Mechanical Turk closes Sept 30 — check your ML pipeline

Amazon finally put a date on it. Mechanical Turk closes permanently on September 30, 2026 — about five weeks out — and it does not go alone. The same notice ends SageMaker Ground Truth and Amazon Augmented AI (A2I). If any part of your machine learning pipeline routes to a human for labeling or review, that step now has an expiry date, and most teams who wired it in have not looked at it in two years.

What actually happened

Amazon posted the closure notice on August 25. The wording is standard AWS sunset language: "We regularly evaluate our programs, tools and services and make adjustments based on those assessments. Following an assessment, we've made the decision to close AWS Mechanical Turk, effective September 30, 2026," per PYMNTS. The service launched in 2005 and at its peak matched more than 500,000 workers to Human Intelligence Tasks — labeling, transcription, surveys, the unglamorous human substrate under a decade of machine learning.

The operational details are in Amazon's own closure FAQ, and they matter more than the eulogy:

• HIT submission stops September 30, 2026. Unsubmitted HITs expire automatically.
• Work submitted by that date gets the normal 30-day approval window; anything not approved or rejected auto-approves.
• Bonuses can be awarded through October 30, 2026.
• Requesters "should expect a full balance refund within 30 days." Verify your payment info before then, not after.
• Transaction history stays available until January 28, 2027. That is your window to export labeling provenance, not a permanent archive.
• SageMaker Ground Truth and Amazon Augmented AI close on the same date.

This is the second half of a phased shutdown. AWS stopped accepting new Mechanical Turk customers on July 30, reported by TechCrunch at the start of July, and closed Ground Truth and A2I to new registrations the same day. The message then was that existing customers keep access. Seven weeks later, access has a deadline.

Why the Mechanical Turk closure matters for your business

Most small teams never opened the MTurk console, so the instinct is to scroll past this. Don't — check A2I first.

Amazon Augmented AI is the buried dependency. A2I is the service you wire in when a model's confidence score drops below a threshold and the record needs a human before it commits: flagged invoices, ambiguous document extractions, product images that don't classify cleanly, moderation edge cases. It was the sanctioned way to put a human gate inside an AWS pipeline. If someone on your team built that gate — or a vendor built it for you — it stops working on September 30. What happens to a record that hits the low-confidence branch after that is a question your code answers by accident right now.

Three things to do this week, in order:

Grep for it. Search your infrastructure and application code for a2i, augmented-ai, sagemaker-groundtruth, mturk, and HumanLoop. Check CloudTrail for calls to those APIs over the last 90 days. Absence of code is not absence of use — a vendor integration or a SageMaker pipeline definition can call it without the string ever appearing in your repo.

Export the provenance, not just the labels. You probably have the labeled dataset. You likely don't have a clean record of who labeled it, when, at what agreement rate, and under which instructions. That record lives in transaction history and disappears January 28, 2027. If you ever have to defend a model's training data — to a customer, an auditor, or a court — that export is the artifact.

Decide what the low-confidence branch does now. The lazy fix is to raise the threshold so fewer records need review, which quietly converts a human check into an unreviewed automated decision. That's the wrong trade and it's the one teams make under deadline pressure. Route those records to a queue your own staff works, even if the queue is a table and an internal page. A human gate you own beats a managed one that just got a shutdown notice.

The broader lesson is cheaper to learn here than elsewhere: maintenance mode is a countdown, not a plateau. When a vendor stops taking new customers and stops shipping features, the honest read is that a closure date exists internally and hasn't been published yet. Treat the "existing customers keep access" line as a timer starting, and spend the runway migrating instead of hoping.

Sources: Amazon Mechanical Turk closure FAQ, PYMNTS, TechCrunch.