Large language models (LLMs) now answer before the teacher does. This theoretical paper asks what kind of mediation an LLM fails to provide, and what the teacher's role becomes once the machine answers first. Drawing on Vygotskian accounts of mediation, on Feuerstein's criteria for mediated learning as presented by Kozulin, on Tharp and Gallimore's means of assisting performance, and on research into teacher and AI complementarity, it argues that an LLM supplies information without intention, without adjustment to the learner and without responsibility for development. What remains for the teacher is therefore not content but the government of the division of labour between learner and machine. The paper proposes six teacher moves that reallocate that work: setting the division of labour, withholding help, diagnosing what the learner can almost do, graduating feedback, problematising machine output, and making the unassisted level visible. Together these define a second order mediation, in which the teacher mediates the learner's relation to the tool rather than to the content alone. Implications are drawn for EFL writing and for teacher preparation in Algerian universities.