Scholarlink

<p>As artificial intelligence transforms the landscape of social science research, large language models (LLMs) like ChatGPT present both unprecedented opportunities and unprecedented challenges. This study explores the application of ChatGPT as “surrogates” or computational substitutes for human participants in sociological and socio-psychological research. By simulating responses to complex socio-cultural issues, we investigate how well ChatGPT can replicate human attitudes toward immigration, gender stereotypes, and LGB parenting attitudes. We utilized a general simulation model employing detailed demographic prompts to generate synthetic participant responses, assessing their accuracy and political biases. When addressing topics in selected socio-psychological domains (gender, migration, and family values), ChatGPT tended to produce responses aligned with liberal-prosocial norms, which likely reflect alignment and safety-training objectives rather than stable ideological commitments. Across all simulations, ChatGPT reproduced the direction and magnitude of human attitudinal trends with high predictive accuracy (low MAE and MAPE). However, explanatory power (R<sup>2</sup>) remained modest across models, indicating that the LLM can reliably approximate aggregate patterns but not the deeper structure of attitudinal variance. These findings suggest that LLMs may serve as useful proxy participants for distributional replication, though not yet for mechanistic inference in attitudinal research. Since GPT-based models are trained predominantly on Western and Anglophone textual data, their simulated attitudes may reflect culturally specific semantic patterns rather than globally representative ones. Thus, the present findings illustrate the potential of LLMs for attitudinal simulation within Western-linguistic contexts, while underscoring the need for culturally diversified models in future research. This research underscores the necessity for critical evaluation of AI-generated data in social science contexts and calls for further refinement of LLM methodologies.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

ChatGPT as a research proxy: simulating human attitudes in social science research

  • 期刊论文

  • 发表日期：

    2026年1月7日

查看全文
  • Antonina Rafikova,
  • Anatoly Voronin

摘要

As artificial intelligence transforms the landscape of social science research, large language models (LLMs) like ChatGPT present both unprecedented opportunities and unprecedented challenges. This study explores the application of ChatGPT as “surrogates” or computational substitutes for human participants in sociological and socio-psychological research. By simulating responses to complex socio-cultural issues, we investigate how well ChatGPT can replicate human attitudes toward immigration, gender stereotypes, and LGB parenting attitudes. We utilized a general simulation model employing detailed demographic prompts to generate synthetic participant responses, assessing their accuracy and political biases. When addressing topics in selected socio-psychological domains (gender, migration, and family values), ChatGPT tended to produce responses aligned with liberal-prosocial norms, which likely reflect alignment and safety-training objectives rather than stable ideological commitments. Across all simulations, ChatGPT reproduced the direction and magnitude of human attitudinal trends with high predictive accuracy (low MAE and MAPE). However, explanatory power (R2) remained modest across models, indicating that the LLM can reliably approximate aggregate patterns but not the deeper structure of attitudinal variance. These findings suggest that LLMs may serve as useful proxy participants for distributional replication, though not yet for mechanistic inference in attitudinal research. Since GPT-based models are trained predominantly on Western and Anglophone textual data, their simulated attitudes may reflect culturally specific semantic patterns rather than globally representative ones. Thus, the present findings illustrate the potential of LLMs for attitudinal simulation within Western-linguistic contexts, while underscoring the need for culturally diversified models in future research. This research underscores the necessity for critical evaluation of AI-generated data in social science contexts and calls for further refinement of LLM methodologies.