Amazon is shutting down Mechanical Turk on September 30, ending a 21-year run for the marketplace that arguably did more to build modern machine learning than any single piece of software did. The notice on the site is corporate to the point of blankness: “We regularly evaluate our programs, tools and services and make adjustments based on those assessments.”
The shutdown was telegraphed. New customer signups closed on July 30, alongside two adjacent AWS products — SageMaker Ground Truth and Augmented AI — that did roughly the same job with more enterprise packaging. Existing requesters and workers get until the end of September.
MTurk launched in 2005 as a place to post Human Intelligence Tasks: the small, tedious jobs that were trivial for a person and impossible for a computer. Transcribe this audio clip. Is there a cat in this photo. Rate whether these two sentences mean the same thing. Bezos named it after an 18th-century chess automaton that turned out to have a man hidden inside, and called the service “artificial artificial intelligence.” The joke was that the intelligence was real and the artificiality was the interface.
The labelled data came first
Almost every capability people find impressive in a 2026 model rests on a foundation of human judgment that someone had to buy by the piece. Image recognition needed millions of labelled photographs before it worked at all — ImageNet, the dataset that kicked off the deep-learning decade, was annotated by Turkers. A generation of social-science papers used MTurk as its subject pool because it was the cheapest way to get a few hundred human responses by Friday.
Then the tools those workers helped build showed up on their own screens. A 2023 study by researchers in Switzerland estimated that up to 46% of MTurk workers were using large language models to complete text tasks. That finding is the whole story in one number: the marketplace existed to supply human data, and its supply had quietly become synthetic. Once you cannot tell whether you are buying a person’s judgment or a model’s, you are no longer buying the thing you came for.
Our take: This is not AI killing a job category so much as AI dissolving the boundary that made the category sellable. The demand for human annotation did not vanish — Scale AI, Mercor and Prolific are all growing by recruiting for exactly this work. What changed is that buyers now pay a premium for verified, credentialed, expensively-sourced humans, because commodity human data at pennies a task can no longer be trusted to be human. The floor of the market disappeared. The ceiling got more expensive.
What to watch
- Anyone with a pipeline on MTurk has five weeks. If you run data collection, research panels or QA through it, the migration window closes September 30. Prolific is the closest substitute for survey and research work; Scale and Mercor sit at the higher-skill end.
- Academic replication takes a hit. Thousands of published studies used MTurk as their sampling frame. Reproducing them on a different panel is not a like-for-like exercise, and the field has not settled on a successor.
- Watch the price of verified humans. The differentiator in data labelling is no longer volume, it is provenance — proof that a person with a specific credential actually did the work. That is where the margin has moved.
The Turk was a hoax with a man inside. The service named after it ran for two decades on the same premise, honestly disclosed. It is closing because the machine finally got good enough to sit in the box on its own — and good enough that nobody could be sure it hadn’t already.