台大心理系

Artificial intelligence (AI) is increasingly deployed in culturally diverse environments, yet little is known about its intercultural capability or how this capability shapes human-AI collaboration. Using the Cultural Intelligence (CQ) framework as the theoretical lens, we investigate the intercultural capability of leading large language models across four studies. First, AI models exhibited distinct intercultural capability profiles, with cognitive CQ showing the greatest variation across models. However, AI models' self-reports showed limited correspondence with independent evaluations by CQ experts, suggesting that self-described intercultural capability is a poor indicator of demonstrated competence. Second, both American and Chinese users consistently perceived AI models developed within their own cultural context as possessing greater intercultural capability. Comparisons with expert evaluations further showed limited correspondence between users' perceptions and expert-evaluated intercultural capability. Third, perceived AI intercultural capability enhanced cross-cultural effectiveness only when users possessed high CQ, supporting a complementarity account rather than a substitution account of human-AI collaboration. Together, these findings suggest that the value of AI's intercultural capability is realized not by AI systems alone but through the interaction among AI intercultural capability, human CQ, and culturally situated perceptions. This research advances a relational perspective on culture in human-AI collaboration and provides a foundation for understanding intercultural capability in AI.