Yet another algorithmic bias: A Discursive Analysis of Large Language Models Reinforcing Dominant Discourses on Gender and Race

LLMs reproduce subtle racial and gender stereotypes that quantitative methods alone miss: The study found that Black women were consistently portrayed as tied to ancestry and resistance, while white women appeared in self-discovery narratives. These patterns emerged through qualitative analysis of AI-generated stories, demonstrating that automated bias detection tools overlook nuanced ways discrimination manifests in natural language.

Current bias correction attempts by LLMs are often superficial: When researchers prompted the models to correct biases, the systems made surface-level revisions while maintaining the same problematic underlying meanings. This suggests that simply asking AI systems to "be less biased" does not address the root ideological issues embedded in their training data.

Language models reinforce social immobility and essentialization: The research shows that LLMs perpetuate "crystallized discursive representations"—essentially freezing certain groups into narrow, stereotypical roles. Black women were locked into narratives about their past and struggle, while white women were granted agency and personal growth, reflecting and reinforcing real-world inequalities.

Qualitative analysis is essential for responsible AI development: The paper argues that manual, discursive analysis should complement automated bias detection methods. This human-centered approach helps developers and users understand how and why biases appear in AI outputs, enabling more targeted and meaningful mitigation strategies than numbers alone can provide.