Honest Ignorance as an Ethical Principle for Child-Oriented AI Design

Current large language models (LLMs), trained through reinforcement learning from human feedback(RLHF), produce outputs that are systematically confident, agreeable, and fluent—regardless ofaccuracy. This paper argues that this design poses a distinctive ethical problem for children, whoseepistemic vigilance mechanisms depend on the correlation between confidence and reliability thatcharacterizes human communication. Drawing on the philosophy of trust (Baier, Jones, Hawley,McLeod), vulnerability theory (Mackenzie, Rogers, & Dodds; Schweiger & Graf), the ethics ofepistemic simulation (Frankfurt, Hicks et al.), epistemic justice (Fricker), the capability approach(Nussbaum, Fricker), children's rights philosophy (Schapiro, Lansdown, Feinberg), and the concept oftransitional paternalism (Manson, Tucker), I develop "Honest Ignorance" as a normatively groundeddesign principle for child-oriented AI. The principle consists of four requirements: (1) provideknowledge without performing reasoning, (2) systematically express uncertainty, (3) franklyacknowledge errors, and (4) bridge to human informants. Each requirement is derived from thenormative framework rather than merely motivated by empirical findings. The paper positions HonestIgnorance as transitional paternalism—a justified, temporary, developmentally calibrated constraint thatcultivates the very epistemic capabilities that will eventually render it unnecessary. The historicalprecedent of leaded gasoline—where a population-level neurotoxin damaged children's cognitivedevelopment for decades before regulation—illustrates the cost of waiting for certainty before acting onconverging evidence of developmental harm.