On July 30, 2026, Amazon quietly stopped letting anyone sign up as a new Mechanical Turk requester. Four weeks later, the company confirmed what that move had already telegraphed: the platform will close for good on September 30, 2026. Twenty one years after Jeff Bezos launched a marketplace for what he called “artificial artificial intelligence,” the real thing has made the joke unnecessary.
The official notice was short and procedural, the kind of language companies use when they want a decision to sound routine rather than symbolic. “We regularly evaluate our programs, tools and services and make adjustments based on those assessments,” Amazon told users. “Following an assessment, we’ve made the decision to close AWS Mechanical Turk, effective September 30, 2026.” Two sibling services built on the same idea, SageMaker Ground Truth and Amazon Augmented AI, are closing on the same date, which means Amazon is not trimming one product. It is walking away from human-labeled data infrastructure entirely.
That is a strange thing for the company that arguably invented the modern AI training pipeline to do in the middle of the biggest AI buildout in history. Understanding why requires going back to what Mechanical Turk actually was, and to a chess robot that has nothing to do with software.
A Chess Robot Gave the Whole Idea Its Name
In the late 1700s, a Hungarian inventor built a mechanical chess player called the Turk. It toured Europe beating statesmen and nobles, and it looked, convincingly, like an automaton. It was not. A skilled human chess player was folded inside the cabinet the entire time, working the arms by hand while the audience marveled at a machine that supposedly thought for itself.
When Amazon needed a name for a platform that paid people small amounts to do tasks computers could not yet handle, transcribing audio, tagging images, sorting survey answers, someone on the team reached for that story. Bezos described the resulting product as “artificial artificial intelligence”: a system that looked automated from the outside while running on invisible human labor underneath. It was, by his own framing, a joke with real economics attached.
The joke worked because it was accurate. Mechanical Turk launched in 2005 as a bare bones marketplace of “Human Intelligence Tasks” that typically paid a few cents apiece. At its peak it connected more than half a million workers to companies that needed data labeled, content moderated, or surveys answered at a scale no in house team could match. Academic researchers adopted it too, turning MTurk into one of the most widely used subject pools in the history of social science, with thousands of peer reviewed studies built on data collected a few cents at a time.
The Shutdown, By the Numbers
Amazon has not framed this as a failure so much as an acknowledgment that the world MTurk was built for no longer exists. Here is the timeline of how the platform got from “declining slowly” to “closing entirely” in the space of two months.
| Date | Event |
|---|---|
| 2005 | Mechanical Turk launches. Bezos calls it “artificial artificial intelligence.” |
| 2011 to 2018 | Platform reaches its peak, serving 500,000-plus workers and becoming a standard tool in academic research. |
| 2023 | Swiss researchers estimate up to 46% of MTurk workers are using AI models to complete their tasks. |
| July 30, 2026 | Amazon closes MTurk to new requester accounts. |
| Sept 30, 2026 | Mechanical Turk, SageMaker Ground Truth and Amazon Augmented AI all close permanently. |
Quick take
- Mechanical Turk closes permanently on September 30, 2026, after 21 years.
- SageMaker Ground Truth and Amazon Augmented AI shut down the same day.
- Amazon says the call followed “an assessment,” with no further detail offered.
- New requester signups stopped a month earlier, on July 30.
- Existing studies and requester accounts keep running until the September deadline.
Five weeks of notice is not a lot when a platform’s API sits embedded in academic data pipelines, enterprise QA workflows and startup MVPs that never expected the underlying service to disappear. Every one of those dependencies now breaks at the same moment.
The Workers Were Already Using AI to Do the Work
The most telling detail in this whole story is not the shutdown itself. It is the 2023 finding that up to 46% of Mechanical Turk workers were already quietly using AI models to finish the tasks they were paid to do by hand. People hired to feed training data into machine learning systems had started feeding machine learning output right back into the pipeline, undetected, for a fraction of what the labor was supposedly worth.
A platform built to supply the human judgment that machines couldn’t fake was, within two decades, quietly staffed by the machines it was supposed to be training. The automaton had climbed back inside its own cabinet.
That is not a footnote. It is close to the actual reason Mechanical Turk stopped being useful. If a meaningful share of “human” responses were AI generated, the data quality guarantee that made MTurk valuable to researchers and requesters in the first place had already quietly eroded years before Amazon made the closure official.
