Human-AI collaboration is rapidly transforming decision-making across domains, yet it introduces a complex set of cyber-social risks that extend beyond traditional technical concerns such as bias and opacity. This paper synthesizes prior literature to provide a structured understanding of these risks across five dimensions: cognitive, emotional and psychological, social and organizational, trust and information, and governance and accountability. We examine how challenges such as cognitive overload, AI anxiety, invisible labor, epistemic misalignment, and responsibility gaps emerge from the interaction between human cognition, system design, and institutional contexts. We also review current mitigation strategies, such as explainable AI, uncertainty visualization, frictional interaction design, team coordination frameworks, and governance models. Furthermore, we critically analyze their limitations, emphasizing ongoing trade-offs and unresolved tensions. Building on this analysis, we outline key future research directions aimed at improving trust calibration, supporting cognitive and emotional sustainability, strengthening human-AI team dynamics, and advancing accountable governance. Together, this work contributes a comprehensive, multi-level perspective on the risks and design challenges of human-AI collaboration, and provides a foundation for developing more responsible, human-centered, and sustainable AI systems.

PDF