Mechanical Turk: human work ordered through an API
On the evening of 2 November 2005 Amazon opened beta testing of a marketplace where a requester posts a small task through a programmatic call and a person completes it. Developer documentation, a WSDL file and a Requester Tool Kit were already in place at launch.
Why it matters
Until then human work inside a program had to be organised apart from the program itself. Here it became one more call to a service: a pipeline could order a thousand image captions the way it ordered computation. The next fifteen years of labelled data rest on that.
A task is called a HIT, a Human Intelligence Task. The marketplace as described in the 2010 paper: payment for a single task starts at 1.2 cents and runs up to a few dollars for something longer, such as transcribing an audio clip, and a worker can preview a task before accepting it. The record claims nothing about how many workers were on the platform on the launch day - Amazon did not say. The nearest measured figure from a primary source is more than 400,000 registered workers and 50,000 to 100,000 tasks available at any one time, as the company reported it around 2010. The idea was described as "artificial artificial intelligence". The phrase is not an outsider's: Jeff Barr and Luis Felipe Cabrera of Amazon used it in a 2006 article in ACM Queue, which is where the research literature quotes it from.