Web Services confirmed this week that Mechanical Turk — the crowdsourcing marketplace that built the modern AI training industry by hiding human labor behind the curtain of products marketed as intelligent — will stop accepting new customers on July 30, 2026. TechCrunch’s report on the closure AWS has moved the platform to its official “Services in Maintenance” list, the holding pen for products being eased toward a full shutdown. No new features will be added. The date 21 years of invisible human computation ends is now on the calendar.
The timing closes a loop that is almost too perfect to be accidental. A platform built on humans pretending to be machines has been made obsolete by machines pretending not to need humans — and accelerated to that end by the workers themselves, who quietly started using AI to do their jobs.
Named for a Hoax, Built on One
In 1769, Hungarian engineer Wolfgang von Kempelen unveiled a chess-playing machine that astonished European courts, defeating Napoleon Bonaparte and Benjamin Franklin before anyone realized the truth: a human chess master sat hidden inside the cabinet, controlling the pieces through a system of levers. Von Kempelen’s “Mechanical Turk” was one of the most spectacular frauds in the history of technology — a hoax so elegant that it ran for decades.
When Amazon launched its crowdsourcing marketplace in November 2005, naming it Amazon Mechanical Turk was a deliberate wink. The service allowed companies to outsource small digital tasks — completing CAPTCHA challenges, transcribing audio clips, flagging inappropriate content, verifying restaurant listings — to a global pool of workers, while those companies’ products appeared to run on machine intelligence alone. Amazon’s internal team coined the phrase “artificial artificial intelligence” to describe what the platform made possible: a computational service that routed tasks to humans for the portions that algorithms could not yet handle.
The parallel with von Kempelen was self-aware from the start. MTurk workers were not represented by names but by numbers. Communication between requesters and workers was fully depersonalized. The human labor that powered the service was, by design, invisible. Jaron Lanier on MTurk’s design observed that MTurk’s architecture “allows you to think of the people as software components” in a way that conjures “a sense of magic, as if you can just pluck results out of the cloud at an incredibly low cost.”
What Lanier named in 2005 became the defining business model of a decade. MTurk workers earned anywhere from a cent to a few dollars per task. Hara et al.’s landmark wage analysis of 3.8 million completed tasks found that workers earned a median of approximately $2 an hour, with only 4% earning more than the federal minimum wage of $7.25. Workers were classified as independent contractors, receiving no minimum wage protections, no overtime, and no benefits. Amazon collected a 20% commission on every successfully completed Human Intelligence Task.
At its peak, MTurk had more than 500,000 registered workers from over 190 countries. By 2018, research showed that while over 100,000 workers remained available on the platform at any given time, only around 2,000 were actively working. The rest had moved on. The platform was already hollowing out before the final turn.
Mechanical Turk and the Machine Learning Pipeline
MTurk’s first decade was defined by its role in the gig economy it helped invent, predating Fiverr and Upwork by years. Its second decade was defined by something more consequential: it became the primary industrial mechanism for generating the human-annotated training data that large AI models required.
The concept is now central to how AI is built. Supervised machine learning requires humans to label examples — this image contains a stop sign, this sentence expresses negative sentiment, this product review is fabricating a claim. MTurk supplied that labor at scale, cheaply. When Amazon integrated the platform with its SageMaker machine learning service in 2018, it was repositioning MTurk not as a gig marketplace but as an AI supply chain component: the source of the human judgment that taught machines to see, read, and reason.
More specifically, MTurk became a primary channel for Reinforcement Learning from Human Feedback — the technique that aligns large language models with human preferences. RLHF trains a “reward model” on human preference data: a person compares two AI outputs and says which is better, and the model learns to predict which responses humans prefer. MTurk workers generated enormous quantities of that preference data. Their judgments — which summary is clearer, which response is more helpful, which answer is more accurate — were encoded into reward models that shaped the behavior of deployed AI systems.
