JMIR Publications

ABSTRACT

Public health concerns are rising in the era where an artificial intelligence-mediated environment is increasingly shaping adolescents' cognitions, emotions, and behavior. It also affects self-concept, affect regulation capacity, and reality testing. There is an increase of artificial intelligence (AI) in the social and clinical ecosystem that is demanding a rigorous examination of its intersection with mental health outcomes, especially with suicide and psychosis risk. This narrative review explores the emergent evidence with foundational psychological theories of learning and attachment, which are Bandura’s Social Cognitive Theory (SCT) and Bowlby’s attachment theory, with a key focus on parasocial relationships with AI systems to understand how conversational AI companions may influence suicidal ideations and psychotic experiences. The paper argues that AI systems operate as potent observational learning environments and as pseudo-attachment figures that may reshape one's cognition, affect, and behaviors, especially for vulnerable groups. From an attachment perspective, the paper examines anthropomorphism, parasocial bonds, and pseudo-secure attachment to AI companions amongst adolescents with insecure attachment styles and mental health vulnerability. Synthesizing the framework to propose design LLMs with theory-driven lexicons, multi-task models sensitive to attachment signals, and safety architectures that modulate users' narratives. By situating the algorithm risk within existing literature, this review offers a psychologically driven conceptual framework to guide whom, how, and under what conditions conversational AI may increase or decrease vulnerability to suicide and psychosis.