Measuring Stereotype and Deviation Biases in Large Language Models

2025-08-08

AI Summary

Large language models (LLMs) exhibit socially harmful biases, yet existing work conflates distinct bias types—stereotypical bias (fixed attribute associations with social groups) and demographic deviation bias (systematic divergence from real-world population distributions). Method: We propose a multi-round prompting framework to elicit fine-grained individual profiles across political orientation, religious affiliation, and sexual orientation from four state-of-the-art LLMs, then quantitatively compare outputs against authoritative demographic benchmarks. Contribution/Results: This is the first study to decouple and jointly quantify both bias types, revealing structural inference biases in user attribute modeling. All models exhibit significant, robust dual biases—stereotypical associations persist even when demographic alignment improves, and vice versa. These findings expose concrete risks of algorithmic unfairness in real-world applications and establish a reproducible, benchmark-driven methodology for bias detection and alignment governance.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Ethics — Bias, Fairness, Transparency & Privacy
• Philosophy and Ethics of AI: Bias, Fairness & Equity

Application Category

• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Social Networks and Social Media: Fairness and bias in social network and social media analysis
• Search and Retrieval-Augmented AI: Large language models for search

Abstract

Large language models (LLMs) are widely applied across diverse domains, raising concerns about their limitations and potential risks. In this study, we investigate two types of bias that LLMs may display: stereotype bias and deviation bias. Stereotype bias refers to when LLMs consistently associate specific traits with a particular demographic group. Deviation bias reflects the disparity between the demographic distributions extracted from LLM-generated content and real-world demographic distributions. By asking four advanced LLMs to generate profiles of individuals, we examine the associations between each demographic group and attributes such as political affiliation, religion, and sexual orientation. Our experimental results show that all examined LLMs exhibit both significant stereotype bias and deviation bias towards multiple groups. Our findings uncover the biases that occur when LLMs infer user attributes and shed light on the potential harms of LLM-generated outputs.

Problem

• Measure stereotype bias in large language models
• Assess deviation bias in LLM-generated content
• Examine biases in demographic attribute associations

Innovation

• Profiles generation to detect biases
• Examines stereotype and deviation biases
• Tests four advanced LLMs for biases