LLMs show transformed, not transferred, bias across English and Swahili

A new research paper analyzes the cross-lingual bias present in large language models like GPT-5.2 and Gemini 2.5 Flash. By submitting symmetric English and Swahili prompt pairs, the study found that biases transform rather than transfer between languages, with significant shifts in stereotype rates and refusal behaviors. The research highlights that English-centric bias audits are insufficient for ensuring equitable performance in multilingual LLM deployments. AI

IMPACT Highlights the inadequacy of English-only bias audits for multilingual LLM deployment, necessitating new evaluation strategies.

RANK_REASON Research paper analyzing bias in LLMs across languages. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• alphaXiv
• arXiv
• CatalyzeX
• Connected Papers
• DagsHub
• English
• Gemini 2.5 Flash
• Gotit.pub
• GPT-5.2
• Hugging Face
• Litmaps
• ScienceCast
• Swahili

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →