data day after day, teaching them, correcting them, helping them learn. Until one day, the machines had learned enough.
Then the machines didn’t need them anymore.
The wind-down
On August 25, 2026, Amazon posted a short notice on the MTurk site. After internal review, it wrote, AWS Mechanical Turk would shut down on September 30, 2026.
The wording was corporate and restrained. No sentimentality, no retrospective, no thank-you to the workforce. Just the standard formulation: we regularly evaluate our programs, tools, and services and make adjustments based on those evaluations.
Nobody in the industry was surprised. MTurk had been in visible decline for years. According to Krista Pawloski, now with the worker advocacy group Turkopticon, Amazon had clearly pulled back its investment. No new features, no interface improvements, task volume dropping, active workers shrinking. Back in July 2026, Amazon had already stopped accepting new customer signups. Anyone paying attention could see where it was headed.
Two clear trend lines drove the closure.
The first is AI capability. The most typical human intelligence tasks on MTurk, things like image classification, text transcription, and basic content moderation, were tasks that only humans could reliably do a decade ago. Today’s AI models do them faster, cheaper, and often better. When a machine can tell a cat from a dog on its own, you don’t need to pay tens of thousands of people to attach labels.
The second is the rise of specialized annotation. Even where AI still needs human help (and it does), the help it needs no longer looks like “anyone will do, pay them a few cents.” Training frontier AI models requires deep expertise. A new generation of annotation companies, Scale AI, Mercor, Prolific, has grown up around this need. They recruit annotators with specific domain knowledge. Lawyers labeling legal documents. Doctors labeling medical images. Senior engineers reviewing code for security issues. This flavor of training, known as RLHF (reinforcement learning from human feedback), demands more than following simple instructions. It demands professional judgment.
MTurk sold the cheapest, least specialized human labor. “Anyone will do” was once its core value proposition. But once AI absorbed the “anyone will do” work, and the remaining human work required “only you can do this” expertise, MTurk got squeezed from both ends. The low end went to machines. The high end went to specialists.
An awkward finding accelerated the collapse. A 2023 academic study found that up to 46 percent of workers on MTurk were using AI models to complete their tasks. The data meant to train AI was being generated by AI. Annotators used ChatGPT to do labeling, which then fed the training data for the next ChatGPT. The recursion was absurd, and it hollowed out MTurk’s central promise. The whole point of the platform was authentic human judgment, and internally that promise had collapsed.
When “anyone will do” becomes “only you”
On the surface, MTurk closing is just one tech company retiring a product that stopped paying its rent. Zoom out, though, and it’s telling a deeper story about where the value of human labor is moving.
Back to the basic question. In an age of increasingly capable AI, what work is left for people?
MTurk’s answer, in its heyday, was “do the simple things machines can’t.” In 2005, getting a computer to identify whether an image contained a cat was a hard technical problem. Any human with functional eyes could do it in a second. That created the MTurk model. Outsource whatever machines are bad at, pay by the piece, run at scale.
That answer worked for nearly two decades. It’s stopped working. Not because the underlying idea was wrong, but because the pool of “simple things machines can’t do” has drained. Image recognition, speech transcription, basic translation, content categorization. All the tasks that once required human “intelligence” are now handled by AI models. Faster, cheaper, and without breaks.
But this doesn’t mean humans have become irrelevant to the AI value chain. The opposite. The human role is undergoing a sharp upgrade.
The MTurk-era equation was: volume × simplicity = value. You didn’t need expertise. As long as you were a person with functioning eyes, ears, and reading comprehension, you had economic value. Five hundred thousand people, each doing hundreds of penny-priced tasks per day, together formed a massive data production engine.
The post-MTurk equation is becoming: expertise × judgment = value. AI no longer needs oceans of simple annotation. It needs smaller volumes of high-quality expert feedback. A doctor spending half an hour reviewing an AI-generated diagnostic recommendation. A lawyer taking an hour to evaluate a contract draft an AI produced. A senior engineer spending two hours doing a security review of AI-written code. These are tasks that AI can’t do on its own (at least not yet), and cheap crowdsourcing can’t do either. Only someone who actually knows the field can do them.
This is what economists call the “raising of the floor.” When automation eliminates the lowest-paid, most commoditized work in a market, it doesn’t eliminate the value of human judgment. It pushes that value upward, concentrating it in higher-skill tiers. The humans still involved in AI training are fewer in number, but each one carries more weight, higher barriers to entry, and better pay.
