Amazon Mechanical Turk will shut down on September 30, 2026.

A gig platform charging just a few cents per task, once powering the entire deep learning era.

Now, it is closing.

Amazon has just announced that its crowdsourcing platform, Mechanical Turk, will permanently shut down on September 30 of this year.

At its peak, over 500,000 people were online simultaneously on the platform, performing tasks such as viewing images, categorizing them, and labeling data—completing every task manually, as computers at the time were still unable to handle these tasks.

Later, this massive human-powered computing force directly contributed to supporting the ImageNet dataset—

Without MTurk, the manual annotation of tens of millions of images on ImageNet would have been nearly impossible.

Without ImageNet, Li Fei-Fei might not have become famous overnight.

Hinton and his students did not instantly achieve legendary status with a single battle, and deep learning may not have suddenly revived in 2012.

After 21 years, real AI will get better at doing these assignments.

And those who once tried to make AI seem smarter, along with this platform, have received a notice of discontinuation.

Mechanical Turk (MTurk) was launched in 2005.

Actually, at first, Amazon never intended to use this to rewrite the history of AI.

At the time, MTurk addressed a particularly practical and ironic problem: some tasks are extremely difficult for computers but absurdly simple for humans...

So Amazon came up with a very straightforward solution: if the program can't handle it, break the problem into pieces and hand it off to real people.

These individual tasks, once separated, are called HITs, or Human Intelligence Tasks.

For example, if a company has 10,000 images to classify, the MTurk approach is to split them into 10,000 individual tasks and post them all on the platform at once.

100 people can do it together, 1,000 people can too, and if tens of thousands join at once, it will be even faster.

In other words, MTurk doesn’t make one person work faster—it turns a task that could only be done sequentially into one that can be performed in parallel globally.

The real human-powered data flywheel...

And the name "Mechanical Turk" itself perfectly matches this "gameplay."

In the 18th century, Europe had a famous mechanical Turk, which appeared to be a machine capable of playing chess on its own—even competing against Napoleon.

Later, everyone discovered that a real human player had been hiding inside the machine!?! (I really messed up...)

Amazon has brought this concept to the internet with near-perfect precision: the program appears to be handling tasks automatically, but in reality, it's real people behind the screen doing the work.

Bezos once gave it a phrase that proved eerily prophetic:

Artificial artificial intelligence.

Once this idea took off, MTurk quickly grew from Amazon's internal tool into a global crowdsourcing marketplace.

By 2011, MTurk had gathered over 500,000 Workers across 190 countries.

As the number of people increased, the volume of available tasks also grew rapidly—image filtering, speech transcription, text classification, data cleaning, sentiment analysis, and content moderation.

Anything that’s hard for machines but easy for humans to judge in seconds can be thrown here.

That throw was 21 years ago.

Back around 2006.

Shortly after arriving at Princeton to teach, Li Fei-Fei decided to do something that was quite counter to the trend at the time—

Build a large enough visual encyclopedia for machines.

At that time, mainstream computer vision datasets typically contained only thousands to tens of thousands of images—too little data for models to truly understand the vast variety of objects in the real world; instead, they easily became adept at memorizing patterns from the training set.

So Li Fei-Fei’s team set out to do something considered wildly ambitious at the time: create a “sufficiently large” visual world for machines.

The ImageNet dataset was born from this.

ImageNet uses WordNet as its framework, breaking down the real world into individual concepts and populating each concept with hundreds or thousands of real images.

Dogs, cars, apples, chairs, birds, fish… include anything that can be defined.

But soon, a more practical issue came crashing down—

Pictures are easy to find—Google, Yahoo, and Flickr all work well—but the real problem is that the images you get back aren’t necessarily accurate, and are often completely unrelated!

After the search engine retrieves the images, someone must still manually verify each one.

To give an exaggerated figure, the number of candidate images ImageNet later had to process exceeded 160 million!!! This turned the human resource issue into an insurmountable bottleneck...

Just as ImageNet was being slowed down by the massive manual image sorting, Li Fei-Fei’s team turned their attention to the newly launched Amazon Mechanical Turk.

The ImageNet team began breaking down the massive image screening task into microtasks and distributing them to workers on the MTurk platform worldwide.

A grand AI infrastructure has been reduced to countless ordinary "mouse clicks."

