Amazon Mechanical Turk Closure: Dates, Payments and AI Labor

A marketplace for tasks computers couldn't do

Amazon Mechanical Turk launched in 2005 as a marketplace where businesses posted small digital tasks and a crowd of workers completed them for a few cents apiece: transcribing audio, tagging images, filling out surveys, checking whether two product listings described the same item. Jeff Bezos called it "artificial artificial intelligence," work that people could do easily and computers could not, routed through an API as if it were a machine service. The name itself borrowed from an 18th-century chess automaton that appeared to play on its own but was secretly operated by a hidden human. [2]

The platform grew out of Amazon's own need to clean up its e-commerce catalog and, at its peak, drew more than 500,000 registered workers. Its more consequential use turned out to be outside retail: MTurk became a standard way for researchers and companies to produce the labeled datasets early machine learning systems trained on, before dedicated AI-training-data vendors existed. [2]

What is closing, and when

Amazon stopped accepting new Mechanical Turk customers on July 30, 2026, continuing the service for existing requesters but limiting further work on it to security and availability fixes. On August 25, 2026, Amazon confirmed the platform would permanently close on September 30, 2026, and that the closure also covers SageMaker Ground Truth and Amazon Augmented AI, the two services built for labeling data and routing human review into machine learning pipelines. [3][1]

Amazon's wind-down notice sets out the mechanics rather than a rationale: workers keep being paid for completed HITs on the existing schedule, bonuses can still be issued through October 30, 2026, requesters have 30 days after closure to approve or reject pending work, transaction history stays visible until January 28, 2027, and unused requester balances are refunded within 30 days. [1]

The work didn't disappear, it moved

Amazon's decision followed years of MTurk's own workforce becoming an unreliable source of human-generated data for exactly the purpose it was increasingly used for: training AI. By one estimate cited in coverage of the closure, between a third and nearly half of Mechanical Turk workers were using large language models to complete tasks meant to produce human-labeled data, undermining the reliability of the output for the AI training and evaluation work that had become the platform's main remaining use. [3]

Newer, specialized platforms including Scale AI, Mercor and Prolific built businesses specifically around recruiting and vetting workers for AI training and evaluation, rather than treating it as one category of task among many. Coverage of the closure describes these vendors drawing workers away from MTurk with work built for the AI economy, while Amazon's service went years without meaningful investment. [2]

Atlas interpretation: Mechanical Turk's model was a generalist crowd: post any small task, draw from the same undifferentiated pool of workers, and let requesters sort out quality themselves. That fit surveys and catalog cleanup well enough. AI training and evaluation is a specific enough job, with its own quality controls, worker vetting and, in some cases, subject-matter expertise, that dedicated vendors could out-compete a generalist marketplace at the one job MTurk had drifted toward being known for. The underlying demand for human-labeled data did not shrink; it moved to companies built around that demand specifically, even as Amazon's do-anything version of the same idea shut down. [2][3]