Amazon Mechanical Turk, the crowd-work platform that supplied human labels to a generation of machine learning research, is in AWS's maintenance stage: closed to new customers, no new features planned, existing customers unaffected. That is what AWS's own documentation says. What AWS does not say anywhere retrievable is that the service shuts down on September 30, 2026 - a date circulating widely in coverage that no AWS page supports.
Key facts
- AWS documentation, verbatim: Mechanical Turk "is no longer open to new customers," existing customers "can continue to use the service as normal," and AWS does "not plan to introduce new features."
- No AWS page states a full shutdown date. The service's front door at mturk.com remains live.
- AWS gives no stated reason for the change in any verifiable document.
- Primary sources: AWS SageMaker AI workforce documentation and AWS Services in Maintenance.
The maintenance stage is a routine AWS product lifecycle designation. Its policy, confirmed on the Services in Maintenance page, is exactly the three things quoted above: no new onboarding, existing customers continue, no new functionality. AWS moves services into it regularly. A June 30, 2026 service availability announcement describes a batch of services moving to maintenance with a July 30 cutoff for new customers, including "multiple Amazon SageMaker AI features" - though Mechanical Turk is not named in that summary.
So the verified story is duller than the one being told: a fifteen-year-old service was put out to pasture without explanation. The story people want is that AI killed it. That story is an inference, and the research it leans on says something more specific and more interesting than the headline version.
The number everyone cites is "46% of Mechanical Turk work is done by AI." It comes from Artificial Artificial Artificial Intelligence: Crowd Workers Widely Use Large Language Models for Text Production Tasks, from EPFL's data science lab, with code published at epfl-dlab/GPTurk.
What the researchers actually did: they reran one specific task on Mechanical Turk - summarizing medical research abstracts - while logging keystrokes and training a classifier to distinguish machine-written from human-written text. Across 46 submissions, they estimated that between 33% and 46% were produced with language model assistance: 21 of 46 at the upper bound, 15 of 46 on the conservative estimate. A large share of submissions involved pasting rather than typing.
That is one task, with 46 submissions, deliberately chosen because summarizing text is close to the easiest thing to hand to a chatbot. The authors say so themselves and caution against generalizing to tasks that are less language-model-friendly. It is not a platform-wide rate, and it never was. A follow-up study found roughly 30% uncued use and, more usefully, that instructing workers not to use language models combined with copy-paste friction cut usage roughly in half.
Here is why this matters far beyond one platform's product lifecycle.
Crowd platforms were the mechanism by which "human judgment" entered machine learning. Human preference labels for alignment training, human relevance judgments for search evaluation, human annotations for benchmarks - a great deal of it was purchased in small increments from workers paid roughly a dollar a task. The entire epistemic value of that data rests on it being human.
If a meaningful share of it is a language model's output passed through a human's clipboard, then models are being trained and graded on other models' text while everyone involved believes otherwise. That is a contamination problem that looks exactly like clean data from the buyer's side. It also compounds: a benchmark validated on contaminated labels certifies models that agree with the contaminating model. Related: synthetic data and how AI is benchmarked.
The economics behind it are not mysterious. A worker paid about a dollar for a hundred-word summary who can produce it in fifteen seconds instead of ten minutes has an obvious incentive, and no meaningful enforcement stands against it. The follow-up study's finding that simple friction halves the rate is the most actionable result in this whole area, and it is barely cited compared to the scary number.
Two honest caveats. First, the causal link between AI contamination and Amazon's decision is unsupported - AWS states no reason, and a service closed to new customers after fifteen years is a common enough outcome without any AI explanation. Second, the specific figures from both EPFL papers could not be re-verified against their full texts in this pass and are carried from secondary summaries; the papers themselves are linked above and worth reading directly before quoting the numbers.
The durable takeaway is not the shutdown rumor. It is that if you buy human-labeled data, the question you should be asking is not whether the platform will still exist next year. It is whether the labels were ever human.
Originally published on Ground Truth, where every claim is checked against the primary source.
