Reproduce “Clustering Discourses: Racial Biases in Short Stories about Women Generated by Large Language Models” on OpenResearch

About this paper

Gustavo Bonil, João Gondim, Marina dos Santos, Simone Hashiguti, Helena Maia, Nadia Silva, Helio Pedrini, Sandra Avila

Researchers from UNICAMP and UFG developed a mixed-methodology, combining computational clustering with qualitative discourse analysis, to investigate racial biases in short stories about women generated by Llama-3.2-3B-Instruct in Portuguese. Their analysis revealed distinct narrative patterns, with Black women predominantly featured in stories of social overcoming or ancestral mythification, and white women in narratives of subjective self-realization and artistic sensitivity.

Key insights

• LLMs consistently generate racially distinct narrative patterns, confining Black women predominantly to themes of collective struggle/social overcoming or ancestral mythification, while white women are afforded more varied narratives of subjective self-realization.
• These patterns, even when seemingly empowering, can be limiting, as they narrow the representational possibilities for Black women compared to the broader, more introspective roles given to white women.
• The integrated mixed-methodology effectively scales qualitative inquiry into large datasets, providing both computational evidence of distinction and nuanced discursive interpretations of bias.

Summary by alphaXiv.