Amazon confirmed Tuesday that Mechanical Turk will permanently close on September 30, 2026 — giving existing requesters and workers thirty-five days to migrate workflows, verify payment settings, and retrieve any data they need before the platform goes dark. The announcement also revealed a detail buried in the MTurk FAQ closure page that the company did not spotlight: SageMaker Ground Truth and Amazon Augmented AI will close on the same date, meaning Amazon is not merely retiring one aging product but exiting the human data-collection infrastructure market entirely.

The prior coverage of this story — Amazon’s July announcement that it would stop accepting new customers on July 30 — documented the historical arc: a platform built on humans faking machines, consumed by machines that learned to fake humans. That story has been told. What has not been told is what the next five weeks require of the researchers, enterprise annotation teams, and workers who are still on the platform today.

Amazon Is Leaving the Human-Data Market, Not Just Retiring a Product

The September 30 date is not the end of one product in an otherwise intact AWS portfolio. The official closure FAQ confirms that the Amazon Mechanical Turk Worker type will no longer be available when creating SageMaker labeling jobs or Amazon Augmented AI human review workflows, also effective September 30, 2026.

Amazon had positioned SageMaker Ground Truth as the managed successor to MTurk’s open-crowd model. Where MTurk routed tasks to an anonymous global pool, Ground Truth used an active-learning architecture: a labeling model trained incrementally on confirmed annotations, auto-labeled the examples it could classify with high confidence, and routed only low-confidence cases to human review, cutting labeling costs by up to 70% compared to fully manual annotation. Ground Truth Plus added audit trails, confidence scores, and worker-agreement metrics — the quality management layer that MTurk’s architecture never offered.

That, too, is gone after September 30. Amazon’s message to the market is that it has decided human-in-the-loop annotation infrastructure is not its business anymore. The annotation tier that mattered economically — the commodity layer of anonymous micro-tasks — has been automated. The managed expert layer is better served by third-party specialists. Amazon is exiting both.

What the September 30 Deadline Requires

The official MTurk FAQ lays out the practical logistics that affected users need to act on now:

For workers: Verify payment preferences in the MTurk Worker Portal immediately. Set up a payment method and a transfer schedule in Account Settings if you have not already done so. All completed and approved HITs will be paid out on your normal disbursement schedule. Requesters can award bonuses until October 30, 2026 — thirty days past the platform’s closure date. Your transaction history will remain accessible until January 28, 2027.

For requesters: Verify your payment information in the MTurk Requester Portal to ensure refunds can be processed. You should expect a full refund of your pre-paid balance within thirty days of closure. HIT submission closes on September 30; any HITs submitted before that date have thirty days — until October 30 — for approval or rejection. HITs not acted on within that window will be auto-approved.

Teams running active annotation pipelines should treat September 30 as the final operating day, not a soft transition point. The platform has been in maintenance mode since July, meaning no new features and minimal investment. What it offers between now and September 30 is diminishing returns on workflows that need to migrate regardless.

What MTurk Built, and Why That History Now Has a Hard Endpoint

The machine’s most consequential hour came when computer vision researcher Fei-Fei Li needed to label millions of images at a scale her team could not afford by hiring students. A chance hallway conversation at Princeton introduced her to MTurk. “He showed me the website,” Li later recalled, “and I can tell you literally that day I knew the ImageNet project was going to happen.” Her team crowdsourced the labeling of 3.2 million images across 5,247 categories through MTurk, producing the dataset published in 2009 that would power AlexNet’s breakthrough at the 2012 ImageNet challenge and launch the deep learning era that all current AI systems descend from.

MTurk later became the primary industrial channel for Reinforcement Learning from Human Feedback — the technique that made large language models coherent enough to deploy. In this human feedback fine-tuning process, workers compare pairs of AI outputs and express preferences; a reward model trained on those preferences guides the fine-tuning of the base model. MTurk workers generated those preference pairs at scale. Their judgments — which summary was clearer, which answer more helpful, which response less harmful — were encoded into the reward models shaping ChatGPT, its predecessors, and their contemporaries.

The recursive failure that ended the platform: by 2023, researchers at EPFL estimated that 33 to 46 percent of workers on an abstract summarization task were using large language models to complete their assignments and submitting the output as human judgment. The humans who were supposed to provide ground-truth signal were routing tasks through the very AI systems researchers were trying to train and evaluate. The distinction between human judgment and machine output — the platform’s entire premise — had eroded from within.

Who Is Replacing Amazon Mechanical Turk

The data-labeling industry that absorbs MTurk’s customers looks nothing like it. Analysts tracking the market describe three distinct wage tiers that have emerged as the commodity layer was automated away:

At the bottom, the bulk commodity tier ran $1–$12 an hour for repetitive labeling and simple classification — this is the tier that MTurk occupied, and it has been largely absorbed by AI auto-labeling pipelines and synthetic data generation. Above it, the mid-tier holds domain professionals in finance, medicine, and law at $20–$85 an hour. At the top, credentialed PhDs and specialists command $85–$200 an hour for frontier reasoning data that current models cannot yet replicate. MTurk was built for the bottom tier. That tier no longer needs a human marketplace.

