This survey explores the intersection of AI and social science, categorizing it into two directions: AI for social science and the social science of AI. AI for social science uses AI as a powerful tool to enhance various stages of social science research, such as literature searching, hypothesis generation, and data analysis. The social science of AI examines AI agents as social entities with human-like cognitive and linguistic capabilities. The paper provides a comprehensive framework to understand the distinctions and connections between these two directions, summarizes state-of-the-art experiment simulation platforms, and highlights the potential and challenges of integrating AI into social science research. Recent advancements in large language models (LLMs) have significantly influenced the field, enabling more efficient and effective research methods. In AI for social science, LLMs are used for hypothesis generation, where they assist in literature review, hypothesis proposing, and data analysis. They can generate hypotheses, evaluate their quality, and provide insights into complex social phenomena. However, they face challenges such as generating unreliable information, sensitivity to prompts, and limited context length. In hypothesis verification, LLMs are used in experimental research to simulate human behavior, allowing researchers to study complex systems without ethical concerns. They can act as experiment assistants or as proxies for human behavior in simulations. In survey research, LLMs serve as proxies for specific human subpopulations, interactive interviewers, and result analysis tools. They can generate synthetic survey data, improve the accuracy of responses, and analyze qualitative data. However, they may introduce biases and require careful consideration of their limitations. In nonreactive research, LLMs are used for content analysis and existing statistics analysis. They can perform text classification, sentiment analysis, stance detection, and hate speech detection. They are also used for text generation tasks, such as natural language descriptions and future predictions. However, they may struggle with conversational data and complex expert taxonomies. The paper concludes that LLMs have significant potential to transform social science research, but their application requires careful consideration of ethical, technical, and methodological challenges. Future research should focus on improving the accuracy, reliability, and fairness of LLMs in social science applications.This survey explores the intersection of AI and social science, categorizing it into two directions: AI for social science and the social science of AI. AI for social science uses AI as a powerful tool to enhance various stages of social science research, such as literature searching, hypothesis generation, and data analysis. The social science of AI examines AI agents as social entities with human-like cognitive and linguistic capabilities. The paper provides a comprehensive framework to understand the distinctions and connections between these two directions, summarizes state-of-the-art experiment simulation platforms, and highlights the potential and challenges of integrating AI into social science research. Recent advancements in large language models (LLMs) have significantly influenced the field, enabling more efficient and effective research methods. In AI for social science, LLMs are used for hypothesis generation, where they assist in literature review, hypothesis proposing, and data analysis. They can generate hypotheses, evaluate their quality, and provide insights into complex social phenomena. However, they face challenges such as generating unreliable information, sensitivity to prompts, and limited context length. In hypothesis verification, LLMs are used in experimental research to simulate human behavior, allowing researchers to study complex systems without ethical concerns. They can act as experiment assistants or as proxies for human behavior in simulations. In survey research, LLMs serve as proxies for specific human subpopulations, interactive interviewers, and result analysis tools. They can generate synthetic survey data, improve the accuracy of responses, and analyze qualitative data. However, they may introduce biases and require careful consideration of their limitations. In nonreactive research, LLMs are used for content analysis and existing statistics analysis. They can perform text classification, sentiment analysis, stance detection, and hate speech detection. They are also used for text generation tasks, such as natural language descriptions and future predictions. However, they may struggle with conversational data and complex expert taxonomies. The paper concludes that LLMs have significant potential to transform social science research, but their application requires careful consideration of ethical, technical, and methodological challenges. Future research should focus on improving the accuracy, reliability, and fairness of LLMs in social science applications.