Mechanisms of Social Bias in Large Language Models: Disentangling Linguistic Framing from Group Specification

An investigation of whether social bias in large language models is driven by the linguistic framing of a prompt or by the explicit specification of a social group.

This project investigates the mechanisms behind social bias in large language models, separating the effect of a prompt’s linguistic framing from the effect of explicitly specifying a social group. The goal is to better understand when and why model outputs diverge across groups.

Authors: Vishwanath Emani Venkata, Silvia Téliz

Presented at IC2S2 2026 (International Conference on Computational Social Science), University of Vermont, Burlington, VT, July 28–31, 2026.