simulating_identity_propagating_bias_emnlp2025

This project investigates whether persona-prompting in large language models affects the level of linguistic abstraction, a known marker of stereotyping. Using the Linguistic Expectancy Bias framework, we analyze outputs from six open-weight LLMs across three prompting conditions, comparing persona-driven responses to those of a generic AI assistant. To support this study, we release Self-Stereo, a dataset of self-reported stereotypes collected from Reddit. Abstraction is measured using concreteness, specificity, and negation. Results show that persona-prompting has limited impact on linguistic abstraction, underscoring concerns about the ecological validity of personas as proxies for socio-demographic groups and the risks of reinforcing stereotypes when modeling marginalized voices.