Amazon Mechanical Turk Shutdown: Why It's Closing

Amazon is shutting down Mechanical Turk, the crowdsourced-labor marketplace it launched in 2005, with the platform set to close permanently on September 30, 2026. The company confirmed the wind-down after stopping new-customer sign-ups on July 30, and framed the decision as the result of an internal assessment of its programs and services. The closure ends one of the longest-running experiments in human-powered digital work — and a service whose data helped train much of the machine-learning era that is now displacing it.

The move was reported across trade press on August 25 and 26, 2026. It marks the retirement of a platform that founder Jeff Bezos once described as “artificial artificial intelligence” — a phrase that captured its core idea: software calls a human, invisibly, to do the parts a computer could not.

How Mechanical Turk worked

Mechanical Turk — commonly shortened to MTurk — was a marketplace for what Amazon formally called Human Intelligence Tasks, or HITs. Businesses and researchers posted small, discrete jobs: labeling an image, transcribing a snippet of audio or video, moderating a piece of content, completing a survey, or judging whether two records referred to the same thing. Distributed workers picked up those tasks and were typically paid a few cents each.

The name was a reference to the 18th-century “Mechanical Turk,” a chess-playing automaton that concealed a human operator inside. Amazon’s version inverted the trick openly: the platform let a software system farm out the steps it could not yet automate to a crowd of people, then stitch the results back into an automated pipeline. At its peak the marketplace connected requesters with more than 500,000 registered workers worldwide.

For nearly two decades that model filled a specific gap. Machine-learning systems need labeled examples to learn from, and in the years before large models could generate or annotate their own training data, MTurk was one of the cheapest and fastest ways to produce it at scale. A great deal of the labeled data that trained early computer-vision and natural-language systems — including the human preference judgments that underpin techniques like reinforcement learning from human feedback — passed through platforms built on exactly this kind of micro-task labor.

What else is closing

The shutdown does not stand alone. On the same July 30 date that MTurk stopped accepting new customers, Amazon also closed two adjacent services to new users: SageMaker Ground Truth, its managed data-labeling product, and Amazon Augmented AI (A2I), which routed low-confidence model predictions to human reviewers.

Together the three services formed Amazon’s stack for getting humans into machine-learning workflows — raw micro-tasks at the bottom, a managed labeling layer above it, and human-in-the-loop review bolted onto model inference at the top. Retiring all three signals that Amazon is stepping back from operating general-purpose human-labor infrastructure, not merely sunsetting one aging product. SageMaker Ground Truth, built in 2018, had been positioned as the more modern successor, using machine learning to pre-label data and cut annotation costs by as much as 70% compared with fully manual review; its closure to new users suggests even the managed version no longer fit Amazon’s plans.

Why now: AI ate the platform it was built to fake

The immediate cause is competitive and technological. A new generation of specialized data-labeling vendors — including Scale AI, Mercor, and Prolific — has captured much of the demand that once flowed to open micro-task marketplaces, offering vetted workforces, quality controls, and domain expertise that a raw HIT board could not match. As frontier-model training shifted toward higher-quality, expert-annotated data and complex preference labeling, the economics moved away from anonymous crowds paid pennies per task.

There is a sharper irony underneath the shutdown. A widely cited June 2023 study by researchers at EPFL found that between 33% and 46% of MTurk workers were themselves using large language models to complete text-production tasks. The paper — pointedly titled “Artificial Artificial Artificial Intelligence” — described the finding as a “canary in the coal mine” for the integrity of human-generated data. The platform built to supply the human judgments that machines lacked was, by then, partly staffed by people quietly routing those judgments back through machines. The value of a crowd-sourced “human” label collapses once you can no longer be sure a human produced it.

That dynamic matters well beyond Amazon. The entire supply chain for AI training data rests on provenance — knowing where labels came from and trusting that they reflect genuine human input. As models grow capable enough to generate plausible annotations, the distinction between human and synthetic data blurs, and platforms that cannot guarantee provenance lose their reason to exist. The industry’s move toward licensed, verified data sources — visible in deals like the Getty Images–OpenAI arrangement — is the other side of the same shift.

What it means

Mechanical Turk’s closure is a small line item for a company Amazon’s size, but it is a clean marker of a larger transition in how AI systems are built.

The micro-task era is ending, not pausing. For twenty years, the default way to inject human judgment into software at scale was to atomize it into pennies-per-task work and distribute it to an anonymous crowd. That model is being replaced on two fronts at once: by AI systems capable enough to handle the tasks themselves, and by specialized vendors selling curated, higher-trust human labor for the tasks that still require people. MTurk sat in the shrinking middle — too commoditized to compete on quality, too human-dependent to compete on cost.

Winners: the dedicated labeling companies — Scale AI, Mercor, Prolific and their peers — that now own the demand for high-quality, verifiable annotation, along with the AI labs willing to pay for provenance they can stand behind. Losers: the workers who relied on MTurk for supplemental income, who lose an established venue on relatively short notice, and the long tail of academic researchers who used the platform as cheap infrastructure for surveys and behavioral studies. Amazon itself absorbs a modest reputational cost for retiring a Bezos-era icon, but frees itself from operating a labor marketplace that no longer aligns with where its AWS and AI business is headed.

What to watch next. The provenance problem MTurk’s own workers exposed does not disappear with the platform — it moves upstream to whoever supplies training data now, and the market will increasingly reward vendors that can prove their labels are genuinely human. Watch how displaced workers migrate to competing platforms, whether Amazon points existing requesters toward a narrower internal offering, and how the labeling vendors that inherited this demand handle the same contamination risk at larger scale. The task that killed Mechanical Turk — telling human work apart from machine work — is now everyone’s problem.