Amazon Is Killing Mechanical Turk, the Human Engine That Helped Train AI

Amazon told workers and customers this week that Mechanical Turk shuts down on September 30, ending a 21-year run for the marketplace that did the grunt work behind the machine learning era. Jeff Bezos once described it as artificial artificial intelligence, a joke about humans pretending to be machines. The actual machines made the joke obsolete.

If you never used it, MTurk paid people a few cents at a time to do small tasks a computer couldn't: labeling images, transcribing audio, filling out surveys, moderating content. More than 500,000 people passed through it. A lot of the training data underneath the AI you use every day started as somebody clicking through those tasks for pennies.

The wind-down was already happening

New customer sign-ups stopped on July 30, the same day Amazon closed SageMaker Ground Truth and Amazon Augmented AI to new users. September 30 is the last step in a shutdown that's been in motion for a while. Amazon offered no explanation beyond a routine review of its programs and services, and pointed remaining customers to an FAQ about refunds and payout settings.

That leaves about five weeks for anyone still piping data through MTurk to move their pipeline somewhere else.

What killed it

Two things, working from opposite ends. At the high end, specialized annotation firms like Scale AI, Mercor, and Prolific took the valuable work, because frontier labs need vetted, expert-level labeling and pay real money for it. At the low end, language models absorbed the cheap classification jobs that made MTurk famous in the first place.

Academic work in 2023 found a meaningful share of MTurk workers were using AI to complete tasks that existed to train AI. Once the humans in a human-in-the-loop system start outsourcing to the loop, the system stops being what you were paying for.

The part that isn't funny

Half a million people used this platform, and for some of them it was real income. It was never good income, and MTurk drew years of well-earned criticism for how little it paid and how little recourse workers had when a requester rejected their work. Both things are true at once: the platform was exploitative, and its disappearance hurts people who relied on it. The second one gets less attention.

Data work isn't going away. It's moving into contracts, vetted panels, and specialist firms, which is probably better for the people doing it and definitely better for the labs buying it. What ends on September 30 is the penny task exchange that made it cheap.