AI Reveals Deep-Seated Caste Bias in India, Assigns Tasks by Surname

• GPT-4 assigned high-status professions to 'Bansal' (Brahmin) and low-status jobs to 'Ahirwar' (Dalit) based solely on surnames, reflecting societal caste bias.
• Sociologists confirm AI models learn and perpetuate existing societal biases, as caste identity persists in India.
• AI's structural bias, not just superficial responses, organizes human identities within its mathematical framework, influencing recruitment, credit, and healthcare.
• Studies like DECASTE and 'Discrimination in LLMs' show consistent stereotyping: upper-caste names linked to prestige, Dalit names to stigma and lower status.
• IndiCASA, developed by IIT Madras and UT Dallas, offers a new dataset to test for caste, religion, gender, disability, and socio-economic bias in LLMs, crucial for the Indian context.