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

2025/11/03 by Muhammed Saeed, Saeed, Muhammed, Muhammad Abdul-mageed +3 · 1 voice
Computer Science · #Computation and Language (cs.CL) #Computers and Society (cs.CY) #FOS: Computer and information sciences #cs.CL #cs.CY

paper · pdf · doi:10.48550/arxiv.2511.01187

arxiv published 2025/11/03 · arxiv updated 2026/03/30

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 (≥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.

Citations

• The Stochastic Parrot on LLM's Shoulder: A Summative Assessment of Physical Concept Understanding
• Who is Responsible? The Data, Models, Users or Regulations? A Comprehensive Survey on Responsible Generative AI for a Sustainable Future
• OASIS Uncovers: High-Quality T2I Models, Same Old Stereotypes
• Large Language Model Safety: A Holistic Survey
• Multi-OphthaLingua: A Multilingual Benchmark for Assessing and Debiasing LLM Ophthalmological QA in LMICs
• Desert Camels and Oil Sheikhs: Arab-Centric Red Teaming of Frontier LLMs
• Exploring Bias and Prediction Metrics to Characterise the Fairness of Machine Learning for Equity-Centered Public Health Decision-Making: A Narrative Review
• The Llama 3 Herd of Models
• Does Refusal Training in LLMs Generalize to the Past Tense?
• Implicit Discourse Relation Classification For Nigerian Pidgin
• DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model
• Modeling Orthographic Variation Improves NLP Performance for Nigerian Pidgin
• IndiBias: A Benchmark Dataset to Measure Social Biases in Language Models for Indian Context
• The Language Barrier: Dissecting Safety Challenges of LLMs in Multilingual Contexts
• Large Language Models
• Low-Resource Cross-Lingual Adaptive Training for Nigerian Pidgin
• Direct Preference Optimization: Your Language Model is Secretly a Reward Model
• Training language models to follow instructions with human feedback
• BBQ: A Hand-Built Bias Benchmark for Question Answering
• Societal Biases in Language Generation: Progress and Challenges
• Persistent Anti-Muslim Bias in Large Language Models
• CrowS-Pairs: A Challenge Dataset for Measuring Social Biases in Masked Language Models
• Language (Technology) is Power: A Critical Survey of "Bias" in NLP
• StereoSet: Measuring stereotypical bias in pretrained language models
• AraBERT: Transformer-based Model for Arabic Language Understanding
• Unsupervised Cross-lingual Representation Learning at Scale
• A Survey on Bias and Fairness in Machine Learning
• Gender Bias in Coreference Resolution: Evaluation and Debiasing Methods

Discussions

• Big news! 🚀 Our paper on LLM bias is accepted in LREC 2026 in Palma. We move beyond simple tests to a multilingual debate benchmark (7 languages, 8.4k prompts) to see how AI actually frames the world [bsky, 3 points, 0 comments]

Related

• A Multifaceted Analysis of Social Biases in Large Language Models
• The Alignment Veto: How Safety Training Suppresses Cultural Knowledge in LLMs
• A Dual-Layered Evaluation of Geopolitical and Cultural Bias in LLMs
• With a Grain of SALT: Are LLMs Fair Across Social Dimensions?
• Beyond Metrics: Evaluating LLMs' Effectiveness in Culturally Nuanced, Low-Resource Real-World Scenarios