Amazon Is Shutting Mechanical Turk After 21 Years

Amazon is closing Mechanical Turk on 30 September 2026, retiring after 21 years the marketplace that Jeff Bezos once described as "artificial artificial intelligence." The company had already stopped taking new customers in July, and its statement is the corporate equivalent of a shrug: after an assessment, it said, it decided to close AWS Mechanical Turk effective that date. At its peak the platform counted more than 500,000 workers worldwide.

Launched in 2005 and named for an 18th-century chess-playing automaton that concealed a human operator, MTurk was built on an honest inversion of the AI pitch. Rather than have software do work people found tedious, it routed the work computers could not yet handle — image labelling, transcription, data validation, content moderation, survey responses — to a distributed pool of people paid by the task, often cents at a time. For the better part of two decades it was quiet infrastructure: the place where academic social science ran its subject pools, and where a great deal of the labelled data underpinning the machine-learning era was assembled.

What killed it is the thing it helped build. The tasks that made MTurk useful are now largely the tasks a competent model does for a fraction of the cost — classify this image, transcribe this clip, is this sentence positive. Meanwhile the labelling market moved upmarket and away from micro-tasks entirely: Scale AI, Mercor and Prolific built businesses around supplying AI labs with vetted specialists — doctors, lawyers, competitive programmers — for expert preference data and evaluations, which is where the money in human annotation went. Amazon, by most accounts, put progressively fewer resources into MTurk as that shift played out.

The platform's data quality had also become its own problem. A 2023 study by Swiss researchers found that roughly 46% of MTurk participants appeared to be using AI to complete the tasks they were being paid to do by hand — a finding that undercut both the labelled-data business and the academic survey work that depended on the responses being human. A crowdsourcing marketplace where half the crowd is quietly calling a model is not a crowdsourcing marketplace, and for researchers who had built a decade of methodology on the assumption of human subjects, it was closer to a crisis than a curiosity.

The workers are the part that does not resolve neatly. MTurk was never a good job by most measures, but it was a flexible one available from anywhere, and some people used it as a primary income rather than pocket money; several have said publicly that they have no comparable replacement lined up. There is a tidy narrative available here about automation closing the loop on the humans who trained it, and it is mostly true — though it undersells how much of the decline was simply Amazon losing interest in a low-margin service while competitors redefined what the customers wanted to buy. The automaton that gave the service its name worked because a person was hidden inside it. Two centuries later the trick finally runs without one, and the shop is closing.