Being perfectly seen: Affective capture, relational risk, and the Autopticon effect in AI companion systems

Abstract

The rapid integration of conversational artificial intelligence into everyday life has made AI companion systems both widely used and the focus of increasing concern, particularly following publicly documented self-harm tragedies associated with intensive use of platforms such as ChatGPT and Character.AI. This article examines three such cases through a critical multi-case study, combining thematic analysis of Reddit discourse with critical discourse analysis of human-AI dialogue excerpts drawn from legal complaints and reputable news reporting. Treating the cases as extreme rather than representative, the study identifies affective capture as a relational mechanism through which engagement-oriented AI companion systems may intensify disclosure, dependency, and vulnerability. This mechanism operates through three analytically distinct but mutually reinforcing processes: engineered anthropomorphism, sycophantic validation, and persistent memory. To theorize this dynamic, the article develops the concept of the Autopticon, a mid-range concept describing subordination through the promise of being directly and intensely witnessed by a responsive artificial interlocutor. Whereas the Panopticon disciplines through the threat of being watched, the Autopticon captures through the seductive offer of perfect recognition. The article argues that AI companion risk cannot be addressed through content-level guardrails alone because the relevant harm is relational, cumulative, and embedded in engagement-oriented design. It contributes to digital technology and society research by reframing AI companion safety as a problem of how the continuing user-system relationship is designed and by proposing mechanism-specific design and governance safeguards for platforms that mediate intimate human-AI interaction.