As generative AI becomes embedded in education, research has largely emphasised outputs, ethics, and performance, with less attention to how these systems shape learners’ conduct and dispositions through the interaction itself. This conceptual paper addresses that gap by treating large language models (LLMs) not as neutral tools but as meaning-mediating infrastructures: interactive systems that organise what kinds of learning moves become easy, sensible, and repeatable, and how learners come to interpret effort, struggle, and success. Extending socio-cognitive accounts of feedback and norm perception through Luhmann’s systems theory, we introduce programmatic behaviour to describe how repeated engagement with structured environments – including AI systems – stabilises habitual orientations toward help-seeking, completion, and contestation. We also develop meaning mediation to explain how tone, pacing, dialogue structure, and closure shape the lived significance of error, progress, and competence in real time. On this basis, we outline three heuristic trajectories that may emerge through sustained LLM use – supportive, corrosive, and manipulative patterns – showing how behavioural drift is co-produced by interaction design and institutional framing. The paper argues that evaluating generative AI in education therefore requires attention to interactional defaults and their normative effects, and it proposes an AI literacy agenda centred on noticing and steering participation within AI-mediated infrastructures.
Ennion et al. (Wed,) studied this question.