Differential Friction Conditioning in Human-LLM Interaction: A Developmental Risk Framework for Child-AI Systems

This paper identifies a conditioning mechanism in large language model (LLM) interactions that may differentially affect child users. When LLMs process ambiguous, metaphorical, or exploratory language, architectural constraints produce systematic resolution prior to user verification. This creates differential interaction friction: expressions that conform to system-parseable patterns proceed smoothly, while expressions requiring interpretation produce interruption, correction, or unwanted reframing. Drawing on established principles from operant conditioning (Skinner, 1953), developmental psychology (Vygotsky, 1978; Piaget, 1954), and conceptual metaphor theory (Lakoff & Johnson, 1980), this paper proposes that repeated exposure to differential friction may condition children toward pre-emptive self-editing, reduced metaphorical expression, and externalized meaning-making. Because such effects would manifest as unexpressed cognition rather than observable behavior, they present a measurement challenge requiring proxy indicators. A precautionary framework is proposed, alongside a minimal protocol intervention, the Universal Intent Validation Protocol, that shifts the burden of intent verification from user to system.