From being scheduled to scheduling
MTurk’s working model had an uncomfortable feature. Workers were entirely passive. You logged in, saw what tasks were available, picked one, finished it, got your few cents, and refreshed to find the next one. You didn’t know where your annotations would end up. You didn’t know if your judgment was good. You didn’t know who you were training or for what. You were a cog. More precisely, you were the hidden chess player inside the Mechanical Turk cabinet, except this time you couldn’t even see your opponent.
That model made sense for a long time because it solved a real efficiency problem. But it also carried an implicit devaluation of human labor. Your value wasn’t in who you were, what you knew, or what unique judgment you could offer. It was simply in being a person, a carbon-based creature that happened to be better than the computers of the time at certain perceptual tasks.
MTurk closing marks, in a sense, the end of that model.
The replacement isn’t about whether humans are still in the AI production pipeline (they still are). It’s about a fundamental change in their position and role. From “being scheduled by machines to do simple tasks” to “actively applying professional knowledge to evaluate, correct, and guide machines.”
In an RLHF workflow, human experts aren’t feeding the AI data. They’re vetting the AI. They look at multiple responses generated by the model, judge which is better, explain why, identify what’s off. That judgment isn’t purchasable for a few cents. It requires professional training, domain knowledge, and critical thinking.
Which points to something worth every knowledge worker considering. In the AI age, your work value depends less and less on what you can do, and more and more on what you can judge.
What you can do (typing, categorizing, translating, writing basic code) is being rapidly absorbed by AI. What you can judge (whether this proposal holds up, whether this code has security implications, whether this diagnosis makes sense, whether this creative direction works) is becoming scarcer and more valuable.
The chess player finally steps out of the cabinet
Back one more time to that 18th-century wooden figure.
The reason the hoax lasted decades is that audiences believed machines could think. Nobody suspected a human was hidden inside. In their worldview, human worth couldn’t possibly be so low that a person would spend hours crouched in a cabinet moving pieces for a fake machine. It was too absurd.
But for the past two decades, that’s precisely what millions of MTurk workers did every day. Hidden behind the machine, playing the role of API output, contributing human intelligence at pennies per unit. Their work was essential to AI’s development. They themselves were nearly invisible. No names, no faces, no visible contribution. Amazon called it Human Intelligence Tasks. The suppliers of that intelligence were treated in the system like interchangeable parts.
Now the cabinet is closing.
Not because humans no longer matter. Because the way humans matter is changing at the root.
Before, your value came from being a person. A pair of eyes that could read images, ears that could parse speech, a brain that could grasp meaning. That was enough. Going forward, your value comes from what kind of person you are. What specialized knowledge you hold, what judgments you can render that others (including AI) cannot, what layer you operate at when working alongside AI rather than being replaced by it.
For anyone still worried about AI taking their job, the MTurk story offers a more precise coordinate than the generic anxiety. What gets replaced isn’t “human work.” What gets replaced is “work that doesn’t require you specifically to be you.”
If your work value rests entirely on being able to perform a task, and that task can be described, decomposed, and repeated at scale, then you are squarely in the replacement zone. That was the position of hundreds of thousands of MTurk workers.
But if your work value rests on judgment, creativity, professional knowledge, and human insight, on the irreplaceability of “only you can make this call,” then AI’s advance actually makes you more valuable. The volume of AI output requiring evaluation and guidance only grows, and the number of people capable of that evaluation doesn’t grow just because AI is better.
The 1770 cabinet was a story about disguise. A human hidden inside a machine, pretending to be the machine thinking. The 2005 version, MTurk, was the same story on the internet. Humans hidden behind an API, pretending to be a system processing. In 2026, the story that ran for two and a half centuries finally reached its turning point. Humans didn’t disappear from the machines. Humans finally got the chance to step out of the cabinet, on their own terms, with their own expertise and their own judgment, standing next to the machine instead of crouching inside it.
That, more than anything else, is what Mechanical Turk closing actually means.
Related reading
- Clockwise Is Shutting Down: Should You Migrate to Motion or Reclaim in 2026?
- In 2026 I Moved My Projects Off Vercel: The Alternatives That Actually Work
- Power Automate AI vs AWS Bedrock Agents vs Google Vertex AI Agent Builder: Which Enterprise AI Automation Platform in 2026?
- AI Video Generators After Sora: Runway vs Pika vs Kling vs Vidu (2026)
Stay updated with our latest AI insights
