There’s a particular kind of irony that only the technology industry can manufacture: building a workforce of half a million humans to train the very machines that will make them obsolete. Amazon Mechanical Turk, which launched in 2005 and is permanently shutting down on September 30, is that story in miniature.
The platform, operated under Amazon Web Services, matched “Turkers” across 190 countries with Human Intelligence Tasks — or HITs. Data labelling. Transcribing audio. Answering surveys. Categorising images. Each task paid a few cents. The name came from an 18th-century chess-playing automaton called The Mechanical Turk, which appeared to be a thinking machine but had a human chess master hidden inside. The metaphor was always apt.
Jeff Bezos once described the service as “artificial artificial intelligence” — a remarkably self-aware label for a system that farmed out tasks computers couldn’t do to people who needed the money.
The slow decline nobody noticed
Mechanical Turk wasn’t killed by a single decision. It was eroded by years of neglect. According to Krista Pawloski of the data-worker advocacy group Turkopticon, Amazon had been pulling back resources for years. Fewer improvements, less investment, while rivals like Scale AI, Mercor and Prolific arrived with better interfaces, fairer pay, and AI-focused recruitment pitches.
The platform tried to reposition itself as a data annotation source for Amazon’s own SageMaker ML platform, promising the fastest turnaround for human review jobs with workers available around the clock. But the clock, as they say, was ticking. Last month, Amazon stopped accepting new customers. Workers on Reddit saw it coming.
The humans were already being replaced by bots
Here’s the part that might surprise nobody who has paid attention to AI for the last few years: a 2023 study by Swiss academic researchers found that up to 46% of Mechanical Turk workers were already using AI models to complete their tasks. The platform existed because certain jobs were easy for humans and hard for machines. The gap closed. And the workers, who needed the income more than Amazon needed their particular labour, adapted by using the very technology the platform was training.
The quality implications of that were significant — data labelled by an LLM, submitted by a human, used to train another LLM — but perhaps not as significant as the broader point. The human layer of AI development was always temporary. It was a bridge between what machines could do in 2005 and what they could do in 2026.
What comes next
Amazon’s closure notice is characteristically understated. A banner across the MTurk homepage reads: “Following an assessment, we’ve made the decision to close AWS Mechanical Turk, effective September 30, 2026.” Workers are pointed to an FAQ about resolving billing issues. That’s the whole statement.
There was no separate press release, no tribute to the quarter-century of crowdsourced labour. The service that powered much of the early boom in AI training data — the human annotation that taught models to recognise cats, transcribe speech, and moderate content — is just being turned off.
The 500,000 workers who logged in, completed a few cents’ worth of HITs, and moved on are going to need to find other platforms. Scale AI, Mercor and Prolific are picking up the demand for human-in-the-loop data work. But the scale was never going to stay — AI is now good enough at most annotation tasks that the human premium has collapsed.
Computer scientist Jaron Lanier wrote about Mechanical Turk years ago, warning that it was designed to turn people into software components for larger business initiatives. He wasn’t wrong. The workers were always the hidden mechanism inside the machine — the human chess master in the wooden cabinet. Now the machine can play chess by itself.
It’s the kind of neat arc that makes you wonder whether anyone at Amazon felt a flicker of “well, this is inevitable” when they pressed send on that banner. Or whether it was just another quarterly assessment with a spreadsheet outcome.
[Sources: The Next Web, TechSpot, CNBC]
