A Multifaceted Analysis of Social Biases in Large Language Models

A new study finds that four widely used large language models continue to exhibit social biases across politics, ideology, geopolitical alliance, language, and gender, despite alignment efforts intended to make them neutral and impartial [1]. The research, submitted by Xulang Zhang and revised through June 2026, examined models identified as Qwen, DeepSeek, Gemini, and GPT [3]. The authors probed political neutrality through news summarization, ideological biases via news stance classification, and geopolitical leanings by simulating United Nations General Assembly voting patterns [2]. Language bias was assessed using multilingual story completion, while gender affinities were measured against responses to the World Values Survey [2]. Results showed that although the models are aligned to be neutral, they still display distinct biases and affinities [1]. The gender analysis revealed that GPT showed a significant affinity for women’s values, while the other models were mostly gender neutral but demonstrated a slight tendency to align more with women’s values than men’s [3]. The study also noted that Qwen and DeepSeek gave answers reflecting contracting values, indicating a lack of a firm stance and a determinate sense of value [3]. The findings arrive as the broader research community continues to formalize how social bias is defined and measured in language models. A 2024 survey in Computational Linguistics consolidated definitions of social harm and proposed taxonomies for bias evaluation metrics and datasets, distinguishing between embedding-based, probability-based, and generated-text-based approaches [5]. That work also categorized mitigation techniques by intervention stage, including pre-processing, in-training, and post-processing methods [5]. The new multifaceted analysis adds empirical weight to concerns that alignment techniques alone do not eliminate underlying model predispositions [1][3].