Amazon shuts Mechanical Turk to new customers

Amazon announced that Mechanical Turk will stop accepting new customers effective July 30, 2026. The crowdsourcing platform that spent more than two decades supplying AI training pipelines with human-labeled data is being wound down — supplanted by the automation it helped create.

Key takeaways

Mechanical Turk closes to new customers July 30, 2026; existing users are unaffected
The platform launched in 2005, handling data labeling, content moderation, and survey tasks
AWS plans no new features — maintenance limited to security and availability
A 2023 study found that 33–46% of MTurk workers used LLMs to complete assigned tasks
Amazon is directing customers to SageMaker Ground Truth and Ground Truth Plus

A model AI built — and then replaced

Mechanical Turk — named after an 18th-century chess-playing automaton that concealed a human operator inside — launched in 2005 as Amazon's answer to a specific problem: where do you find cheap, scalable labor for tasks machines hadn't yet mastered? The platform broke large projects into microtasks — image labeling, transcription, content review — paying workers fractions of a dollar per completed unit.

In the pre-generative AI era, it was one of the industry's cornerstones. Computer vision models, text classifiers, speech recognition systems — nearly all required labeled data gathered through exactly this kind of model. MTurk supplied scalable, relatively low-cost labor to labs like the Allen Institute for AI, universities, and corporations across the industry.

The inflection point came in 2018, when Amazon repositioned MTurk as a SageMaker tool for neural network data annotation. The platform was supposed to be a bridge between human understanding and machine learning. But the paradox surfaced in 2023: an analysis found that between 33% and 46% of MTurk workers were using large language models to automatically complete the very tasks they were being paid to do as humans. Human labor inside an AI pipeline was quietly being replaced by AI.

Why there was no shortcut

Amazon tried to give MTurk a second life through multiple iterations. Integration with SageMaker, new task categories, a developer API — none of it stopped the erosion of both customers and workers. The market evolved: specialized data providers like Scale AI and Labelbox emerged with higher quality standards and more rigorous annotation management. Simultaneously, synthetic models and generative data augmentation began reducing demand for purely human labeling across many categories.

The AWS message is at once honest and terse: after "careful consideration," the decision was made to close the platform to new customers, with no planned new features, and maintenance limited to security and availability. That is standard language for a product in life-support mode that has not found a new formula.

As the alternative, Amazon points to SageMaker Ground Truth — its own fully managed data labeling tool — and Ground Truth Plus, a version with an external workforce managed by Amazon. Both are more deeply integrated with AWS AI infrastructure and offer more robust quality monitoring than open crowdsourcing through MTurk.

The symbolism of a closing era

Shutting MTurk to new customers is more than a routine product end-of-life. The platform became a symbol of early AI's underlying philosophy: machines are nearly sufficient, but we need humans to fill the gaps. That model held for two decades. It generated thousands of academic studies, dozens of ethical controversies about pay and labor conditions, and — paradoxically — vast quantities of training data that enabled AI models to cross the capability threshold at which platforms like MTurk become unnecessary.

There is also an irony worth noting: MTurk was one of the hidden foundations for AI products that marketed themselves as fully automated. The term "Mechanical Turk" became shorthand in the industry for products that look like AI from the outside but are actually powered by humans inside. In an era when that boundary has been genuinely dissolved by real automation, a new problem has emerged: models used to complete MTurk tasks are contaminating the pools of human-labeled data intended for further AI training.

Why this matters

The closure of Mechanical Turk to new customers signals a structural shift in how the AI industry sources training data. For years, MTurk was a lower-cost alternative to internal annotation teams — available on demand, without long-term contracts. Its decline confirms that this model does not scale to the demands of the current generation of models: insufficient data quality, limited quality control, too much risk of contamination by synthetic responses. At the same time, it is pushing the market toward two poles: proprietary in-house data teams like those built by OpenAI, Anthropic, and Google, and specialized B2B platforms with full quality management. Small companies and startups lose their cheapest entry point to labeled crowdsourcing. Academia loses easy access to large-scale survey data — MTurk was a standard instrument in hundreds of studies in experimental psychology and behavioral economics.

What's next

AWS will maintain the platform for existing customers without a stated shutdown date — no full closure date has been announced
SageMaker Ground Truth Plus becomes the default path for enterprise customers seeking managed data labeling workforces
Research studies relying on MTurk data will likely require renewed methodological scrutiny given the degree of LLM contamination in samples collected between 2022 and 2026