On November 2, 2005, Amazon Web Services announced Amazon Mechanical Turk, a web service that let a computer program hand off work to human beings. The official announcement frames it as a missing piece of automation: “What if a computer program could ask a human being to perform a task and return the results? What if it could coordinate many human beings to perform a task?” Amazon described this as integrating “Artificial Artificial Intelligence” directly into software - the name a deliberate nod to the 18th-century chess-playing automaton that concealed a human operator.
The mechanics were simple. Employers, called requesters, posted small units of work called Human Intelligence Tasks, or HITs - things like transcribing audio, judging whether an image matched a label, or rating search results. Workers picked up HITs for a few cents each and returned the answers, while the requester’s software treated the crowd as if it were a single function call. Adoption was immediate; in the AWS News Blog, Jeff Barr wrote that “we went from standing still to 100 miles per hour in less than 48 hours,” and within days the launch was on Slashdot.
Mechanical Turk became foundational plumbing for machine learning. The same human-labeling workflow it pioneered - many low-paid workers each making a tiny judgment, with multiple votes per item for quality - is how large labeled datasets get built. ImageNet, the dataset that launched the deep-learning era, was assembled on Mechanical Turk a few years later.
Why business readers should care: Mechanical Turk made explicit a pattern that still underpins modern AI - “automated” systems are frequently backed by distributed human labor that the end user never sees. Understanding where that labor sits, what it costs, and how it is governed is part of understanding any AI product’s real supply chain.