Amazon Is Shutting Down Mechanical Turk Because Its Own Workers Started Cheating With AI

Amazon told Mechanical Turk customers this week that the platform closes for good on September 30, ending a marketplace that ran for 21 years and once counted more than half a million workers worldwide, according to Tech Startups. Jeff Bezos launched it in 2005 and called it "artificial artificial intelligence," a joke that aged into something closer to a warning label. Amazon's public explanation was the blandest possible line about periodically reviewing its programs and adjusting course. The real reason is more specific, and more interesting, than that.

Mechanical Turk sold one thing: cheap, anonymous humans answering questions a computer could not yet answer, a few cents at a time. That business model needed one property to hold, that the person on the other end of a Human Intelligence Task was actually doing the thinking. By 2023, per figures reported by TechCrunch and corroborated by Stack Archive, somewhere between roughly a third and nearly half of crowdworkers in one study were running their answers through large language models before submitting them. Amazon had already quietly stopped accepting new Mechanical Turk customers back on July 30, easing the platform into a wind down rather than closing it outright, per TechCrunch's reporting from that period. The people paying for verified human judgment were increasingly buying AI output laundered through a human middleman, at a markup, with none of the accountability of either a real annotator or a real model provider.

That is a quality collapse, not a labor collapse. Turkopticon advocate Krista Pawloski told reporters the platform had been fading for years, as Amazon put fewer resources into fraud controls and worker vetting while both researchers and workers drifted to competitors, per the Tech Startups report. Those competitors did not sit still. Scale AI, Mercor, Surge AI, and Prolific built businesses on the opposite premise: identifiable specialists, sometimes programmers, scientists, and other domain experts, evaluating and correcting frontier models for a fee that reflects real expertise rather than a microtask rate. That is where the demand for human judgment went. It did not evaporate. It moved to vendors who could still prove a human did the work.

Here is the honest steelman against reading this as a tidy market success story. A marketplace that gave over half a million people around the world a way to earn money, however small, is closing on two months' notice with a form-letter explanation. If you are inclined to distrust large, unregulated platforms, this looks like the pattern to worry about: a company builds a labor market, extracts value from it for two decades, then sunsets it the moment it stops being useful, with no obligation to the people who built its reputation one task at a time. A critic could reasonably ask why a company this large gets to walk away that cleanly, and whether workers deserved more runway.

That criticism has a real point about notice and dignity, and I am not going to wave it off. But it does not change what actually killed the product, and forcing Amazon to keep Mechanical Turk running would not have fixed the underlying defect. The platform's problem was not merely decline, it was fraud baked into the data it sold. Every dataset built on a pool where close to half the answers might secretly be AI generated needs independent re-verification before anyone can trust it, and that verification cost stacks on top of the original per-task price. A buyer paying five cents a task for a contaminated answer is not getting a bargain. They are paying twice, once for the bad data and again to catch it. Rational customers stop buying that on their own, no mandate required, and the specialized platforms that absorbed the demand are proof the underlying market for human judgment in AI pipelines is not shrinking. It just stopped tolerating anonymity as a substitute for accountability.

The practical lesson for anyone running human review inside an AI workflow, whether that is content moderation, support escalation, or a human sign-off step in a compliance process, is that "a person looked at this" is not a fact you can assume. It is a claim you have to verify, the same way you would audit a vendor's invoice. I walk clients through exactly this kind of check when we map where a human is actually load-bearing in a process versus decorative. If you are not sure which of your review steps are real, let's talk.