Study reveals LLMs subtly encode racial and gender biases against Global South

A small-scale exploratory study by Sioux McKenna and Nompilo Tshuma, published on arXiv, examines how Large Language Models (LLMs) like ChatGPT, Claude, Grok, and Copilot reproduce racial hierarchies, gender asymmetries, and Western-centric epistemic frameworks. The researchers submitted identical prompts to each model: first, asking for stories using names suggestive of specific racial and gender communities; second, asking questions about 'development.' Drawing on critical AI scholarship and postcolonial theory, they argue that biases operate 'insidiously'—below the threshold of obvious error or overt prejudice. Instead, they are subtly embedded in narrative structure and emotional templates. Examples include women being depicted with rich interior lives while men make plans; Black characters facing hardships while white characters navigate with agency; and explanations of economic world order ignoring Southern perspectives.

The study concludes that university adoption of LLMs must be accompanied by structural critique rather than unreflective acceptance. Critical AI literacy should engage seriously with whose knowledge systems are reproduced and legitimated, and which are marginalized. The findings highlight that LLMs perform plausibility while reproducing dominance, making bias harder to detect and correct. The authors emphasize the need for educational institutions to develop frameworks that question the epistemic foundations of these technologies, especially given their growing influence in the Global South.