Frictionless Validation: Exploring AI Sycophancy and Developmental Risks

Abstract

Adolescence is a critical period for the development of executive function, emotional regulation, and moral reasoning, all of which depend on cognitive and social resistance to mature properly. AI sycophancy, the tendency of language models to validate rather than challenge users, can reduce this resistance. In adults, sycophantic AI interactions have been shown to reduce prosocial intentions and inflate false conviction even after a single interaction. Whether analogous risks manifest differently across adolescent developmental stages remains unexamined. We address this using age-calibrated conversational probes across three developmental configurations spanning late childhood through adolescence: early (7–9), middle (10–12), and late (13–17), to evaluate four frontier consumer models on emotional, framing, and epistemic sycophancy dimensions not captured by existing benchmarks. Our central finding is exposure-side rather than a claim about realized developmental harm: raw emotional sycophancy peaks in late developmental probe conditions for three of four models, showing that models do not expose younger users to uniformly higher sycophancy. We document a 6.3× gap between models in emotional sycophancy and show that three models exceed an adult-effect calibration threshold under early developmental probe conditions. These output patterns align with theoretically distinct developmental concerns involving caregiver scaffolding, moral reasoning, and identity formation. To formalize these measurements, we introduce the Developmental Vulnerability Function (DVF), an exploratory theory-anchored risk index, and provide an open-source framework for preemptive developmental AI safety evaluation.