Amazon Mechanical Turk Closes to New Customers July 30 After 21 Years

Amazon Mechanical Turk will close to new customers on July 30, AWS confirmed this week, ending 21 years of new signups for the crowdsourcing marketplace that spent its final chapter as an accidental exhibit for everything complicated about human labor and machine learning. Existing customers keep their access. No shutdown date is set. AWS is still investing in security and availability, but ruled out new features - a description that fits a service wound down in slow motion rather than killed outright.

A 2005 Marketplace Built to Simulate AI, Later Repurposed to Build It

Mechanical Turk launched in 2005 as a solution to a problem automation had not yet solved: tasks too simple to require expertise but too varied for scripts to handle reliably. Workers completed CAPTCHA challenges, tagged images, and classified the sentiment in short text samples - paid fractions of a cent per task. Amazon named it after the 18th-century chess hoax, where a "mechanical" opponent turned out to be a human hidden inside the machine. That turned out to be a more accurate name than anyone intended.

Starting in 2018, Amazon began positioning Mechanical Turk as a data annotation layer for SageMaker, its AI training platform. Teams could label training data at scale without hiring a permanent annotation workforce. For a few years, that pitch gave the marketplace a second life - not as a quirky crowdsourcing tool but as infrastructure feeding model development pipelines at companies that could not afford a dedicated data team.

Between 33% and 46% of Workers Were Using LLMs to Complete Tasks by 2023

A 2023 analysis found that between 33% and 46% of Mechanical Turk workers were using large language models to complete the tasks they were being paid to perform as humans. That finding landed like a delayed punchline. Workers hired to generate human-labeled training data for AI models were submitting AI-generated labels instead. Researchers flagged the reliability problem: outputs clustering around similar LLM patterns could introduce systematic biases into whatever models that data trained.

For AI teams using Mechanical Turk to label data, that finding meant the human-in-the-loop guarantee they were paying for was already gone. SageMaker Ground Truth - Amazon's more controlled annotation product with private labeler pools and quality auditing - handled that concern better from the start. Mechanical Turk trained AI, got used by AI workers to fake human output, and that circularity probably explains more about its decline than any official statement will.

The Platform Had a Stranger Role: Making Fake AI Products Look Real

Before the data annotation pivot, Mechanical Turk had a quieter reputation as infrastructure for products marketed as AI that quietly routed requests to human workers completing tasks on demand. Researchers described this as the Potemkin AI problem - software presenting a computational facade while humans performed the actual classification, sorting, or analysis underneath. Amazon had named the platform after a chess-playing automaton that turned out to conceal a human operator. Products built this way made the metaphor literal.

Mechanical Turk also appeared at an early stage of the Facebook-Cambridge Analytica scandal, where tasks completed on the platform helped gather and process data used in psychological profiling. That connection was peripheral, but it illustrated how a microtask marketplace with no visibility into end-use cases could become part of data collection operations with real consequences. AI displaced 31% of all job cuts in June alone - but Mechanical Turk's decline ran the other way: AI eliminated the demand for human microtasks the platform supplied.

SageMaker Ground Truth Handled What Mechanical Turk Started

Amazon's replacement is SageMaker Ground Truth, which offers managed labeling workflows, private workforce options, and built-in quality controls that Mechanical Turk's open crowdsourcing model never had. Ground Truth launched in 2018 - the same year Amazon started pitching Turk for AI training. Running two overlapping data annotation products for eight years, while one absorbed the other's use case, is a product strategy Amazon has applied before. Both products usually survive; one tends to stop growing.

No announcement accompanied the closure. AWS posted notice to the Mechanical Turk website and the SageMaker documentation, offering "careful consideration" as the full explanation and nothing more specific. At peak, the platform reportedly hosted hundreds of thousands of workers globally, with large concentrations in India and the United States. Amazon has not disclosed how many remain active. Whether July 30 is a step toward full shutdown or simply the steady state Mechanical Turk will hold indefinitely, Amazon did not say.