Why Privacy Is Taught, Not Just Protected: A Relational Model for Children's Use of Large Language Models.

Large language models (LLMs) such as ChatGPT, Claude and Gemini are entering classrooms as dialogical, anthropomorphic systems, raising questions about how classroom framing shapes children's expectations about disclosure. Dominant approaches frame privacy in education either as data-protection compliance or as individual competence; both underestimate the contextual character of information flows and the trust dynamics that conversational systems set in motion. This conceptual paper develops a relational model of pedagogical privacy for AI-supported learning environments. It reads together three bodies of theory: Nissenbaum's contextual integrity, work on calibrated trust in human–machine interaction, and research on the anthropomorphising of digital systems, situated within relational and networked-privacy scholarship. Privacy is modelled as the outcome of three interlocking relations: a context-relational dimension (child–context), a trust-relational dimension (child–system) and a pedagogical-relational dimension (system–pedagogical setting). The model specifies how instructional framing may connect children's interpretation of anthropomorphic cues with expectations about understanding, confidentiality and appropriate disclosure. Its contribution is to make these relationships explicit for LLM-supported learning environments and to distinguish the expectations guiding disclosure from the normative assessment of the resulting information flows. A contrasting classroom case illustrates the mechanism. From the model we derive three design principles for LLM-supported learning environments (transparency, trust calibration and reflexive negotiation) and set out an empirical research agenda for the field. Privacy, on this account, is taught, not only protected. Preprint (not peer-reviewed)