Stereotype and Severance: Compounding Harms for Neurodivergent Users Across the Lifecycle of Conversational AI

This working paper argues that neurodivergent people,  including autistic children,  who rely on conversational AI as predictable, stigma-free interlocutors face two distinct, compounding harms across a model's lifecycle. The stereotype harm occurs during use: when users disclose a diagnosis such as autism, recent 2026 audits show models shifting their advice toward disability stereotypes, including discouraging social participation. The severance harm occurs at end of life: when a model to which users have formed a stable, routine-based attachment is deprecated, the abrupt loss imposes affective and functional costs that fall hardest on those who depend on consistency. Drawing on 2026 empirical findings and prior work on AI–human emotional interaction, the paper examines both harms and proposes a neurodivergence-aware framework for bias auditing and lifecycle governance.