Mechanical Turk Closes September 30 and Takes SageMaker Ground Truth With It. You Have 34 Days.

Amazon said on Tuesday that Mechanical Turk shuts down on September 30, 2026, twenty-one years after launch. The part that actually breaks things is buried in the same notice: SageMaker Ground Truth and Amazon Augmented AI close on the same date. That is not one aging marketplace being sunset. That is AWS walking out of the human-data-collection business and closing the door.

If you have a pipeline with a human review step in it, you have 34 days.

The signal came a month early and most people missed it

Amazon stopped accepting new MTurk customers last month. Workers on the platform read that as the end, and they were right. That is the reliable tell across every deprecation I have tracked on this blog: the product stops taking new signups well before anyone publishes a date. Google's Content API for Shopping did it. The OpenAI Assistants API did it. When a service quietly closes the front door, the back door is already scheduled.

Worth writing that down as a monitoring rule, because it is cheap: for every third-party service you depend on, check once a quarter whether a brand new customer can still sign up. If the answer is no, you have roughly a quarter to move, not a week.

What MTurk actually was, and why it died

MTurk launched in 2005. Jeff Bezos called it "artificial artificial intelligence," which was funnier before it became literally the business model of the AI industry. Businesses posted small tasks called Human Intelligence Tasks: labelling images, transcribing audio, filling out surveys, checking whether two product listings were the same thing. At its peak the platform had more than 500,000 workers.

It died from two directions at once. Models got good enough to do the easy tasks, so the volume of trivial HITs collapsed. And a generation of better-funded labelling companies, Scale AI, Mercor, Prolific among them, took the workers and the serious contracts for the hard tasks. MTurk was left with the middle, which is not a place any product survives. Amazon's own investment in it visibly tapered off years ago.

The three groups this actually hits

Academic and market researchers. MTurk has been the default subject pool for online studies for fifteen years. Prolific has been eating this segment already and is the obvious destination, with a better reputation on participant quality and worse pricing. If you have a study in the field right now, September 30 is a real deadline for data collection, not a soft one.

Anyone using Augmented AI for human-in-the-loop review. This is the group I would worry about, because A2I was wired into production workflows: low-confidence Textract or Rekognition outputs get routed to a human reviewer, the corrected result flows back. That is not a marketplace you swap, it is an integration you rebuild. Thirty-four days is thin for that.

Solo operators with a "we send the weird ones to a person" step. More of us have this than admit it. Content moderation on user submissions, receipt parsing that fails, a classification model that punts on 4% of cases. If yours was routed through MTurk because it was the cheapest way to buy five minutes of human attention, you now need either a different vendor with a real minimum spend, or you need to do it yourself.

What I'd actually do

If you are the third group, my honest recommendation is to check whether you still need the human step at all, and to check it with a measurement rather than a hunch. Take the last 200 items you routed to a human, run them through a current frontier model with a proper rubric, and compare against what the humans returned. Two years ago that comparison went badly. It may still go badly for your specific task, which is a useful thing to learn for $3 of tokens. But the reason MTurk is closing is that for a large class of tasks it now goes fine, and if yours is one of them, the migration is a prompt, not a vendor evaluation.

If the comparison goes badly, the replacement stack depends on volume. Small and irregular: Prolific for research-shaped work, or an actual freelancer on retainer, which is more reliable than a marketplace at low volume anyway. Larger and steady: Scale, Mercor, or Surge, all of which will want a contract conversation rather than an API key and a credit card.

Either way, export first. Do the boring thing this week: pull your completed HIT data, your qualification records, your worker lists, and your A2I human review outputs out of AWS and into your own storage. Deprecations rarely come with a good export tool, and the export path is always in worse shape in the last week than it is today.

Where I could be wrong about the "just use a model" advice

The obvious objection is that swapping human labellers for a model is exactly how you get a dataset that quietly encodes the model's own errors, and then train on it, and then wonder why quality flattened. That risk is real and it is not hypothetical. For anything that produces training data or ground truth, keeping a human in the loop is the point of the human in the loop, and my $3-of-tokens test is not a substitute for it.

The second objection is about labour, and it deserves to be said plainly rather than skipped. Five hundred thousand people used this platform, many of them for real income, and "the models can do it now" is a sentence with consequences for them that do not show up in anyone's migration checklist. I do not have a solution to offer. I do think the people writing about this deadline should at least name it instead of treating it as a pure infrastructure story.

Author

Lukas

@lukcombinator

Sources

• Amazon shutting down Mechanical Turk platform on Sept. 30, 2026
• Amazon service Bezos once called 'artificial artificial intelligence' is shutting down
• Amazon Mechanical Turk Will Close September 30, Shutting Down SageMaker Ground Truth Too
• Amazon is shutting down Mechanical Turk after 21 years as AI reshapes crowdsourced work