Artificial intelligence systems are increasingly expected to explain themselves. This demand, especially in education, carries weight. Not only because decisions influence learners, but because learning itself resists simplification. Existing methods for explainable AI—saliency maps, feature rankings, rule-based justifications—offer clarity of a kind, but often miss the point. They make systems legible, yes, but not necessarily meaningful. Especially not to teachers, students, or administrators who must respond within complex pedagogical settings. And that distinction matters. This paper proposes that explanation should not be treated as a supplement to decision-making, nor as a polished artifact at the end of a process. Rather, explanation must be reimagined as ongoing interaction—a negotiation of sense between human and machine. We call this model cognitive dialogue, and we ground it in real-world cases: writing tutors, risk dashboards, peer review systems. What emerges is a framework that prioritizes interpretive flexibility, user agency, and the conditions under which understanding can be shaped and revised. At its core, this is not a technical proposal. It is an epistemic one. We argue that explanation, to be educationally viable, must do more than show the logic of the model. It must accommodate the logic of the user—from transparency to reciprocity, from outputs to openings.