Amazon is closing Mechanical Turk to new customers on July 30, and the AI industry it helped build has already moved on.
Amazon Web Services has quietly added Mechanical Turk to its "Services in Maintenance" list, the internal label AWS uses for products it plans to wind down. Starting July 30, 2026, no one will be able to sign up as a new requester on the platform. Existing customers keep their access, and AWS says it will keep patching security and availability, but it has confirmed there are no new features coming. Amazon hasn't said why.
You don't need a memo to read the signal. Mechanical Turk has been around since November 2005, older than Fiverr and most of the freelance marketplaces it eventually competed with, built on a premise that sounded almost like a joke at the time: pay strangers a few cents each to do tasks computers couldn't. The name itself was borrowed from an 18th-century chess-playing automaton that turned out to be secretly operated by a hidden human. The irony landed early and never really left.
For most of its life, MTurk was where researchers sourced cheap human judgment: sentiment labeling, CAPTCHA solving, survey panels, the unglamorous connective tissue of machine learning before machine learning had a name recognizable outside a lab. In 2018, Amazon repositioned the platform inside SageMaker as an annotation pipeline for AI training data, and for a few years that gave the aging service a second life. Then the thing it was feeding started eating its own tail.
A 2023 academic study found that somewhere between 33% and 46% of MTurk workers were using large language models to complete the very tasks meant to train and evaluate those models. Ask a Turker to write a product review, rate a chatbot response, or flag toxic text, and there's a real chance a chatbot wrote the answer, not the person being paid for it. That's not a labor problem you patch with better pay. It's a data integrity problem, and it's one Amazon's own AI teams would have run into every time they pulled a dataset off the platform.
So it tracks that Amazon has been building its own way around the issue. SageMaker Ground Truth, the in-house successor, gives Amazon tighter control over who does the labeling and how the work gets verified, instead of leaving quality to an open marketplace with no real gatekeeping. Killing new sign-ups to MTurk while keeping Ground Truth alive is Amazon choosing the tool it can actually audit.
MTurk's slow exit lands in the middle of a rougher stretch for the whole human-data-labeling trade. Uber dismissed two senior leaders at its AI data-labeling unit, Uber AI Solutions, on July 1, according to Bloomberg: Naga Kasu, senior director of engineering, and Pankaj Kamat, director of product, both more than a decade into their Uber tenure. The unit only launched in 2024 and acquired Segments.ai last October, and it had already lost the executive who oversaw it, Gus Fuldner, back in January. An Uber spokesperson said the business is "seeing strong momentum" and didn't explain the departures.
Scale AI, the company that turned data labeling into a venture-backed category, has its own credibility problem. Meta paid $14.3 billion last June for a 49% stake and pulled Scale co-founder Alexandr Wang in to run its superintelligence lab. Rival AI labs that had relied on Scale to build training data started asking whether a Meta-owned vendor could be trusted to label data for Meta's competitors. Scale still posted about $2 billion in revenue for 2025, its best year on record under interim chief Jason Droege, but the neutrality question hasn't gone away.
Put those three stories side by side and a pattern shows up. The open, anyone-can-log-in marketplace model that MTurk pioneered is losing out to purpose-built shops that promise quality control, provenance, and a clean chain of custody for training data. Frankly, that's the trade the industry has been quietly making for two years: less flexible, more accountable.
The people who lose the most are the ones without a Ground Truth contract or a seat at Scale's table. Academic labs, small startups, and independent researchers used MTurk because it was cheap, fast, and required no sales call. That long tail doesn't have the budget for an enterprise labeling vendor, and nobody has said who inherits them once MTurk stops taking new customers. For now, they have three weeks left to get an account before the door closes for good.
Also read: South Korea Turns an AI Chip Tax Windfall Into a New National Fund • Alibaba and Tencent Chose to Fund Kling AI Instead of Beating It • Chip stocks tumbled this week as investors doubt the AI spending payoff
