Mechanical Turk's maintenance mode exposes AWS's AI gap

Why it matters

AI labs still need human judgment, expert demonstrations, and evaluation data. AWS had one of the earliest scaled workforces for that job, but its maintenance move cedes the faster-growing layer to Scale, Mercor, micro1, and other data specialists.

AWS will close Amazon Mechanical Turk to new customers on July 30th, putting the 20-year-old crowdsourcing marketplace into maintenance at the same moment AI labs are paying heavily for the human data and expert feedback that MTurk once made available at web scale.

This one lands personally. I was on the AWS Startup Team from 2014 to 2018, where I got an inside look at AWS's early AI strategy, including the OpenAI deal AWS did not land and Microsoft ultimately won. Even then, MTurk looked like an underutilized asset that could have been massive with the right product vision.

The change surfaced in a July 3rd Register report and is now posted at the top of MTurk's own homepage: "Amazon Mechanical Turk will be closed to new customers, effective July 30, 2026. Existing users will not be impacted by this change." AWS's broader June 30th service availability notice puts Mechanical Turk among a batch of SageMaker AI features moving to maintenance, alongside Ground Truth, A2I, Clarify, Debugger, GeoSpatial, Model Monitor, Role Manager, and Studio Lab.

AWS defines maintenance plainly in its general reference documentation: customers cannot onboard, existing customers can continue using the service, AWS will keep operating and supporting it, and AWS will not enhance or add functionality. That is a quiet retirement track, even if AWS has not given MTurk a full shutdown date.

The timing is the important part. MTurk launched on November 2nd, 2005, when Amazon described it as a way for software to call on humans for tasks computers could not reliably do. The original AWS announcement called the idea "Artificial Artificial Intelligence," a phrase that aged into a better description of the AI industry than Amazon could have intended. Two decades later, frontier AI systems still depend on people to label data, judge answers, create evaluations, write demonstrations, and expose model failures. The market did not disappear. Amazon's version of it aged out.

AWS did bolt MTurk onto its machine-learning stack. Its current MTurk page still pitches the service for machine-learning development, human-in-the-loop validation, content moderation, data deduplication, survey participation, and research. AWS's SageMaker Ground Truth documentation says Ground Truth can use workers from Mechanical Turk, a chosen vendor, or a private internal workforce, and separately says the Mechanical Turk workforce includes more than 500,000 independent contractors worldwide. Another AWS page says MTurk usually gives Ground Truth and A2I users the fastest turnaround because the workforce is available globally, 24 hours a day, 7 days a week.

That should have been Amazon's advantage: distribution on both sides of the marketplace, native AWS billing, an API surface developers already understood, and the credibility to sell human data workflows to researchers and enterprises already building on SageMaker. Instead, AWS is telling new customers to stop coming through the front door while the rest of the AI data market has moved toward screened experts, evaluation pipelines, workflow capture, and domain-specific training tasks.

The contrast is now easy to see. Mercor says it is building "the layer between human expertise and frontier models," with domain experts training models and enterprises using its evaluation infrastructure. In October 2025, Mercor announced a $350 million Series C led by Felicis, with Benchmark, General Catalyst, and Robinhood Ventures participating, at a $10 billion valuation. micro1 describes itself as a data lab for frontier model training, agent evaluations, RL environments, and robotics data. Outlier, operated by Scale AI, says it connects experts with AI companies to provide human feedback for language models and claims more than 100,000 experts across 50 countries.

Scale AI gives the clearest read on the asset Amazon let drift. In June 2025, Scale said Meta's investment valued Scale at more than $29 billion and sent founder Alexandr Wang to work on Meta's AI efforts. Meta's deal was a purchase of proximity to data pipelines, human evaluation capacity, and the operating knowledge behind them. Amazon had a version of that infrastructure first.

MTurk's decline did not happen only because Amazon failed to market it. The product had real quality problems. A 2023 Texas A&M paper, "Who Broke Amazon Mechanical Turk?", compared the same survey task across multiple years and found unusable data rose from 2.4% in October 2013 to 88.8% in June 2022 under the authors' cleaning method. A separate 2023 arXiv paper estimated that 33% to 46% of MTurk workers used large language models while completing a text production task. The old promise of a reliable human behind each microtask became harder to defend once workers could cheaply route the task back through AI.

Those problems made reinvention more urgent. AWS could have turned MTurk into a verified expert network, added stronger identity and skill screening, built native mobile workflows for workers, given requesters better provenance controls, and sold model-evaluation infrastructure as a first-class AWS AI product. AWS had the requester base, the cloud relationships, and the brand permission to own the data source layer for teams training and evaluating models. The June 30th availability notice shows AWS choosing maintenance instead.

Even AWS's suggested path is muddier than it looks. The MTurk homepage points machine-learning users toward SageMaker Ground Truth and Ground Truth Plus. AWS's June 30th notice, however, says Ground Truth is also moving to maintenance for new customers starting July 30th, while Ground Truth Plus reached end of support on June 30th. AWS still has a Ground Truth product page for model customization work, but the old self-serve ladder from MTurk to managed labeling is no longer a growth story.

For existing MTurk requesters and workers, the immediate impact is limited because AWS says existing users can keep using the service. For the AI market, the message is larger. Amazon built one of the original marketplaces for human computation, then watched a new generation of founder-led AI data companies turn human judgment into one of the most valuable inputs in model development. AWS did not lose the AI data market because the need for people went away. AWS lost position because the product stayed closer to a 2005 web marketplace than a 2026 AI supply chain.