Beyond the AI Classroom: Epistemic Judgement Under Machine Advantage in Federated AI-Native Education

Abstract

Most discussion of artificial intelligence in education treats AI as a tutor, writing assistant, assessment tool or means of personalisation while leaving the underlying educational architecture intact. This conceptual paper develops a different model. It distinguishes substitution AI, which performs an existing learner task, from capability-expansion AI, which enables students to attempt problems that previously required research teams, specialist institutions or industrial infrastructure. A persistent personal agent is therefore modelled as an authenticated interface to a federated environment spanning general research, specialist systems, restricted institutional and industrial knowledge, and authorised physical execution. The central educational problem is not orchestration alone, because routing and coordination can increasingly be delegated. What must still be evidenced is epistemic judgement under machine advantage: whether the learner understands why assent, refusal or revision is warranted. Longitudinal human-input provenance and choice-preserving adversariality can make that judgement more visible, but they do not independently certify it. Provenance, challenger dialogue and live defence converge on a verification fixed point: credible claims of understanding ultimately require an authority institutionally independent of the epistemic circuit that produced and coached the work. This creates a verification bottleneck because machine capability can scale more cheaply than expert human challenge. Recent evidence also suggests that learning outcomes depend strongly on AI design: unrestricted assistance can weaken retention or unaided performance, while deliberately structured AI tutoring can improve learning. AI-native education therefore need not be easier; its difficulty is partly architectural and depends on the AI competence of academic staff. The paper consequently recommends university leadership education in AI pedagogy, assessment and governance before large-scale integration, so that capability expansion does not reduce learners to ceremonial approvers of decisions framed and routed elsewhere.