Cognitive Debt and the Regulatory Blind Spot: Bridging Neurocognitive Evidence, Practitioner Observation and the EU AI Act on AI in Education

Adoption of large language models in education has reached a scale that the European Union’s principal regulatory instrument for high-risk artificial intelligence (Regulation (EU) 2024/1689, the AI Act) was not designed to anticipate in full. This paper argues that the regulatory architecture governing educational AI under Annex III, point 3(b), and specifically the human oversight requirements of Article 14, addresses synchronic risks at the moment of decision but does not address the diachronic risk of cognitive debt: the structural erosion, across the lifecycle of sustained user engagement, of the cognitive substrate that meaningful human oversight presupposes. We make this argument across three integrated lines of evidence. First, we synthesise the convergent neurocognitive literature and identify four mechanisms through which cognitive debt accumulates: cognitive offloading, atrophy through disuse, transfer-appropriate processing failure and engagement asymmetry. Second, we report longitudinal practitioner observations gathered by the first author across twelve years of software-engineering management roles spanning the pre- and post-LLM transition, suggesting that the experimental findings reproduce at the scale of professional practice. Third, building on a recent analysis of automation bias published in this journal, we identify what we term the cognitive blind spot of Article 14: the assumption, structurally embedded in the provision, that the supervisor retains a cognitive substrate that the supervised activity, performed sustainedly under the regime, progressively erodes. We conclude by deriving operational implications of a developmental and substitutive distinction for institutions, providers, and regulators, and by indicating empirical and policy work required to address the gap before the high-risk obligations enter into force.