Amazon's 20 year crowdsourcing platform, where humans did small online tasks for pennies, stops accepting new customers on July 30. By 2023, much of its remaining workforce used AI to do the work itself.
Amazon's Mechanical Turk, the 20-year crowdsourcing platform where paid humans did the small online tasks that automated systems still could not handle, stops accepting new customers on July 30. The platform had been hollowing out for years. By 2023, researchers estimating using survey data found that between 33% and 46% of active Turkers were completing tasks with the help of large language models.
For most of its life Mechanical Turk ran in the opposite direction. It launched in 2005 as a tiny labor exchange Amazon called "artificial artificial intelligence," borrowing its name from an 18th-century chess automaton whose trick was a hidden human inside the cabinet. Through the 2010s it became the default venue for crowdsourced data annotation: image labels, sentiment tags, and the small human judgments that became training data for the first wave of commercial AI. In 2018, Amazon repositioned the platform as the on-demand human workforce for SageMaker Ground Truth, the AWS service that sells labeled data to machine learning pipelines. AWS says it will keep existing customers running past the cutoff with only security and availability work.
The marketplace had already been thinning. Phelim Bradley, chief executive of Prolific, a research-recruitment company founded explicitly as an MTurk alternative, wrote on X that the closure reflected years of quality decline, not a sudden decision. CloudResearch, the vendor whose MTurk Toolkit most academic labs used to access the worker pool, said it is retiring the toolkit because researcher usage had fallen from thousands per day at peak to single digits per day; most academic and enterprise customers had already moved to its Connect, Engage, and Sentry products. AWS' public framing on its own SageMaker documentation page was that the decision followed "careful consideration," with no new features planned.
The MTurk closure sharpens a segmentation that was already visible. A wave of bot traffic, fraud, and worker flight had pushed academic survey work onto vetted alternatives; Prolific's founding premise was effectively what MTurk had stopped reliably being. Enterprise data-labeling demand, by contrast, has consolidated inside AWS and other vendors with managed workforces and quality guarantees. The marketplace that goes dark on July 30 is the open-call, no-curation version, which had become the smallest slice of the labeled-data market by usage and the largest by reputational damage.
TechCrunch and The Register confirmed the July 30 cutoff independently. No Amazon strategic memo is on the public record; the analytical interpretation is being supplied by competitors and migration partners, whose incentives are visible, and the people closest to the marketplace read the announcement as punctuation on a multi-year process.
The mechanism survives the platform. The closure is not proof that human-in-the-loop AI is finished; it is a clean, documented instance of one specific pattern, where a workforce is consumed by the thing it helped build. The watch item is what AWS announces in MTurk's place before or after the cutoff: a sold labeled-data service that bundles managed workforces with Ground Truth, or a model-first pipeline with humans only on the edges.