Amazon has quietly stopped accepting new customers for Mechanical Turk, its two-decade-old platform that let businesses pay real humans tiny amounts to do tiny tasks — labelling images, transcribing audio, answering surveys. The writing has been on the wall for a while, but the formal closure to new sign-ups marks the effective end of an era in how the internet thought about human-in-the-loop labour.

MTurk launched in 2005 and was, in its heyday, genuinely novel. It industrialised what Amazon called "artificial artificial intelligence" — tasks that computers couldn't do well, so humans did them cheaply at scale. At its peak it had hundreds of thousands of workers, known as "Turkers," completing millions of microtasks. Academics loved it for cheap survey panels. AI companies loved it for training data labelling.

That second use case is what makes this closure interesting. MTurk didn't die because demand for human-labelled data disappeared. It died because the work moved somewhere else — and, increasingly, because AI is doing the labelling itself.

The irony is thick. MTurk spent years helping companies build the training datasets that trained the AI models that are now replacing the need for MTurk. It's the most literal version of "the workers built the machines that took their jobs" the tech industry has produced.

The broader "human cloud" — platforms like Scale AI, Surge AI, and Remotasks — has consolidated heavily over the past two years. Scale AI, now valued in the tens of billions, essentially ate MTurk's AI-training market by offering higher-quality annotators at higher prices. The commodity end of the market, where MTurk lived, got squeezed from both sides: AI automation eating the simpler tasks, and better-funded competitors taking the complex ones.

For researchers, this matters. MTurk was a fixture of behavioural economics, psychology, and social science research — cheap, fast, and accessible to small labs. Its effective closure will push academic users toward more expensive alternatives or, more likely, toward AI-generated synthetic survey responses, which raises its own validity questions. Several universities have already begun updating their ethics guidelines for studies that relied on Turker panels.

For the AI industry in Australia, the shift is already visible. Australian AI companies — including several in the federal government's AI Connect program — have been moving annotation work to domestic providers or hybrid AI-human workflows. The ACS and CSIRO's Data61 have both flagged data labelling as an emerging area of domestic strategic interest, given the sensitivity of training data for public-sector AI deployments.

What to watch: whether Amazon winds down the existing Turker workforce formally, or simply lets the platform atrophy. MTurk has thousands of active workers, many of whom rely on it for meaningful income in lower-cost economies. Amazon hasn't announced a sunset date — just a "no new customers" sign on the door.

That's not a closure. It's a slow fade. And for a platform that quietly powered so much of the internet's AI layer, it deserves more than a footnote.