Scale AI — which has partnered with major AI labs and, before Meta’s $14 billion stake purchase in June 2025, was the dominant managed-annotation vendor — operates a professional annotation pipeline with domain-specialized workers and AI-assisted tooling. That stake purchase triggered departures by Google, OpenAI, and others concerned about neutrality, redirecting significant annotation volume to competitors.

Mercor emerged as a primary beneficiary. The startup connects frontier AI labs with vetted experts — lawyers, doctors, engineers, PhDs — at roughly $95 an hour for specialized reasoning data, and reached $2 billion in annualized revenue as of June 2026, doubling from $760 million four months earlier. As of July 2026, Mercor was in discussions to raise approximately $500 million at $20 billion valuation — a figure that describes in one number how much capital has moved into the tier above what MTurk served. Whether that round has closed as of this writing has not been confirmed.

Prolific has attracted academic and enterprise researchers who migrated from MTurk after data quality concerns mounted. The platform offers controlled participant pools with demographic targeting and IRB-compatible workflows, though it lacks MTurk’s API depth at the low end. CloudResearch and Qualtrics panels offer additional academic-research alternatives.

For AWS-native teams, the practical successor within Amazon’s ecosystem is gone along with MTurk. Teams that need to stay in the AWS environment for compliance or integration reasons will need to evaluate managed annotation vendors with AWS Marketplace integrations, as Amazon has no announced replacement for the MTurk Worker type in SageMaker workflows after September 30.

What the Closure Means for Academic Research Built on MTurk

Thousands of published papers in psychology, behavioral economics, and natural language processing built their datasets through MTurk. The closure does not retroactively invalidate those studies, but the Veselovsky et al. 2023 research created an unresolved question about any MTurk-derived data collected after November 2022, when ChatGPT became publicly available. Any study that used MTurk annotation in 2023 or 2024 cannot determine from the data alone what fraction of labeled examples were produced by AI rather than human judgment.

A 2025 study published in Royal Society Open Science found that MTurk workers with the platform’s “master” qualification almost never missed attention checks and showed high reliability — but that non-master workers had significant quality issues. The closure removes even that option. Academic teams that have not yet migrated to Prolific or CloudResearch for IRB-compatible human-subjects research should treat the September 30 date as a hard research-design deadline.

What Is Replacing Amazon Mechanical Turk?

The migration path depends on the use case. For machine learning data labeling, Scale AI (Outlier), Mercor, Surge AI, and Labelbox offer vetted professional workforces with quality guarantees suited to high-stakes AI applications. For academic human-subjects research, Prolific and CloudResearch offer controlled participant pools with demographic targeting, though neither fully replicates MTurk’s API flexibility at the low cost end. For basic annotation tasks, auto-labeling pipelines built on models like GPT-4o and comparable systems have largely automated the work that MTurk’s commodity tier handled.

For workers who earned income from MTurk, the five-week window is a soft transition rather than a hard cutoff for your existing account — but the platform has been declining for years, and the community consensus documented in worker forums is that available tasks had already collapsed. Verify your payment transfer schedule now to ensure you receive all earned wages before the platform closes.

Frequently Asked Questions

When does Amazon Mechanical Turk permanently close, and what should existing users do first?

MTurk will permanently close on September 30, 2026. The single most important action for both workers and requesters is to verify payment settings immediately: workers should confirm a payment method and transfer schedule in the MTurk Worker Portal; requesters should verify payment information in the Requester Portal to ensure balance refunds are processed. Requesters can approve or reject HITs until October 30, 2026. Transaction history remains accessible until January 28, 2027.

What else is closing on September 30 that most coverage has missed?

The official MTurk FAQ confirms that the Amazon Mechanical Turk Worker type will no longer be available in SageMaker Ground Truth labeling jobs or Amazon Augmented AI human review workflows, also effective September 30, 2026. Amazon is not retiring one product — it is exiting the human data-collection infrastructure market in its entirety. Teams using SageMaker Ground Truth workflows that route to MTurk workers need to migrate those pipelines, not just their direct MTurk API integrations.

Why is Amazon shutting down Mechanical Turk?

Amazon has offered no public explanation beyond “careful consideration” and “an assessment.” The structural reasons are visible: the commodity annotation market that MTurk occupied has been automated by AI auto-labeling pipelines, while the higher-value expert annotation market is served by managed platforms — Scale AI, Mercor, Prolific — that MTurk’s open anonymous architecture was not designed to compete with. A 2023 EPFL study found that 33 to 46 percent of workers on some task types were submitting AI-generated responses as human judgment, which dissolved the platform’s core value proposition. Amazon did not explain whether EU Platform Work Directive compliance obligations — which apply to microtask platforms and take effect in December 2026 — factored into the timing, though reporting from The Next Web notes the closure lands three months before that deadline.

What did Amazon Mechanical Turk actually build?

The most consequential output was ImageNet — the 3.2 million-image labeled dataset that computer vision researcher Fei-Fei Li’s team built through MTurk crowdsourcing, published in 2009 and recognized in competition in 2012. ImageNet triggered the deep learning revolution. MTurk also supplied the human preference data underlying Reinforcement Learning from Human Feedback — the fine-tuning technique that aligned ChatGPT and its contemporaries with human intent. The platform that supplied the judgment to build modern AI was ultimately displaced by the AI it built.