The SQuAD benchmark dataset, one of the most widely cited benchmarks in natural language processing, was built substantially through MTurk. Thousands of published academic papers in psychology, behavioral economics, and cognitive science relied on MTurk for human subject data. The platform was not peripheral to the AI industry — it was load-bearing infrastructure.
Why 33–46% of Workers Were Using AI to Do Their Jobs
The fracture arrived in June 2023, when the EPFL paper estimating LLM use by Veniamin Veselovsky, Manoel Horta Ribeiro, and Robert West captured the recursive absurdity perfectly: “Artificial Artificial Artificial Intelligence.” Their method combined keystroke detection with a synthetic-text classifier trained to distinguish human-written responses from LLM-generated ones. Applied to real MTurk responses on an abstract summarization task, it estimated that between 33% and 46% of workers were using large language models to complete their assignments and submitting the output as human work.
The platform that had spent years providing human intelligence to train AI had become a place where workers used AI to simulate the human intelligence the platform claimed to supply. The Gilardi et al. PNAS study published the same month found that ChatGPT already outperformed MTurk crowd workers on text annotation tasks by approximately 25 percentage points on average — and did so at roughly one-thirtieth the per-annotation cost. The economics that had made MTurk attractive had dissolved in both directions simultaneously: workers were gaming it, and AI had made the underlying labor cheaper to obtain directly.
The Veselovsky paper’s implications extend beyond the platform itself. Any study, benchmark, or RLHF preference dataset built on MTurk data collected after November 2022 — when ChatGPT became publicly available — now carries a material uncertainty about whether the “human” signal it captures was genuinely human. Researchers who published findings using MTurk annotation in 2023 and 2024 cannot know, from the data alone, what fraction of their labeled examples were produced by the same models those labels were used to train or evaluate. This is not a minor data quality concern; it is a structural circularity in the evidence base. The field has not yet grappled with it in any systematic way.
What Amazon Closes, and What It Replaces With
AWS’s official announcement is characteristically minimal: the decision was made after “careful consideration,” and existing workers and requesters “can continue to use the service as normal.” No final sunset date for existing users has been announced. No explanation has been offered for what happens to worker reputation histories or outstanding requester balances when the platform eventually closes entirely.
The practical replacement is already in place. AWS’s Ground Truth documentation explains that SageMaker Ground Truth, launched by Amazon in 2018, uses an active learning architecture that works differently from MTurk at a structural level: a labeling model trains incrementally as annotations accumulate, auto-labeling examples it can classify with high confidence and routing only low-confidence cases to human review. The result is a hybrid system that can reduce the human labor required for a labeling job by up to 70% compared to fully manual annotation. Ground Truth Plus, the managed-service tier, provides an end-to-end quality management layer — audit trails, confidence scores, worker agreement metrics — that MTurk’s anonymous open market never offered.
The difference is not just price or features. MTurk’s founding architecture assumed that the right way to obtain human judgment was to expose a task to an open anonymous marketplace and accept whatever came back. Ground Truth’s architecture assumes that human judgment is most reliable when it is routed, managed, and quality-controlled — with the AI handling the easy cases so human review is reserved for the hard ones. The open crowd model that MTurk pioneered is being retired not because it failed but because the industry it built outgrew it.
Beyond Ground Truth, teams migrating off MTurk have access to Scale AI and Labelbox, Surge AI, and Defined.ai, which offer specialized annotation pipelines with domain expertise, privacy controls, and quality guarantees suited to regulated industries and high-stakes AI applications.
What Is Replacing Amazon Mechanical Turk?
For teams currently running active MTurk pipelines, July 30 is a hard deadline. New requester accounts will not be accepted after that date, meaning any team that needs to expand its annotation capacity, onboard new projects, or recover from an account suspension after July 30 will have no MTurk path. “No new features” in AWS terminology has a known trajectory: it means the service is being maintained at minimum cost while the transition period plays out, not that the service is stable indefinitely.
