Amazon Is Closing MTurk. Its Humans Were Already Using AI.

TL;DR

Amazon Mechanical Turk will permanently close on 30 September 2026, after 21 years. AWS gave no reason beyond "following an assessment." The same date removes the Mechanical Turk worker type from SageMaker Ground Truth and Amazon Augmented AI, both of which are already closed to new customers. If your labeling or human-review pipeline still points at the public crowd ARN, you have about a month to move it.

The dates that matter

Amazon's closure FAQ is short and unusually specific, so here is the whole thing without the corporate padding:

• 30 September 2026: HIT submission closes. Any unsubmitted HIT expires automatically.
• 30 October 2026: last day requesters can award bonuses. Requesters keep the standard 30-day window to approve or reject HITs completed up to that date.
• No action taken: HITs auto-approve. Requester balances refund in full within 30 days. Worker payments follow the existing schedule.

The AWS documentation carries the same notice, plus the practical detail: arn:aws:sagemaker:{region}:394669845002:workteam/public-crowd/default is the workteam ARN that stops resolving. If that string appears anywhere in your CreateLabelingJob or CreateFlowDefinition calls, that job dies on 1 October. Your alternatives are a private workforce, a vendor workforce from AWS Marketplace, or an outside platform.

Note what AWS did not say. Ground Truth and A2I both already carry the note "no longer open to new customers" and "we do not plan to introduce new features." Amazon is not just retiring a labor marketplace. It is walking away from human-in-the-loop data infrastructure as a product line, right at the moment human data became the scarcest input in frontier training.

The workforce was never the size the docs claimed

AWS's own Ground Truth page advertises "the Amazon Mechanical Turk workforce of over 500,000 independent contractors worldwide." That number has been sitting in the documentation for years. It is a registration count, not a workforce.

When Difallah, Filatova, and Ipeirotis ran a 28-month survey collecting more than 85,000 responses from 40,000 unique participants, they estimated over 100,000 workers available on the platform, and over 2,000 active at any given time. Three-quarters were in the United States, 16% in India.

Two thousand people, awake at various hours across two continents, were the actual substrate under a decade of academic behavioral research, content moderation pipelines, and computer-vision training sets. That is roughly one mid-sized office building holding up a meaningful slice of the empirical social sciences.

Two dollars an hour

The economics were never good. Hara and colleagues built a browser plugin that recorded real task timings and published a data-driven analysis of worker earnings. Their finding: workers earned a median hourly wage of roughly $2, and only 4% cleared the US federal minimum of $7.25.

The interesting number is the third one. The average requester pays more than $11 an hour. The money leaves the requester's account at a plausible rate and arrives at roughly a sixth of it, because the worker eats the unpaid time: hunting for tasks worth taking, loading HITs that turn out to be traps, and absorbing rejections with no appeal.

The part that actually killed it

In 2023, Veselovsky, Horta Ribeiro, and West at EPFL ran a text summarization task on MTurk and then checked whether the humans had written the summaries themselves. They estimated that 33% to 46% of crowd workers used LLMs to complete it. The paper's title is the best joke anyone has made about this industry: "Artificial Artificial Artificial Intelligence."

Sit with the loop for a second. Bezos launched MTurk in 2005 and pitched it as "artificial artificial intelligence": a service where a computer requests work and a human does it, inverting the normal arrangement. Eighteen years later, roughly two in five of those humans were pasting the task into a language model and returning the output. The service that sold human labor disguised as software was quietly selling software disguised as human labor disguised as software.

This is not a moral failure of low-paid workers. It is what a $2/hour piece rate rationally produces. But it is fatal to the product, because the entire value of MTurk to an ML pipeline was that the label came from a person. Once you cannot tell, you are paying a premium for laundered model output, and you would rather just call the model.

The market did not shrink, it split

The easy read is that AI ate the data-labeling job. The evidence says something more specific: the cheap end evaporated and the expert end got expensive. Bounding boxes, sentiment tags, and transcription are now cheaper and more consistent from a model than from a stranger paid in cents. What is not cheaper is a working radiologist, a securities lawyer, or a senior Rust engineer sitting down to produce reasoning traces a model cannot fake.

That work went to Scale AI, Mercor, and Prolific, platforms built on identity verification, credential checks, and hourly rates that are two orders of magnitude above the MTurk median. MTurk's design, anonymous workers doing atomized tasks at piece rates with no reputation that travels, is precisely the wrong shape for that market.

Krista Pawloski of Turkopticon, the worker advocacy group that spent years documenting MTurk's rejection and pay practices, told CNBC the service had already been in decline. Amazon is closing something that mostly stopped mattering, which is a quieter ending than the platform deserved.

What to do this week

• Grep for the ARN. Search your infrastructure for public-crowd/default and account 394669845002. Anything that matches breaks on 1 October.
• Audit anything using Ground Truth or A2I at all. Both are closed to new customers and frozen for features. This is a service on the glide path, not a stable dependency.
• Re-examine old MTurk-sourced datasets. If you collected free-text labels after roughly late 2022, a meaningful fraction may be model output. That matters for eval sets more than training sets, because a benchmark contaminated with model text quietly grades models on their own reflection.

Key Takeaways

• Mechanical Turk permanently closes 30 September 2026; bonuses and HIT approvals run to 30 October, and requester balances refund within 30 days.
• The MTurk worker type disappears from SageMaker Ground Truth and Augmented AI on the same date, and both services are already closed to new customers with no new features planned.
• The advertised "over 500,000" workforce was a registration count. Independent measurement found over 100,000 available workers and over 2,000 active at any moment.
• Median worker earnings were about $2/hour with only 4% above $7.25, while the average requester paid over $11/hour. The difference is unpaid task search and rejections.
• An EPFL study estimated 33-46% of crowd workers used LLMs on a text task in 2023, which destroys the one thing the platform sold: a label that came from a person.
• Data labeling did not die, it bifurcated. Check any free-text dataset you collected from anonymous crowd workers after 2022 before you trust it as an eval.

Sources: Amazon Mechanical Turk closure FAQ, AWS SageMaker docs: Using the Amazon Mechanical Turk Workforce, AWS SageMaker Ground Truth docs, Amazon Augmented AI docs, Hara et al., "A Data-Driven Analysis of Workers' Earnings on Amazon Mechanical Turk" (CHI 2018), Veselovsky et al., "Artificial Artificial Artificial Intelligence" (2023), Difallah et al., "Demographics and Dynamics of Mechanical Turk Workers" (WSDM 2018), CNBC, The Next Web