Take caution in using LLMs as human surrogates.

Area of Science:

• Artificial Intelligence
• Social Sciences
• Computational Social Science

Background:

• Large language models (LLMs) demonstrate human-like text generation capabilities.
• LLMs are increasingly proposed as surrogates for human participants in social science research.
• Fundamental differences exist between LLM processing and human cognition, including embodied experience and survival drives.

Purpose of the Study:

• To evaluate the capability of LLMs to simulate human behavior in social science research.
• To assess the reasoning depth of LLMs in economic decision-making tasks.
• To identify limitations and potential biases when using LLMs as human surrogates.

Main Methods:

• Utilized the 11-20 money request game to test LLM reasoning.
• Compared LLM responses against established human behavior distributions.
• Analyzed various advanced LLM approaches and their performance across different models.

Main Results:

• Most advanced LLM approaches failed to accurately replicate human behavior distributions.
• Failure modes were diverse and unpredictable, influenced by factors like input language and model safeguarding.
• Significant discrepancies were observed between LLM outputs and human decision-making patterns.

Conclusions:

• LLMs, despite human-like responses, do not reliably simulate human behavior due to fundamental cognitive differences.
• Caution is advised when employing LLMs as surrogates or for simulating human behavior in research.
• Further research is needed to understand and mitigate LLM limitations in social science applications.