Actions Speak Louder than Words: Agent Decisions Reveal Implicit Biases in Language Models

TL;DR

This paper introduces a technique to reveal implicit sociodemographic biases in large language models by analyzing their decision-making across different personas, uncovering significant and amplified biases compared to real-world disparities.

Contribution

The study presents a novel method to systematically detect implicit biases in LLMs through sociodemographically-informed decision scenarios, revealing biases that are often amplified in advanced models.

Findings

State-of-the-art LLMs show significant sociodemographic disparities.

More advanced models exhibit greater implicit biases.

Biases uncovered are directionally aligned but amplified compared to real-world data.

Abstract

While advances in fairness and alignment have helped mitigate overt biases exhibited by large language models (LLMs) when explicitly prompted, we hypothesize that these models may still exhibit implicit biases when simulating human behavior. To test this hypothesis, we propose a technique to systematically uncover such biases across a broad range of sociodemographic categories by assessing decision-making disparities among agents with LLM-generated, sociodemographically-informed personas. Using our technique, we tested six LLMs across three sociodemographic groups and four decision-making scenarios. Our results show that state-of-the-art LLMs exhibit significant sociodemographic disparities in nearly all simulations, with more advanced models exhibiting greater implicit biases despite reducing explicit biases. Furthermore, when comparing our findings to real-world disparities reported in empirical studies, we find that the biases we uncovered are directionally aligned but markedly amplified. This directional alignment highlights the utility of our technique in uncovering systematic biases in LLMs rather than random variations; moreover, the presence and amplification of implicit biases emphasizes the need for novel strategies to address these biases.