IndiCASA: A Dataset and Bias Evaluation Framework in LLMs Using Contrastive Embedding Similarity in the Indian Context

All major LLMs tested showed measurable bias across multiple demographic categories, with no model achieving bias-free performance, indicating that current large language models carry embedded stereotypes that could affect their outputs in real-world applications.

Disability-related biases proved particularly resistant to current debiasing efforts, suggesting that marginalized groups beyond the most commonly addressed categories (gender, race) require targeted attention in AI safety and fairness work.

Existing bias evaluation methods fail to capture culturally specific stereotypes, as demonstrated by the need for IndiCASA's 2,575 context-specific sentences across five demographic axes (caste, gender, religion, disability, socioeconomic status), highlighting that one-size-fits-all bias testing is insufficient for global AI deployment.

Religion-related biases were notably lower across models, likely due to global debiasing efforts, which provides evidence that intentional fairness interventions can work—but also shows that some demographic categories receive more attention than others in AI development.

High-stakes applications require rigorous, culturally-informed bias evaluation before deployment, as the paper emphasizes the necessity of testing LLMs in specific cultural contexts before using them in critical domains where biased outputs could cause real harm to vulnerable populations.