The ubiquity and ease of use of large language models makes it easy to overlook the interactional and interpretive processes at play. To understand the attraction of this technology we need to trace its sociotechnical roots. From divination and horoscopes and from ELIZA to present-day large language models, I document how people have been thinking with things, outsourcing judgement, and making sense of interactively presented non-sense. Following the lead of Lucy Suchman to “slow down discourses of the ‘smart’ machines”, I consider the interactional foundations of our engagement with technologies of language. I make the case that the fluid output, fine-tuned overconfidence, and interactive design of these computational artefacts conspire to exploit our interpretive processes and interactional infrastructure, rendering them irresistible to lay people and researchers alike. This means that a deep understanding of processes of human interaction and sense-making will be a foundational resource for the growing arsenal of methods in critical AI literacy. Preprint of a chapter for an edited volume: A Research Agenda for Critical AI Studies. Currently under review, likely to be revised. Your comments are welcome!