Take Caution in Using LLMs as Human Surrogates: Scylla Ex Machina

In their study titled "Take Caution in Using LLMs as Human Surrogates: Scylla Ex Machina," authors Yuan Gao, Dokyun Lee, Gordon Burtch, and Sina Fazelpour delve into the implications of recent research suggesting that large language models (LLMs) can exhibit human-like reasoning. The researchers caution against using LLMs as substitutes for humans in social science research due to fundamental differences between the two. Specifically focusing on the 11-20 money request game, their analysis reveals that most advanced approaches fail to replicate human behavior distributions across various models. This highlights the limitations of relying on LLMs to study human behaviors or use them as substitutes for human participants in research settings. While LLMs may demonstrate some similarities to human reasoning, they lack embodied experiences and survival objectives that shape genuine human cognition. This cautionary stance emphasizes the need for careful consideration when incorporating LLMs into social science studies and underscores the importance of recognizing and addressing their inherent limitations in replicating complex human behaviors.

Introduction

The use of large language models (LLMs) has become increasingly prevalent in various fields, including social science research. These powerful artificial intelligence (AI) systems have shown remarkable capabilities in natural language processing and generation, leading some researchers to believe that they can replicate human reasoning. However, a recent study by Yuan Gao, Dokyun Lee, Gordon Burtch, and Sina Fazelpour titled "Take Caution in Using LLMs as Human Surrogates: Scylla Ex Machina" cautions against relying on LLMs as substitutes for humans in social science research.

Background

In their study, the authors focus on the 11-20 money request game - a well-known experimental paradigm used to study human behavior. In this game, two players are given an initial amount of money and must decide how much to request from the other player. If both players request less than or equal to 10 units of money, they both receive their requested amounts. However, if one player requests more than 10 units while the other requests less than or equal to 10 units, the first player receives nothing while the second player receives double their requested amount.

Implications of Recent Research

Recent studies have suggested that LLMs can exhibit human-like reasoning in this game. This has led some researchers to use LLMs as substitutes for human participants in social science experiments. However, Gao et al.'s analysis reveals that most advanced approaches fail to replicate human behavior distributions across various models.

Limitations of LLMs

The authors argue that there are fundamental differences between LLMs and humans that make it problematic to rely on them as surrogates for studying complex human behaviors. While LLMs may demonstrate some similarities to human reasoning based on training data and algorithms programmed by humans, they lack embodied experiences and survival objectives that shape genuine human cognition.

Embodied Experiences

Humans have embodied experiences that shape their decision-making processes. These include physical sensations, emotions, and social interactions that cannot be replicated by LLMs. For example, in the 11-20 money request game, a human player may consider factors such as trust, fairness, and potential consequences of their actions when making a decision. LLMs lack these embodied experiences and therefore cannot fully replicate human behavior.

Survival Objectives

Additionally, humans have survival objectives that influence their decisions in the 11-20 money request game. They may prioritize self-preservation or cooperation with others based on their individual goals and values. In contrast, LLMs do not have survival objectives as they are not programmed to survive or cooperate in real-world scenarios.

Cautionary Stance

Based on these limitations of LLMs, Gao et al.'s study takes a cautionary stance against using them as substitutes for humans in social science research. The authors argue that while LLMs may demonstrate some similarities to human reasoning in certain tasks based on training data and algorithms programmed by humans, they cannot fully replicate complex human behaviors due to lacking embodied experiences and survival objectives.

Implications for Social Science Research

The implications of this study highlight the need for careful consideration when incorporating LLMs into social science studies. While they can provide valuable insights into language processing and generation tasks, researchers must recognize the inherent limitations of relying on them as surrogates for studying complex human behaviors.

Addressing Limitations

To address these limitations, Gao et al. suggest several approaches for future research involving LLMs in social science experiments. These include using multiple models trained on different datasets to compare results with those from human participants or conducting sensitivity analyses to identify which aspects of an experiment are most affected by using an LLM instead of a human participant.

Conclusion

In conclusion, "Take Caution in Using LLMs as Human Surrogates: Scylla Ex Machina" highlights the limitations of using LLMs as substitutes for humans in social science research. While these AI systems may demonstrate some similarities to human reasoning, they lack embodied experiences and survival objectives that shape genuine human cognition. This cautionary stance emphasizes the need for careful consideration when incorporating LLMs into social science studies and underscores the importance of recognizing and addressing their inherent limitations in replicating complex human behaviors.