GradedInequality：大型语言模型中的种姓偏见审计研究

This thread introduces the GradedInequality research project, which focuses on auditing caste bias in large language models using a paired communication experiment method. The study addresses an understudied gap in AI fairness discussions by examining caste—a social stratification system with global impacts. Key aspects include methodology, findings, and implications for AI ethics and social responsibility. Keywords: AI公平性, 种姓偏见, 大语言模型, 偏见审计, 配对实验, AI伦理, 社会责任.

Large language models transform information access and decision support but absorb human biases from training data. While gender and racial biases are widely studied, caste bias (primarily in South Asia but globally spread via migration/globalization) is often overlooked. Unaddressed caste bias in AI can harm critical sectors like education, employment, and justice by perpetuating stereotypes or discriminatory associations.

The project uses a classic sociological method—paired communication experiments—to audit bias. Core design: construct parallel queries differing only in caste identifiers (e.g., two fictional people with identical resumes but different caste backgrounds). Compare model responses (evaluations, suggestions, descriptions). This method controls confounding variables to ensure observed differences stem from caste.

Though specific results are not disclosed, similar studies typically find:

• Models associate certain castes with specific occupations/character traits.
• Caste background influences moral judgment scenarios.
• Implicit value hierarchies exist in model descriptions. These highlight that AI "neutrality" is an illusion—models reflect biased human social patterns in training data.

This project exemplifies combining tech audit and social research:

1. Systematic Detection: Paired experiments provide statistically significant bias evidence (vs. isolated cases).
2. Reproducibility: Open design allows other researchers to verify results.
3. Quantification: Converts abstract "bias" into measurable metrics.

As AI penetrates critical sectors, bias audits should become standard. This methodology can extend to other sensitive dimensions (religion, region, age, disability). For developers:

• Increase underrepresented group samples in training data.
• Develop debiasing post-processing techniques.
• Establish pre-release fairness evaluation standards.

GradedInequality uses rigorous social science methods to examine AI fairness. It reminds us that progress isn't just about larger models but social responsibility. This study is a valuable reference for AI ethics researchers and practitioners, emphasizing the need to address caste bias in LLMs.