BiasCause: Evaluate Socially Biased Causal Reasoning of Large Language Models

2025-04-08

AI Summary

This work addresses the problem of implicit social bias in large language models (LLMs) during causal reasoning. We introduce the first evaluation framework specifically designed for assessing social bias in causal inference. Our method employs 1,788 questions spanning eight sensitive attributes, integrates explicit causal graph modeling, LLM-assisted yet human-validated annotation, multi-model comparative evaluation, and analysis of association–causation confusion. Applied to four state-of-the-art LLMs, it identifies 4,135 biased causal graphs. Key contributions include: (1) a novel taxonomy for classifying causal reasoning errors; (2) the discovery of “misattribution bias”—a systematic tendency wherein models erroneously conflate statistical associations with causal relationships and embed group-level stereotypes; and (3) three reusable, actionable strategies for mitigating such bias. This study systematically uncovers the underlying causal mechanisms driving social bias in LLMs, thereby filling a critical gap in understanding how biases emerge and propagate during causal reasoning.

Technology Category

• Reasoning under Uncertainty: Causality
• Machine Learning: Causal Learning
• Cognitive Modeling & Cognitive Systems: Conceptual Inference and Reasoning

Application Category

• Social Networks and Social Media: Fairness and bias in social network and social media analysis
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

While large language models (LLMs) already play significant roles in society, research has shown that LLMs still generate content including social bias against certain sensitive groups. While existing benchmarks have effectively identified social biases in LLMs, a critical gap remains in our understanding of the underlying reasoning that leads to these biased outputs. This paper goes one step further to evaluate the causal reasoning process of LLMs when they answer questions eliciting social biases. We first propose a novel conceptual framework to classify the causal reasoning produced by LLMs. Next, we use LLMs to synthesize $1788$ questions covering $8$ sensitive attributes and manually validate them. The questions can test different kinds of causal reasoning by letting LLMs disclose their reasoning process with causal graphs. We then test 4 state-of-the-art LLMs. All models answer the majority of questions with biased causal reasoning, resulting in a total of $4135$ biased causal graphs. Meanwhile, we discover $3$ strategies for LLMs to avoid biased causal reasoning by analyzing the"bias-free"cases. Finally, we reveal that LLMs are also prone to"mistaken-biased"causal reasoning, where they first confuse correlation with causality to infer specific sensitive group names and then incorporate biased causal reasoning.

Problem

• Evaluates biased causal reasoning in LLMs regarding sensitive groups
• Proposes framework to classify causal reasoning in LLM outputs
• Identifies strategies for LLMs to avoid biased reasoning

Innovation

• Propose framework to classify LLM causal reasoning
• Synthesize questions to test biased causal reasoning
• Analyze strategies to avoid biased causal reasoning