Insidious by Design: Implications of Large Language Model algorithmic bias for the Global South

Subtle biases in LLMs reinforce global power imbalances: The study found that major LLMs (ChatGPT, Claude, Grok, and Copilot) embed racial hierarchies and Western-centered knowledge systems in ways that are not obvious errors but rather "subtly embedded in narrative structure and emotional template"—for example, portraying Black characters facing hardships while white characters demonstrate agency.

Gender stereotypes are systematized across models: When prompted with different names, LLMs consistently generated narratives where women were depicted with "rich interior lives" while men were shown "making plans," revealing how gender asymmetries become baked into the models' default storytelling patterns.

LLM explanations of global economics exclude Southern perspectives: The models failed to incorporate explanations and knowledge systems from the Global South when discussing development and economic issues, instead reproducing Western-centric frameworks as if they were universal truths.

These biases are dangerous because they appear credible: The paper emphasizes that LLM outputs "perform plausibility while reproducing dominance"—meaning biased content sounds authoritative and reasonable, making it more likely to be accepted and spread without critical examination.

Institutions need critical evaluation before adoption: Rather than unreflectively implementing LLMs, universities and organizations should conduct structural critiques examining "whose knowledge systems are reproduced and legitimated, or marginalised and undermined" by these technologies.