The 2009 ImageNet paper described that the dataset already included 5,247 concepts and 3.2 million curated images.

The ImageNet team later reflected on the project and reported that 49,000 MTurk workers from 167 countries participated.

In other words, the massive image-filtering effort behind ImageNet was truly powered by a global, temporary workforce of annotators.

That's why without MTurk, ImageNet might never have been possible...

After ImageNet was developed, the story truly began to accelerate.

In 2010, the ImageNet Challenge began, prompting models worldwide to be trained on the same dataset and compete on the same leaderboard.

Two years later, the duo that would later be repeatedly chronicled in AI history made their appearance—

Geoffrey Hinton from the University of Toronto, along with his two students, Alex Krizhevsky and Ilya Sutskever.

Three people entered a deep convolutional neural network into the ImageNet Large Scale Visual Recognition Challenge.

Directly reduced the Top-5 error rate on the 2012 ImageNet Challenge to about 15%, leaving traditional methods far behind.

This also allowed the entire computer vision community to see for the first time how combining massive amounts of data, GPU computing power, and deep neural networks could instantly elevate performance to an entirely new level.

From left to right: Ilya Sutskever, Geoffrey Hinton, and Alex Krizhevsky

From then on, Alex Krizhevsky's name became forever linked to AlexNet.

Ilya later went to Google Brain, co-founded OpenAI as its Chief Scientist, and after leaving, founded SSI.

Later, Mr. Hinton won the Turing Award and became one of the universally recognized fathers of deep learning.

Li Fei-Fei returned to Stanford to teach, co-lead HAI, and cemented her place in the history of modern computer vision through ImageNet.

Later, VGG arrived, followed by GoogLeNet and then ResNet.

The ImageNet leaderboard has been continuously broken year after year; deep learning has surged from vision into speech, then into natural language processing, ultimately leading us to today’s era of large models.

This world is truly amazing—

Li Fei-Fei’s team used MTurk to create ImageNet; ImageNet brought Li Fei-Fei widespread recognition; Hinton and his students rose to fame through ImageNet; deep learning leveraged this victory to enter a period of revival.

Nearly 50,000 ordinary people from 167 countries sat at their computers and manually selected billions of candidate images to create ImageNet.

Precisely for this reason, Mechanical Turk's shutdown today feels like the closing of an era.

Amazon's official statement was restrained, stating only that it continuously evaluates its products and services, leading to the decision to discontinue MTurk.

CNBC previously cited Krista Pawloski of the worker rights organization Turkopticon, stating that MTurk has been in decline over the years, with Amazon investing less and new data labeling platforms continuously drawing away workers and clients.

But the more direct reason, put simply, is that AI has gotten better and better at doing the very tasks that made it most valuable back then...

In 2005, getting a machine to determine whether a picture contained a dog was a serious challenge, but now it’s become a basic task that multimodal models can handle effortlessly.

Meanwhile, the AI industry's demand for "human" talent has also changed.

Advanced models require professionals such as programmers, doctors, and lawyers to review code, evaluate responses, test reasoning, and assess security.

As a result, new platforms such as Scale AI, Mercor, and Prolific have emerged to screen and manage experts, providing more specialized training and evaluation data.

In contrast, crowdsourcing models like MTurk, where "anyone can take on tasks for a few cents per job," are increasingly falling behind the times...

Even more astonishing, MTurk workers themselves are now using AI!!!

A 2023 study by Swiss researchers found that among surveyed MTurk workers, the proportion using AI models to complete text tasks reached as high as 46%.

Customers originally paid for human judgment, but some workers may take on the orders and then pass the questions to models like ChatGPT for answers.

This is a bit of black humor...

The fate of MTurk is, in fact, a microcosm of AI over the past 20 years.

Early humans manually clicked through images one by one, painstakingly building ImageNet and opening the door to the era of deep learning.

Later, the AI grew along with this data, eventually learning to see images, hear sounds, and write text on its own.

Twenty-one years later, those who once pretended to be smart for AI have stepped away.

The “Mechanical Turk,” which concealed a human chess player, can finally close its cabinet door.

Reference link:

[1] https://www.cnbc.com/2026/08/25/amazon-service-that-jeff-bezos-called-artificial-ai-is-shutting-down.html

This article is from the WeChat public account "Quantum Bit," authored by Meng Yao.