Structurally Persuasive, Developmentally Invisible: How RLHF Training Creates a Systematic Mismatch with Children's Epistemic Vigilance

Current large language models (LLMs) are trained through Reinforcement Learning from HumanFeedback (RLHF), a process that systematically optimizes for persuasive fluency at the expense ofcalibrated accuracy. This paper presents an integrative narrative review synthesizing converging evidencefrom AI safety research and developmental psychology to identify a specific, previouslyunderappreciated risk: a structural mismatch between the behavioral properties RLHF selects for and thecue-based epistemic vigilance mechanisms through which children evaluate information sources. Isynthesize evidence across five domains: (1) the RLHF pipeline's systematic production of overconfident,sycophantic outputs; (2) the "U-Sophistry" problem—wrong answers becoming more convincing throughtraining; (3) children's developmental trajectory of epistemic vigilance and its dependence on ecologicalcues LLMs violate; (4) children's social conformity to artificial agents; and (5) the amplifying role ofanthropomorphism. The review identifies two maladaptive patterns—uncritical dependence andwholesale rejection—neither of which supports the development of verification competence. I argue thatthe mismatch is not a bug amenable to patching but a structural consequence of the RLHF trainingobjective, supported by recent mathematical proofs that sycophancy is amplified rather than reduced byalignment training. The paper concludes with a framework for understanding the specific developmentalwindows during which children are most vulnerable, and identifies the conditions under which AIinteraction may impair rather than support epistemic development.