What Replaces a Marketplace That Helped Train an Entire Industry
Mechanical Turk is not disappearing into a vacuum. A new tier of data labeling companies has already taken the work that matters most to frontier AI labs, and it looks nothing like a marketplace of anonymous workers doing piecework for pennies. Over 75% of AI training data and reinforcement learning environment revenue now flows to just four vendors, and the shift in who gets hired is as significant as the shift in who pays.
| Company | Reported 2026 Valuation | What It Actually Sells |
|---|---|---|
| Scale AI | ~$29 billion, after Meta’s $14.3B stake | Data labeling and RLHF pipelines for frontier labs |
| Surge AI | $15 billion to $25 billion, first outside raise in 2025 | Expert-grade RLHF labeling, bootstrapped and profitable since 2021 |
| Mercor | $10 billion (Oct 2025), talks reportedly near $20 billion | Matches lawyers, doctors and professors to AI labs for expert annotation |
| Prolific | Not disclosed | Verified human panels for model evaluation and red-teaming |
The pattern is unmistakable. Mechanical Turk paid anyone with an internet connection a few cents to label an image. Its replacements pay mathematicians to annotate proofs, lawyers to mark up contracts, and working professors to grade essays, because the frontier labs training today’s models no longer need volume labeling so much as they need expert judgment that a general crowd cannot reliably fake. The work did not vanish. It moved upmarket, and it moved fast: Scale AI’s valuation alone roughly jumped from $2 billion in February 2025 to somewhere near $29 billion eighteen months later.
The same shift is visible in how open model releases are shaping who needs this kind of labeled data at all. When Meta quietly put a 30 billion parameter open-weight model on Hugging Face this month, it handed smaller teams a credible alternative to renting a frontier API, but building anything competitive on top of an open model still means sourcing exactly the kind of high-quality human feedback that Scale, Surge and Mercor now specialize in. Closing the low end of the market while the high end explodes is not a contradiction. It is the same industry sorting itself by what actually moves model quality anymore.
The Researchers Who Never Saw This Coming
Outside the AI industry, the group most exposed by this shutdown is academic. Mechanical Turk became embedded in psychology, economics and political science research precisely because it was cheap, fast and, for most of its life, reliable enough to trust. Losing it with five weeks of notice forces labs mid-study to either scramble for a replacement panel or abandon data collection entirely, and it arrives at an uncomfortable moment for anyone studying how much of the internet’s content is even human made anymore. Independent researchers already estimate that roughly one in three web pages created since ChatGPT launched was not written by a person at all, a statistic that makes the 46% AI-assisted MTurk finding look less like an isolated scandal and more like an early symptom of the same trend spreading through every corner of the internet that depends on distinguishing human output from machine output.
A Bigger Question Sitting Underneath This
There is a harder question tucked inside this story than “what replaces Mechanical Turk.” It is whether the enormous capital now flowing into expert-grade training data, tens of billions of dollars across just four companies, is actually translating into the productivity gains that justify it. A recent National Bureau of Economic Research survey of executives after three years of enterprise AI adoption found that more than 90% reported no measurable effect on their own firm’s employment, and 89% reported none on productivity either. That lines up uncomfortably well with a separate trend already playing out in the consumer market, where ChatGPT crossed a billion weekly users this month even as its share of the AI assistant market kept shrinking. Usage keeps climbing. Measurable payoff, for a lot of the companies actually paying for it, keeps lagging behind.
Where AI training-data money is concentrated, 2026
Share of AI training-data and reinforcement-learning-environment revenue, by vendor category.
What This Means If You Build or Study With AI
For teams and researchers affected by the shutdown
- If your pipeline calls the MTurk API, SageMaker Ground Truth or Augmented AI, you have until September 30 to migrate. Export your data and requester history before then.
- Academic labs relying on MTurk panels should budget for a replacement subject pool now. Prolific and CloudResearch are the most commonly cited direct substitutes for study recruitment.
- If you are hiring for training data at the frontier level, expect to pay for verified expertise, not crowd volume. The vendors winning budget in 2026 sell judgment, not clicks.
- Treat any AI-assisted human data source with more skepticism than you did two years ago. The 46% figure is a reminder that “human labeled” and “human verified” are no longer the same guarantee.
The Bottom Line
Mechanical Turk did not fail because nobody needed cheap human labor anymore. It failed because the line between human and machine labor inside its own workforce had already blurred past the point of being useful, years before Amazon made the call official. The platform that gave the AI industry its founding joke, a marketplace of artificial artificial intelligence, is closing in a world where the artificial part finally caught up to the joke. What comes next is not less human involvement in training AI. It is a smaller number of much more expensive humans, and a much bigger bill for the companies that still need them.
