Amazon has set an expiration date. On July 30, 2026, Mechanical Turk will stop accepting new customers. Existing accounts will keep running, and Amazon says it will maintain the platform's security and uptime. But the company has also made clear that no new features are coming. For a service that helped define an era of internet labor, that is as good as a sunset notice.
The Original Gig Work Marketplace
Mechanical Turk launched in 2005, well before apps like Uber or TaskRabbit turned piecemeal labor into a household concept. Amazon built it as a marketplace for "human intelligence tasks" — small jobs that software could not yet handle. Workers, mostly anonymous and scattered across the globe, logged in to transcribe audio snippets, verify storefront addresses, moderate content, and tag images. Need someone to solve a CAPTCHA, rank search results, or judge whether a product review sounded positive or angry? MTurk was the place.
Prices were often measured in cents. A batch of a thousand image classifications might pay a penny apiece. For researchers, marketers, and bootstrapped startups, this was revolutionary. Suddenly you could run a survey with hundreds of participants in an afternoon, or label training data for a computer vision model without hiring a full annotation staff. The platform became infrastructure. Social scientists ran experiments on it. Tech companies tested interfaces. AI teams built datasets.
Around 2018, Amazon tried to modernize that role. It repositioned Mechanical Turk as a data annotation engine for neural networks, tying it more closely to Amazon SageMaker AI. The idea was to feed the machine learning pipeline: humans would clean, label, and verify the raw material that algorithms needed to learn. MTurk was no longer just odd jobs. It was supposed to be the human backbone for the AI boom.
That boom kept growing. But Mechanical Turk did not grow with it.
When the Human Layer Starts Using AI
The contradiction arrived quietly, then all at once. A 2023 analysis found that between 33% and 46% of MTurk workers were using Large Language Models to complete the tasks they were paid to do by hand. Researchers paid for human judgment. They got ChatGPT instead.
This is not just a story about workers cutting corners. When an annotator uses an LLM to label sentiment in customer reviews, or to classify medical text, or to write dialogue for a chatbot training set, the data ceases to be human-grounded truth. It becomes a photocopy of a machine's best guess. Feed that photocopy back into another model as ground truth, and the loop tightens. Signal degrades into noise. The "human in the loop" becomes a rubber stamp for software that was already running in the background.
For the AI industry, this is a practical crisis. Teams spend thousands of dollars on MTurk contracts believing they are buying diverse, organic human perspective. If a large minority of that labor is automated, the resulting dataset inherits the biases, blind spots, and hallucinations of the LLM doing the actual work. Quality control grows harder. Researchers have to design increasingly complex attention checks and trap questions just to prove a human is really there. That defensive posture raises costs and slows projects down