SEATTLE — The marketplace where humans once performed tasks that computers could not is closing its doors to new customers. Amazon said it will no longer accept new registrations on Mechanical Turk, the crowdsourcing platform it launched in 2005 and that Jeff Bezos once described as “artificial artificial intelligence.” The decision, first reported by TechCrunch, puts the platform into maintenance mode: existing requesters and workers can keep using it, but Amazon will not add new features, and the business that defined an entire category of online work is effectively being wound down.
Mechanical Turk was a product of its time, and its decline is a product of this one. The platform paid people pennies to tag images, transcribe audio, and categorize products — tasks that were trivial for humans and impossibly hard for computers when it launched. At its peak, it connected more than 500,000 registered workers, known as turkers, to a global stream of micro-jobs, and it became a crucial source of the labeled data that trained the first generations of machine-learning models.
The irony is that the technology those workers helped create has now made much of their work obsolete. Large language models can perform the labeling, classification, and text tasks that were Mechanical Turk’s bread and butter, and the quality of AI-generated output has improved to the point where many requesters no longer need human review. Amazon’s own research teams, once among the platform’s biggest customers, now train models on data produced by other means.
The squeeze has been visible for years. Researchers at the Swiss institute EPFL found in a 2023 study that between a third and half of surveyed turkers were completing text-based tasks with the help of large language models — humans being paid to relay machine output. Requesters meanwhile complained of bots and fraud degrading response quality. The economics of the platform deteriorated from both directions: the work that remained was increasingly done by machines, and the human work that was left was increasingly unreliable.
The vacuum has been filled by a new generation of companies. Scale AI, Mercor, and Prolific have built businesses around recruiting workers specifically to train and evaluate AI models, with better pay and more structured pipelines than Mechanical Turk ever offered. These companies have absorbed much of the demand that once flowed through Amazon’s platform, and they operate in a market Amazon has largely ceded — its own SageMaker Ground Truth service remains a distant option rather than a priority.
For the workers who depended on Mechanical Turk as a source of flexible income, the freeze is the end of an era that never quite delivered on its promise. Turking was famously poorly paid — most tasks paid cents and took minutes — and workers long complained about arbitrary rejections, opaque requester behavior, and the absence of any safety net. The platform’s defenders argued it provided income to people in places with few alternatives; its critics said it was a warehouse job without a warehouse.
Amazon has not said what will happen to the platform’s remaining users or when it might close entirely. The company said existing customers can continue to use the service, and that it will keep the infrastructure running while it decides. Business customers are already moving to alternatives, and the remaining requesters are expected to follow.
Mechanical Turk’s place in the history of AI is secure. The crowdsourced labeling it pioneered built some of the most important datasets of the machine-learning era — the ImageNet image collection that helped spark the deep-learning boom was assembled with Turk labor — and the platform became the default answer to a recurring problem: getting reliable human judgment at scale. That problem has not disappeared; the market for human judgment has simply moved to platforms that pay more and specialize in AI training.
Turking built a community as much as an economy. Workers shared strategies in forums, researchers studied the platform’s labor practices, and the word ‘turk’ entered the vocabulary of both AI and the gig economy. The freeze on new registrations does not end the platform overnight — existing workers can keep working — but it ends its future, and with it the sense among workers that the marketplace would be around for the long run.
The work itself is being reinvented. AI companies now pay for human feedback on model outputs, for the preference data that trains reinforcement-learning systems, and for the verification of AI-generated content — tasks that resemble turking but are priced, managed, and delivered differently. The companies that do this work well have become quietly important parts of the AI supply chain, and they have absorbed the role Mechanical Turk once played.
The freeze is a small event for Amazon, which earned more in a single quarter than Mechanical Turk generated in most years. But it closes a chapter in the history of AI itself. The platform was built on the assumption that human judgment would always be needed for the tasks machines could not do; the machines caught up, and the platform that depended on the gap between them is the casualty. The work did not disappear — it moved to companies that pay better, and to AI systems that no longer need to ask.
