Abstract
<p>The integration of generative artificial intelligence into classrooms forces a re-examination of assumptions that education has long left unquestioned. Access is no longer the central issue. What demands scrutiny is why identical tools yield divergent outcomes across learners — and what that divergence exposes about the cognitive and institutional architecture of schooling itself.This paper develops an interpretive framework organised around four structural tensions rendered newly visible by AI's arrival, and proposes a reorientation of educational thinking from tool availability toward cognitive process, instructional sequence, and epistemic accountability. The first tension is a metacognitive gap: the benefits of AI access are mediated not by socioeconomic variables but by students' prior capacity for self-regulation, making equal access a systematic amplifier of existing inequality. The second is a friction paradox: the efficiency logic driving AI-assisted learning directly contradicts the desirable difficulty principles that cognitive neuroscience establishes as necessary for durable retention. The third concerns differential neuroplasticity — the as-yet unanswerable question of what sustained cognitive externalisation does to developing minds, a concern for which no adequate longitudinal evidence yet exists. The fourth is the epistemic collapse of assessment: the foundational inference from observable product to internalised competence has been destabilised in ways that detection tools cannot remedy.These tensions are examined through a two-cycle practitioner inquiry (n=18) at a German-model international school in South America, during preparation for the Deutsches Sprachdiplom I (DSD I). A structured five-step AI-assisted practice sequence produced a mean gain of +0.8 points (Δ SK1→SK2), with marked internal dispersion. Students who improved by two or more points attributed their learning, without exception, to the teacher's initial explanation — the step preceding any AI involvement. Together, these findings suggest that instructional sequence matters more than tool availability. The challenge of preparing education for an uncertain future is not primarily technological. It is epistemological.</p>