Surfacing Subtle Stereotypes: A Multilingual, Debate-Oriented Evaluation of Modern LLMs

TL;DR

This paper introduces a multilingual debate-style benchmark to evaluate how large language models reproduce stereotypes across different languages and sensitive domains, revealing biases that persist despite safety efforts.

Contribution

It presents extsc{DebateBias}, a new multilingual benchmark for assessing stereotype emergence in LLMs, highlighting biases in low-resource languages and across diverse cultural contexts.

Findings

Models reproduce entrenched stereotypes despite safety measures.

Biases are more pronounced in low-resource languages.

Current alignment methods do not fully mitigate biases in open-ended generation.

Abstract

Large language models (LLMs) are widely deployed for open-ended communication, yet most bias evaluations still rely on English, classification-style tasks. We introduce \corpusname, a new multilingual, debate-style benchmark designed to reveal how narrative bias appears in realistic generative settings. Our dataset includes 8{,}400 structured debate prompts spanning four sensitive domains -- Women's Rights, Backwardness, Terrorism, and Religion -- across seven languages ranging from high-resource (English, Chinese) to low-resource (Swahili, Nigerian Pidgin). Using four flagship models (GPT-4o, Claude~3.5~Haiku, DeepSeek-Chat, and LLaMA-3-70B), we generate over 100{,}000 debate responses and automatically classify which demographic groups are assigned stereotyped versus modern roles. Results show that all models reproduce entrenched stereotypes despite safety alignment: Arabs are overwhelmingly linked to Terrorism and Religion ($\geq$89\%), Africans to socioeconomic ``backwardness'' (up to 77\%), and Western groups are consistently framed as modern or progressive. Biases grow sharply in lower-resource languages, revealing that alignment trained primarily in English does not generalize globally. Our findings highlight a persistent divide in multilingual fairness: current alignment methods reduce explicit toxicity but fail to prevent biased outputs in open-ended contexts. We release our \corpusname benchmark and analysis framework to support the next generation of multilingual bias evaluation and safer, culturally inclusive model alignment.