Supervised learning needs annotations: the correct label for each example. Sometimes they can be obtained automatically, but often they must be produced by hand, a labor-heavy process on which the model's quality depends (Chollet, 2021, sec. 6.1.2). Chollet (2021) lists the options: annotate the data yourself, hire a specialized company, or use a platform like Mechanical Turk, cheap and scalable but noisy.
The founding case is ImageNet, the image dataset on which AlexNet was trained. It was built on the lexical hierarchy of WordNet with workers from Amazon Mechanical Turk (Deng et al., 2009). By Fei-Fei Li's estimate, labeling everything she planned by hand would have taken one person about 19 years; Mechanical Turk, which Amazon had built to weed out duplicate product pages, made it possible with some 49,000 people in 167 countries. The platform is named after the 1770 chess-playing automaton that hid a human player (the Turk), and Jeff Bezos described it as "artificial artificial intelligence" (Mollick, 2026).
Today's assistants repeat the pattern. InstructGPT's RLHF relied on about 40 contractors hired through Upwork and Scale AI, mostly in the United States and Southeast Asia (Ouyang et al., 2022, sec. 5.2). An investigation by journalist Billy Perrigo (2023) revealed that workers in Kenya, through the company Sama, earned between $1.32 and $2 an hour labeling violent and abusive texts for ChatGPT's safety filter. Gray & Suri (2019) call this labor, made invisible by the interface, "ghost work."
The point is not only about labor but about epistemology. When we say that a model "learned" to recognize dogs or to be polite, part of that competence is made of human judgments frozen into data. To whom should we attribute the intelligence of a system whose correct behavior was defined, example by example, by thousands of anonymous people?