Amazon Is Closing Mechanical Turk to New Customers, the Human Cloud That Quietly Trained Modern AI

TL;DR

Amazon has added Amazon Mechanical Turk to the list of AWS "services in maintenance" and set July 30, 2026 as the last day it will accept new customers. Existing requesters can keep running jobs for now, and Amazon has offered no public reason for the wind-down. The service, launched in November 2005, is where a generation of humans did the tiny piecework, tagging images, transcribing audio, flagging junk, that got fed into the neural networks now automating that exact work. It named the entire genre of hidden human labor behind "AI," and it is exiting quietly, by maintenance ticket, at pennies a task to the end.

What Amazon actually did

There was no blog post and no keynote. AWS simply slid Mechanical Turk onto its "Services in Maintenance" list this week, the same bureaucratic shelf where products go to be politely forgotten. The concrete change: as of July 30, 2026, no new customers can sign up. Accounts that already exist keep working, and AWS says it will continue to operate and support the service for them.

That "for them" is doing a lot of load-bearing work. A service closed to new customers is a service with a countdown on it, even if the clock is not printed anywhere. Amazon declined to say why, which for a 21-year-old product is its own kind of statement.

A quick history of the humans in the box

The name was always the joke. The original Mechanical Turk was an 18th-century chess "automaton" that toured Europe beating aristocrats, until it came out that a real chess master was hidden inside the cabinet working the arm. Amazon revived the name in 2005 for a marketplace where you post a small job, a "Human Intelligence Task" or HIT, and a distributed crowd of people do it for cents. Jeff Bezos called it "artificial artificial intelligence": software that, when it hits something a computer cannot do, quietly routes the problem to a human and hands the answer back as if the machine did it.

That phrase turned out to be the whole shape of the modern industry. Under most impressive "AI" of the 2010s sat a layer of people tagging, ranking, transcribing, and moderating, paid by the task, invisible in the demo. Mechanical Turk did not invent that labor so much as give it an address and a price.

The part builders should sit with: it trained your models

Here is why this is not just nostalgia. In 2009 the ImageNet dataset, the 14-million-image corpus that lit the fuse on the deep-learning boom, was labeled largely by Mechanical Turk workers drawing boxes and picking categories for pennies apiece. The 2012 result that put deep learning on the map was trained on labels that a crowd of anonymous people produced by hand. By 2018 AWS was pitching MTurk directly as the place to have "humans review and annotate data used to train neural networks," wired straight into SageMaker.

So the loop is uncomfortably tidy. The humans in the box labeled the data. The data trained the models. The models got good enough to do the labeling. It is the ladder-and-the-rung problem: the workforce assembled the staircase that the elevator now makes redundant.

Why it is winding down now

Amazon has not said, but the shape is legible. The company now steers customers toward SageMaker Ground Truth, its managed labeling product that blends automated pre-labeling with human review, plus vetted third-party labeling vendors. A raw, open, pay-a-stranger-a-penny marketplace is a worse fit for how frontier labs actually buy annotation today: curated, quality-controlled, contractually clean, and increasingly machine-assisted.

The crowd side had been eroding for a while too. Workers on forums describe accounts being cut off on short notice with little explanation, and the going rate never really recovered from its glory days. Then came the twist that reads like a short story: in 2023, researchers caught a meaningful slice of Turkers using ChatGPT to complete their text tasks. The humans hired to be more reliable than machines had started quietly subcontracting to a machine, which is either the perfect ending or the moment the whole premise ate itself.

Why a builder should care

• The "human in the loop" was never a metaphor. Behind the datasets you fine-tune on sits paid piecework by real people. MTurk closing does not remove that layer, it just consolidates it into managed vendors you cannot see as clearly.
• Cheap open crowdsourcing is ending, not the labor. If you relied on MTurk for cheap eval labeling, sentiment tagging, or human preference data, your options are now Ground Truth, a specialist vendor like Scale or Surge, or a synthetic pipeline. Budget and quality assumptions change with each.
• Data provenance is getting harder to audit. As labeling moves into closed managed services and machine pre-labeling, "who actually produced these labels, and how" gets murkier right as regulators start asking.
• Watch the synthetic-data tradeoff. Replacing human labelers with model-generated labels is cheaper and faster, and it quietly imports the base model's blind spots into everything you train on top. Free lunch, meet subtle bias.

Key Takeaways

• Amazon moved Mechanical Turk to AWS "maintenance" and stops accepting new customers on July 30, 2026; existing customers keep running for now, with no official reason given.
• Launched in 2005 and named for an 18th-century chess machine with a human hidden inside, MTurk coined "artificial artificial intelligence" and priced the hidden human labor behind AI.
• Its workers hand-labeled foundational datasets including ImageNet in 2009, and AWS later pitched it directly for annotating neural-network training data.
• Amazon now steers labeling to SageMaker Ground Truth and vetted vendors; the raw penny-per-task marketplace no longer fits how labs buy annotation.
• By 2023 some workers were using ChatGPT to do their tasks, closing the loop between the crowd and the models it helped train.

Sources: The Register, Shopifreaks, TechCrunch