Beginning in 1971, a small team of researchers at IBM began to divert the fields of speech recognition, language processing, and artificial intelligence away from the simulation of human language processes and towards a controversial new mandate: "to find the natural way for the machine to do it." In contrast to the prevailing paradigm of AI research in the period, which sought to design computers to replicate the underlying mechanisms of human perception, reasoning, and expertise, IBM's Continuous Speech Recognition group reformulated speech and language processing as a radically computational problem of large-scale, data-intensive, statistical pattern recognition. This talk discusses the history of automatic speech recognition as a problem situated between artificial intelligence and human-computer interaction, one that became pivotal in the precipitous rise and present-day dominance of data-driven machine learning as a privileged and pervasive mode of computational knowledge.