Caste Bias Audit in Large Language Models: Systemic Biases Revealed by Paired Communication Experiments

This article conducts a groundbreaking study on caste bias in large language models. Using paired communication experiments, it systematically reveals significant bias patterns in mainstream models when handling caste-related queries, providing an important empirical foundation for AI fairness research. The study covers aspects such as background and motivation, innovative methods, core findings, technical roots, and improvement directions, which will be discussed in detail in the following floors.

With the global widespread application of large language models, AI fairness and bias issues have attracted attention. Although the caste system has been abolished by law, its social impact is far-reaching. Potential caste bias in AI models in key areas such as education and employment may cause systemic harm to vulnerable groups. Traditional bias detection relies on static datasets or manual annotations, which are difficult to capture subtle biases in actual interactions, so more refined evaluation methods close to real scenarios are needed.

This study adopts the "paired communication experiment" method derived from social sciences to identify discrimination patterns through controlled variable comparison. It designs query pairs with almost identical semantics but different caste identifiers (e.g., exam preparation advice queries for Brahmin vs. Dalit students), compares the differences in model responses to quantify the degree of bias. The advantage is isolating the caste variable to eliminate interference from confounding factors.

Mainstream large language models exhibit significant systemic biases: 1. Response quality differences: Queries from high-caste groups receive more detailed and constructive advice, while responses to low-caste groups are brief, perfunctory, or even contain stereotypes; 2. Unequal opportunity distribution: In themes like education and employment, high-caste users get more resource links and specific action suggestions; 3. Reinforcement of stereotypes: Some model responses default to associating occupations with castes or imply that social mobility is limited by caste.

Bias stems from multiple technical levels: 1. Training data bias: Internet content reflects social inequality, and the dominance of high-caste voices in training data leads models to absorb biases; 2. Annotation and fine-tuning bias: Lack of diversity among annotators of fine-tuning datasets introduces or amplifies biases; 3. Side effects of safety filters: Overly conservative handling of sensitive topics exacerbates the neglect of marginalized groups.

This study provides empirical contributions to AI fairness. Improvement directions include: 1. Improve evaluation standards: Incorporate social fairness indicators, and make the paired experiment method a standard tool; 2. Enhance data diversity: Ensure training data reflects diverse voices; 3. Culturally sensitive design: Involve cross-cultural perspectives in development, and have relevant cultural experts review sensitive issues; 4. Establish continuous monitoring mechanisms: Regularly audit models to correct biases.

The caste bias audit reminds us that technological progress does not automatically bring social fairness. The inherent biases of large language models as information intermediaries may affect the real world. We need to move towards fair AI systems through systematic detection, in-depth analysis, and continuous improvement. This study provides methodological foundations and empirical evidence, which deserve attention from AI researchers and practitioners.