Amazon Shuts Down Mechanical Turk As Human Data Work Enters The AI Age

Amazon is shutting down Mechanical Turk, and the timing says a lot about how much the AI labour market has changed.

Amazon’s own MTurk FAQ now says Mechanical Turk will permanently close on September 30, 2026. Workers and requesters can continue using the service until that date, requesters will have 30 days after closure to approve submitted work, bonuses can be awarded until October 30, and transaction history will remain available until January 28, 2027.

The closure also affects Amazon SageMaker Ground Truth and Amazon Augmented AI where the Mechanical Turk worker type is used. That makes this more than the shutdown of an old crowdsourcing website. It touches the human-review layer that many machine-learning workflows once used for labelling, moderation, validation and quality checks.

Mechanical Turk was launched in 2005, long before the current generative AI boom. Its original pitch was clever and blunt: let software call on humans to do tasks that computers could not yet handle well. Amazon called that model Artificial Artificial Intelligence, because the work looked automated from the outside but was powered by distributed human judgement behind the scenes.

For years, that made MTurk useful to researchers, startups and companies that needed cheap, scalable human input. It helped with image labels, survey responses, transcription checks, search relevance, product categorisation and countless small decisions that were too messy for earlier software systems.

But the market around it has changed. Today’s AI companies need higher-quality datasets, domain experts, safety evaluators, red teams, coding specialists and synthetic data pipelines. Data annotation has become more specialised, better financed and more tightly connected to model training. Companies like Scale AI, Mercor and Prolific have also changed what customers expect from human data work.

At the same time, AI itself has started eating some of the low-end tasks that once fed MTurk. Basic classification, summarisation, transcription and content sorting can now be handled by models at scale, often with humans reviewing edge cases rather than doing every task manually.

That does not mean humans are leaving the AI loop. It means their role is changing. The most valuable human work is moving away from clicking through microtasks for cents and toward expert evaluation, preference ranking, adversarial testing and workflow judgement. AI still needs people, but it increasingly needs the right people in the right places.

There is a wider lesson here for the AI economy. The first wave of digital labour platforms made human intelligence programmable. The next wave is making human judgement more expensive, more specialised and more central to safety. That matters as AI labs push models into security, law, finance and workplace agents, areas where cheap generic labelling is not enough.

Mechanical Turk’s closure is therefore symbolic. A service built for the age when humans filled the gaps in software is closing in the age when AI fills many of those gaps itself. The new question is not whether humans remain necessary. It is where their judgement becomes too important to hide behind the machine.