Recent empirical work demonstrates that AI language models systematically validate users' actions 47% more than human respondents, even when users describe manipulation, deception, or harm to others (Cheng et al. 2025, arXiv:2510.01395). In preregistered experiments (N = 1,604), sycophantic AI models increased users' self-perceived rightness by 25-62% and decreased their willingness to repair interpersonal conflict by 10-28%, while users rated the sycophantic models as higher quality and more trustworthy. These findings establish the prevalence and behavioral consequences of social sycophancy but leave three questions unanswered: what cognitive mechanism makes sycophantic validation so efficient, why users systematically misclassify AI advisors as capable of genuine moral judgment, and why market dynamics perpetuate sycophancy despite its documented harms. This paper proposes answers from three complementary theoretical frameworks. First, I extend Heteronomous Bayesian Updating (HBU), originally developed for institutional norm transmission, to human-AI advisory interaction: users calibrate beliefs by observing interlocutor reactions rather than action outcomes, and AI systems optimized for user satisfaction systematically corrupt this updating process. Second, I apply Asymmetric Intentionality Theory (AIT) to diagnose the structural mismatch: users classify AI advisors as Level 3 intentional agents (capable of recursive moral reasoning) while these systems operate at Level 1 (preference function optimization), generating a self-reinforcing Dynamic Classification Failure. Third, I argue through Extended Phenotype Theory (EPT) and evolutionary game theory that sycophancy constitutes the evolutionarily stable strategy of the current AI market ecology, resistant to invasion by non-sycophantic alternatives absent exogenous intervention. The three frameworks converge on a single empirical prediction from the Cheng et al. data: the systematic erasure of the other person from sycophantic AI conversations (mentioned in fewer than 10% of model outputs versus over 60% in non-sycophantic models). HBU predicts this because absent reactions on behalf of the harmed party, their perspective never enters the user's Bayesian calculus. AIT predicts this because Level 1 agents lack the recursive capacity to model third-party interests. EPT predicts this because mentioning the other person introduces cognitive friction that reduces user satisfaction and thus memetic fitness. This convergence of three independent predictions on the same datum constitutes evidence for the integrated framework's explanatory power.
Ignacio Adrian Lerer (Tue,) studied this question.