The practical migration path depends on the use case. Teams using MTurk for machine learning data labeling should evaluate SageMaker Ground Truth for AWS-native workflows, or a vendor like Scale AI or Labelbox for larger or more specialized programs. Teams using MTurk for academic human-subjects research — surveys, behavioral experiments, cognitive tasks — have the longest migration problem, because the replacements are functionally different: Prolific, CloudResearch, and Qualtrics panels offer more controlled participant pools but lack MTurk’s API flexibility and scale. Academic review boards will need to update their protocols.
Workers who earn income from MTurk should treat the July 30 date as a soft signal rather than a hard cutoff — their accounts are not immediately affected. But the community consensus documented in worker forums is that the platform has been effectively dying for years, hollowed out by bots, by Amazon’s pattern of account closures without explanation, and by the collapse in available tasks as requesters migrated to managed alternatives. Workers who have not already diversified their platforms should begin doing so now.
Reading the Closure
The 18th-century Mechanical Turk was exposed when skeptics examined the cabinet and found the chess master inside. Amazon’s version was exposed when researchers examined the outputs and found the machine.
MTurk’s arc is remarkably complete. The platform launched to help AI appear more capable than it was. It matured into the mechanism that made AI more capable than it had been, by supplying the human judgments that trained and aligned the models. And it ended when those models became capable enough to do the labor themselves — including the labor of generating the human judgments that were supposed to evaluate them.
The workers who built the AI industry for cents per task are being replaced by the models they trained. That is not an unfamiliar story in technology. What makes the MTurk chapter unusual is that the workers knew what they were building, knew what they were worth, and built it anyway — because for many of them, it was the best option available in the regional labor markets where they happened to live.
Mechanical Turk’s retirement is, in the end, a clean and legible ending to a story that rarely gets those. The platform did exactly what it was designed to do, and the doing of it made itself unnecessary.
Frequently Asked Questions
Why is Amazon shutting down Mechanical Turk?
Amazon has not offered a public explanation beyond “careful consideration,” but the structural reasons are visible: the AI industry has shifted away from anonymous open-crowd annotation toward managed labeling services with quality controls, audit trails, and domain expertise. Amazon’s own SageMaker Ground Truth offers that managed model within the AWS ecosystem. At the same time, a 2023 study found that 33–46% of MTurk workers had begun using large language models to complete their tasks, compromising the quality of the human-generated data the platform was supposed to supply. The platform built to generate human judgment had become a conduit for AI-generated responses submitted as human ones.
What will happen to AI research that used MTurk data?
Any academic study, NLP benchmark, or RLHF preference dataset that relied on MTurk annotation collected after November 2022 — when ChatGPT became publicly available — carries an unresolved question about the authenticity of the “human” signal it captured. The Veselovsky et al. study estimated that up to 46% of responses on some task types were LLM-generated. Researchers who published findings using MTurk labels in 2023 and 2024 cannot retroactively verify what fraction of their annotated examples were produced by AI. No systematic replication or audit effort has been announced by the field’s major venues.
What is the best replacement for Amazon Mechanical Turk?
The answer depends on the use case. For machine learning data labeling and annotation within AWS, SageMaker Ground Truth is the natural successor — it uses active learning to reduce human review to the cases that genuinely require it, and Ground Truth Plus offers a fully managed quality pipeline. For specialized, high-stakes, or regulated annotation work, Scale AI, Labelbox, and Surge AI offer vetted professional workforces with stronger quality guarantees. For academic human-subjects research, Prolific and CloudResearch offer controlled participant pools with demographic targeting and IRB-compatible workflows, though they lack MTurk’s API depth and scale at the low end.
Is the July 30 deadline a full shutdown?
No. July 30, 2026 is the cutoff for new customer accounts — new requesters and new workers will not be able to register after that date. Existing accounts can continue using the platform. Amazon has not announced a final shutdown date for existing users, and the company says it will continue investing in security and availability. In AWS terminology, “Services in Maintenance” means minimum viable upkeep with no new development — a status that typically precedes full retirement, but on a timeline Amazon has not disclosed.
