Amazon Is Shutting Down Mechanical Turk

Jeff Bezos once described Mechanical Turk as “artificial artificial intelligence”: real humans doing the small, repetitive tasks that made early machine learning possible, dressed up as an API so it felt like calling a model. Twenty-one years later, the actual thing it was pretending to be caught up. Amazon is shutting down Mechanical Turk on September 30, 2026, along with SageMaker Ground Truth’s and Amazon Augmented AI’s Mechanical Turk workforce option. If any part of your AI pipeline routes data labeling, content moderation, or human evaluation through either service, the practical question isn’t whether to move, it’s where, and how fast.

What’s actually closing, and when

Two things are shutting down, on slightly different timelines.

Mechanical Turk itself closes entirely on September 30, 2026. No new HITs (Human Intelligence Tasks), no new workers, no new requesters, full stop.

SageMaker Ground Truth’s and Amazon Augmented AI’s Mechanical Turk workforce option goes away the same day. If you were using Ground Truth’s labeling jobs or A2I’s human review workflows with the MTurk worker type specifically, that option disappears September 30. AWS had already closed new-customer access to Ground Truth’s own managed workforce back on July 30, 2026, so existing customers on that path have effectively been on a runway for two months already. Ground Truth’s other labeling job types and workforce options aren’t part of this shutdown.

The dates that actually matter for anyone with money or work in flight:

Date	What happens
July 30, 2026	New-customer access to Ground Truth’s managed workforce closed
September 30, 2026	MTurk closes. Ground Truth and A2I lose the MTurk worker type
October 30, 2026	Deadline for requesters to approve or reject any outstanding HITs
January 28, 2027	Last date MTurk transaction history remains viewable

Why now

Amazon’s stated reason is an internal review of its programs and services, which is corporate language for “this stopped being worth running.” The more useful explanation is what changed underneath it.

MTurk launched in 2005 selling access to a large, cheap, on-demand human workforce for tasks too unstructured for software at the time: transcription, image tagging, survey completion, content review. That was genuinely valuable when the alternative was hiring temps or doing it yourself. Two things eroded that value over two decades. AI models absorbed a growing share of the tasks MTurk workers used to do, sometimes literally: research from a few years back found a meaningful share of MTurk workers were already using AI tools to complete their own tasks. And a newer generation of labeling companies, Scale AI, Mercor, and Prolific chief among them, built what MTurk never bothered to: vetted worker pools, quality tooling, task-specific workflows, and account management, instead of a bare marketplace where quality control was the requester’s problem.

MTurk didn’t lose to AI directly. It lost to a market that AI made more valuable, and to competitors who took that market more seriously than Amazon did.

What to do if you’re affected

If you’re a requester running active labeling or eval work through MTurk: start migrating now, not in September. Five weeks sounds like enough time until you’re mid-migration and realize your quality baselines don’t transfer cleanly to a new vendor’s worker pool. Scale AI, Mercor, and Prolific are the most common landing spots for AI training data and RLHF-style human evaluation work. For simpler annotation without a managed workforce, open tools like LabelMe or Roboflow Annotate can cover teams willing to manage their own reviewers.

If you’re a requester with outstanding HITs: you have until October 30 to approve or reject them. Don’t let that slip past the shutdown date assuming there’s flexibility after.

If you’re a worker: confirm your payment method and transfer schedule in the Worker Portal before September 30. Your transaction history stays visible through January 28, 2027, but you can’t do anything about a payment problem once the platform is gone.

If you built internal tooling around the MTurk API: budget the migration like any other vendor dependency you didn’t choose to end. It’s the same lesson as any other vendor lock-in: a service can be reliable for two decades and still shut down on five weeks’ notice, which is exactly why the human-in-the-loop layer of an AI pipeline deserves the same abstraction discipline as the model layer. If your eval pipeline calls a labeling vendor’s API directly from application code instead of through an interface you control, this is the week to fix that, regardless of which vendor you land on next.

For teams building AI evals that depend on human judgment as ground truth, this is also a reasonable prompt to ask a harder question: how much of that human-in-the-loop work could move to a smaller, better-vetted internal panel instead of a crowdsourced marketplace, now that the market’s assumptions about cheap anonymous labor have shifted underneath everyone at once.