Scholarlink

<p>Culture is fundamental to individual identity and serves as a backbone for unity, communal progress, and social harmony. Given the widespread adoption of artificial intelligence (AI) tools, such as AI-driven Question Answering (QA) systems, it is essential that these systems respond to cultural queries in an unbiased manner. This paper provides a systematic study of existing bias detection works (published in the last five years) with respect to cultural informatics. This work is unique as it is the first combined survey to cover bias and fairness detection in both large language model (LLM)-centric and vision language model (VLM)-centric studies, and the first survey to focus on South Asian festivals and their representation in these models. This survey combines a <i>statistical objective survey</i> (SoS) and a <i>subjective survey</i>. The <i>SoS</i> analyzes research publications from 2021–2025, revealing a robust upward trend in the research studies focused on bias and fairness in both VLMs and LLMs. The <i>subjective survey</i> explores existing literature across three dimensions: models, metrics, and benchmark datasets used in ethical cultural informatics. This work will be helpful for readers who want to work on building AI systems that can integrate cultural sensitivity and reduce bias for fairness in responsible AI design, with specific reference to the multicultural and festival-oriented content context with respect to South Asian countries.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Bias and fairness in vision–language models and large language models: a survey on South Asian festival representations

  • 期刊论文

  • 发表日期：

    2026年6月2日

查看全文
  • Nitumani Sarmah,
  • Rupam Bhattacharyya

摘要

Culture is fundamental to individual identity and serves as a backbone for unity, communal progress, and social harmony. Given the widespread adoption of artificial intelligence (AI) tools, such as AI-driven Question Answering (QA) systems, it is essential that these systems respond to cultural queries in an unbiased manner. This paper provides a systematic study of existing bias detection works (published in the last five years) with respect to cultural informatics. This work is unique as it is the first combined survey to cover bias and fairness detection in both large language model (LLM)-centric and vision language model (VLM)-centric studies, and the first survey to focus on South Asian festivals and their representation in these models. This survey combines a statistical objective survey (SoS) and a subjective survey. The SoS analyzes research publications from 2021–2025, revealing a robust upward trend in the research studies focused on bias and fairness in both VLMs and LLMs. The subjective survey explores existing literature across three dimensions: models, metrics, and benchmark datasets used in ethical cultural informatics. This work will be helpful for readers who want to work on building AI systems that can integrate cultural sensitivity and reduce bias for fairness in responsible AI design, with specific reference to the multicultural and festival-oriented content context with respect to South Asian countries.