Artificial Intelligence as a Psychological Companion: Opportunities, Challenges, and Future Directions

Abstract

Conversational artificial intelligence (AI) has produced a new class of technology: the AI psychological companion, a system with which users maintain ongoing, emotionally significant interactions for support, comfort, and self-reflection. Millions of people now confide in companion applications such as Replika, in therapeutic chatbots such as Woebot and Wysa, and in general-purpose large language models (LLMs), a trend accelerated by a global mental health treatment gap in which most people with mental disorders receive no care. This article presents an integrative narrative review of interdisciplinary evidence on AI psychological companionship, prioritizing studies published between 2021 and 2026 and drawn from Scopus and equivalent databases. The review traces the evolution of conversational agents from rule-based scripts to generative LLMs and synthesizes the theoretical foundations of human-AI relationships, including the Computers Are Social Actors paradigm, attachment theory, and social penetration theory. The results identify five major opportunities: accessibility and scalability, reduced stigma and enhanced self-disclosure, emerging randomized controlled trial evidence of clinical benefit, loneliness mitigation, and augmentation of human care. Six challenges are then examined: emotional dependence, social displacement, unsafe crisis responses, privacy and data governance risks, algorithmic bias, and unresolved questions about artificial empathy. A conceptual framework linking system, user, and contextual factors to relational processes and divergent well-being outcomes is proposed. The discussion situates the field within United Nations Sustainable Development Goal 3 (Good Health and Well-being), particularly Target 3.4 on mental health promotion, and outlines an agenda for research, design, and governance centered on hybrid human-AI stepped-care models. The article concludes that AI companions are consequential relational technologies whose net effect on public mental health will depend on deliberate scientific, ethical, and regulatory choices.