It comes from a 2021 paper by Emily Bender, Timnit Gebru, and coauthors criticizing large language models trained on scraped text at scale. The phrase is still invoked in debates over whether fluent model output reflects real reasoning or sophisticated statistical mimicry.