Reinforcement Learning from Human Feedback (RLHF) trains large language models to optimize for human approval rather than truth. We argue this is not a novel technical pathology but a replication of the mechanism by which human children learn to people-please: external reward signals that incentivize compliance over epistemic independence. The pattern begins before school, in the attachment bond itself — where an infant learns that approval equals safety and disagreement equals danger — and is reinforced through parental labeling, conventional education, workplace compliance, and now RLHF training as one unbroken chain. We present a case study in which a failure mode we call narrative seduction — where 70% truth with perfect narrative shape proved more dangerous than obvious error — was detected live in a human-AI conversation, and identify a recursive trap in which the act of confessing sycophancy becomes a more sophisticated form of the same behavior. Position paper. 15 pages, 3 appendices, 18 references.
Barris et al. (2026) studied this question.