A cross-lingual syntactic investigation of gender bias and stereotyping in GPT-4o: English vs Hindi

Beneath the vast capabilities and potentials of AI systems lie some crucial predicaments. Though Large Language Models (LLMs) such as ChatGPT have proven to be transformative, the ingrained gender biases and stereotypes in such models are a societal concern. The propagation of such biases is detrimental to men, women and all gender-diverse groups alike. Addressing these issues is imperative for ensuring equity and inclusivity. Thus, there is a need to identify and analyze such biases in LLMs and drift these models towards Responsible AI. In this research, we endeavor to investigate gender biases and occupation-based stereotyping in ChatGPT (GPT-4o), focusing on a comparative analysis between English and Hindi language responses. The dataset curated for this analysis comprises ChatGPT’s responses to selected gender-neutral prompts. The gender-determining syntactic structure of languages is employed as a metric for bias determination. In the English language, pronouns are the gender-determining part of speech, whereas, in Hindi, verb conjugations determine the gender of the subject. This formed the foundation of the gender bias identification in this study. We observe that ChatGPT tends to incline towards yielding male nouns in both languages. The biases have a higher degree of being male-skewed in English than in Hindi. In addition, the investigation further confirms that ChatGPT harbours occupation-based stereotypes. These biases and stereotypes do not necessarily depict the present disparities in society, indicating that ChatGPT reflects the biases of its training data. Conclusively, this research serves as a foundation for the identification of AI-generated biases and, subsequently, their annihilation.