洞中流影:基于大语言模型的硅基样本在社会科学研究中的应用反思  

Floating Shadows in the Cave: Reflections on Using Large-Language-Model-Generated Silicon Samples in Social Science Research

作　　者：胡安宁[1] Hu Anning

机构地区：[1]复旦大学社会发展与公共政策学院

出　　处：《社会发展研究》2025年第1期22-39,M0003,M0004,共20页Journal of Social Development

基　　金：国家社会科学基金项目“社会团结的文化基础研究”(项目编号:22VRC140,主持人:胡安宁)的阶段性研究成果。

摘　　要：无论是质性还是量化,大多数社会科学经验研究均立足于以现实生活中活生生的个体为对象收集信息。但是,以生成式人工智能为基础的硅基样本却对这一传统研究路径带来了潜在的冲击。硅基样本通过提示词的设置让生成式人工智能平台“扮演”具有特定社会人口学特征的人类对象,并以此反馈信息、提供研究数据。本文立足于对现有研究的系统梳理,提炼和总结了硅基样本使用过程中的诸多限制:包括提示词设定的特异性、敏感性、全面性与过度修正;生成数据中的信息扭曲、信息误差和信息冗余;对现实世界数据生成过程的更低限度的描述;对特定数据集的过拟合以及统计推断过程不确定性的增加。在此基础上,本文对当下社会科学领域内硅基样本的使用提出应用反思。Whether qualitative or quantitative,most empirical research in the social sciences relies on collecting information from real-life individuals. However,silicon samples,powered by generative artificial intelligence,pose a potential challenge to this traditional research path. These samples use prompts to instruct generative AI platforms to“simulate”human subjects with specific sociodemographic characteristics,providing research data through their feedback( such as evaluations,judgments,or decisions). While fields like economics,political science,and psychology have already produced a significant body of empirical findings highlighting the unique value of silicon samples,fully leveraging their potential requires researchers to better understand the errors and biases inherent in their use. Based on a systematic review of existing studies,this paper identifies and summarizes various limitations in the use of silicon samples. These limitations include factors such as the specificity, sensitivity,comprehensiveness, and over-correction of prompt settings;information distortion, errors, and redundancy in generated data;the limited description of real-world data generation processes;overfitting to specific datasets;and an increased level of uncertainty in statistical inference processes. Based on this analysis,the paper provides reflections and recommendations on the use of silicon samples in social science research.

关 键 词：生成式人工智能 硅基样本 社会科学研究 范式 

分 类 号：C91[经济管理]