Abstract
Abstract
AI systems increasingly serve as social partners, companions, and romantic partners, yet governance remains fragmented, reactive, and split between clinical applications and anthropomorphic deception. This systematic review examines how engineering, ethics, and policy shape human-AI relationships between 2022-2026 and identifies challenges and opportunities at their intersections. Following PRISMA 2020, seven databases yielded 395 records; 169 studies met inclusion criteria. Reflexive Thematic Analysis and K-means clustering synthesized our findings. We identify the literature is polarized into three distinct clusters: (1) “Anthropomorphism vs. Deception in social-consumer contexts” (n=30) is characterized by empirical research lacking support at governance level despite studying deception and vulnerability, and it highlights a governance absent in such empirical research; (2) “Engagement vs. Safety” (n=83) characterizes highest governance proposals but lowest policy engagement and presents the highest standardization gaps, most of which are mentioned implicitly; (3) “Transparency vs. Trust in health-care” (n=56) has strong policy engagement but largest framing gap. Accountability gaps dominate ethic policy challenges; regulatory lag and principal to practice gaps dominate policy engineering. Value-Sensitive Design is the top engineering-ethics opportunity, yet remains under-implemented. The literature is polarized, leaving anthropomorphic deception and user vulnerability insufficiently addressed. Conceptual work outweighs empirical, limiting actionable regulation. The findings also reveal that current governance efforts are reactive and piecemeal, treating symptoms rather than underlying causes. We argue that a rational, staged, and anticipatory governance framework is urgently needed. A Multi-Stakeholder approach integrating engineering, ethics, and policy is essential for responsible human-centered AI governance.