Amazon Is Closing Mechanical Turk After Twenty Years

When Jeff Bezos launched Amazon Mechanical Turk in 2005, he called it 'artificial artificial intelligence', a way to hand computers the small human judgements they could not yet make on their own. Twenty-one years later the machines have caught up. Amazon told users this week that the service will shut down on September 30.

Mechanical Turk was a marketplace for what Amazon called Human Intelligence Tasks, or HITs. Companies posted small jobs, labelling an image, transcribing a clip, deciding whether two product listings described the same thing, and a global pool of workers picked them up for a few cents each. A great deal of the data that trained the first generation of modern machine-learning systems passed through hands like these. The timing carries an obvious irony, as CNBC pointed out: the platform that helped teach AI to read and see is closing partly because AI can now read and see for itself.

That is not the whole explanation. Mechanical Turk had been fading for years. It stopped taking new customers at the end of July, and the quality of its worker pool had long been a running complaint among the academics and market researchers who leaned on it for cheap survey data. Meanwhile a crop of specialised firms, Scale AI, Mercor, Prolific and others, built better-vetted, better-paid labour pools aimed squarely at training and evaluating frontier models. The work did not disappear. It professionalised, and moved somewhere Amazon was not.

The people behind the tasks

For all the talk of automation, human judgement is still woven through modern AI. Reinforcement learning from human feedback, red-teaming, safety labelling and evaluation all depend on people making calls that models cannot yet make reliably. Data annotation is a larger business today than it was when Mechanical Turk began. What has changed is who does it, under what terms, and for how much.

Mechanical Turk always sat in an awkward spot, celebrated as a clever piece of infrastructure and criticised for pay that often fell below minimum wage once unpaid searching and rejected work were counted in. Its closure closes one chapter of that argument without settling it. The hidden human layer beneath AI is not going away. It is just harder to see than a single website with a login.