![Mechanical Turk: The Machine That Needed Humans](https://cdn.slatesource.com/8/9/8/89839ed9-942c-494c-b111-9b8fddc7c206.webp)

# Mechanical Turk: The Machine That Needed Humans

- [Made in Slatesource](https://slatesource.com/@steph/mechanical-turk-the-machine-that-needed-humans)
- By [Steph](https://slatesource.com/@steph)
- Created on Aug 26, 2026

## Amazon's Human Workforce, Fired by the Machine It Was Built to Train

On September 30, 2026, Amazon will shut down Mechanical Turk, the crowdsourcing platform it launched in 2005. At its peak the service employed more than 500,000 workers across 190 countries. Their job was to answer the questions computers could not: label an image, transcribe a recording, decide whether a sentence was rude. Amazon founder Jeff Bezos called it "artificial artificial intelligence," which turned out to be the most precise description of what was happening. Humans pretended to be machines so that real machines could, one day, do the same jobs. That day arrived. Amazon gave existing customers 35 days notice. New customers had already been cut off on July 30.

![](https://cdn.slatesource.com/8/9/8/89839ed9-942c-494c-b111-9b8fddc7c206.webp)

## What People Actually Did, and What It Paid

A Human Intelligence Task, or HIT in MTurk's vocabulary, was the atomic unit of the platform: a micro-job published by a "Requester" and claimed by a "Worker" for a fee that typically ran from one cent to a few dollars. Insurance companies used MTurk to review claims. Academics ran surveys that would have cost ten times as much in a lab. AI labs used it to annotate millions of images, label training data for speech recognition, and evaluate which of two model outputs was more helpful, a process now called RLHF. The tasks were trivial; the aggregate was not. Amazon's own documentation put the workforce at over 500,000 people working around the clock across 190 countries.

## The People Behind the Tasks

The workers who depended on MTurk full-time were spread across South Asia, Latin America, and the United States. Krista Pawloski, a full-time Turker and organizer at the worker advocacy group Turkopticon, started on the platform in 2008 during maternity leave and moved to it full-time in 2012 after losing her job. She described the work as "meaningful income with flexible hours," comparable to any other gig work but fully remote. Amazon has recommended workers migrate to a third-party solution, without naming one.

[Amazon Mechanical Turk official shutdown notice, September 30, 2026](https://www.mturk.com/?utm_source=slatesource)

## The 46 Percent Problem That Made the Whole Thing Absurd

A 2023 study by researchers in Switzerland delivered a finding that made MTurk's continued existence philosophically untenable. In a controlled experiment, up to 46 percent of MTurk workers appeared to be using AI models to complete the tasks they had been hired to complete as humans. The platform was created because early AI could not do what humans could. By 2023, workers were outsourcing their responses to the very AI systems their earlier work had helped to train. The closed loop was complete. A human labeled an image in 2010. A model learned from it in 2018. A worker fed that model's output back into a new labeling job in 2023. The data quality question this raised was obvious. If nearly half the responses marked "human" were generated by AI, the scientific research built on them was compromised. Requesters began leaving for platforms that could verify human input more reliably. Amazon did not fix the problem. It scheduled a shutdown instead.

> They're concerned now.

## Who Fills the Gap: Fewer Workers, Paid More

The platforms that replaced MTurk do not look like MTurk. Scale AI, Mercor, and Prolific do not want anonymous pennies-per-task crowds. They want domain-specific experts: radiologists to review medical imaging labels, software engineers to evaluate code outputs, lawyers to assess legal reasoning. The pay is higher per hour and the work requires credentials that most MTurk workers never needed. Scale AI was valued at $14 billion in 2024 and counts the U.S. Department of Defense among its clients. The crowd economy that MTurk pioneered is not dead. It has been stratified. The lowest tier, the one MTurk occupied, has been automated away. What remains is a smaller, more expensive, more skilled tier that AI has not yet been able to replace, though the companies building that tier are explicit that replacing it is the long-term plan. SageMaker Ground Truth and Amazon Augmented AI, the AWS services that plugged into MTurk, will also cease to accept new work on September 30.

[Swiss study, 2023: up to 46 percent of MTurk workers using AI to complete tasks](https://arxiv.org/abs/2306.07899?utm_source=slatesource)