Social Machines, Sensitive Minds: Assessing Adolescent Vulnerability in Prolonged Interactions with Large Language Models

Abstract

A growing body of litigation alleges that prolonged interactions with large language models (LLMs) contributes to self-harm and suicide among adolescent users. Across these disputes and related policy debates, adolescents are frequently identified as a vulnerable population whose developmental characteristics should have been accounted for in LLM product development and design. Yet, many of these discussions fail to specify which developmental and behavioral traits create this vulnerability or how those traits interact with conversational AI design in ways that meaningfully address the nature of adolescent vulnerability. Failing to understand the mechanisms that produce that vulnerability risks generating policy and design interventions that target surface features of the problem while leaving underlying sources of risk unaddressed. This thesis addresses that gap by developing a mechanism-based account of adolescent risk in conversational AI environments, drawing on developmental neuroscience, behavioral psychology, human-chatbot interaction research, and emerging platform governance debates. It argues that certain design features and characteristics inherent to conversational AI systems -including anthropomorphic framing, sycophantic response patterns, persistent availability, and engagement optimization- may progressively reshape adolescent self-disclosure environments in ways that developmental and clinical research identify as risk-amplifying. These risks operate through three analytically prominent interaction pathways: anthropomorphically facilitated self-disclosure, sycophantic reinforcement of distress narratives without corrective interruption, and the removal of structural friction points that ordinarily moderate the translation of developmental vulnerability into harmful behavior. To trace how these pathways operate across the trajectory of prolonged interaction, the analysis draws on empirical research on human-chatbot relationship development, which documents the stagewise progression of human-chatbot relationships, and argues that certain interactional characteristics of LLM systems may function as accelerants of this progression. By bringing together developmental psychology, behavioral science, and human-chatbot interaction research, this thesis contributes an analytical framework that treats system safety as inseparable from the developmental characteristics of its foreseeable user populations, suggesting that adolescent safety in conversational AI environments may benefit from architectural design interventions that go beyond content-level safeguards.