A Comparative Study of Cultural Hallucination in Large Language Models on Culturally Specific Ethical Questions

Abstract Rapid advancements in natural language processing have led to the development of highly sophisticated models capable of generating human-like text, yet challenges remain in ensuring that these models produce culturally accurate and ethically consistent responses. The novel concept of this study lies in the comprehensive evaluation of ChatGPT 4o and Gemini 1.5 Flash on culturally specific ethical questions, providing a detailed comparison of their performance across diverse cultural contexts. Automated evaluation metrics, including semantic similarity, cultural relevance, and ethical consistency, were employed to assess the models' capabilities, revealing significant insights into their strengths and limitations. The results indicated that while both models exhibit high cultural relevance and ethical consistency, notable differences in their performance across various regions suggest areas for further improvement. Statistical analysis confirmed the significance of these differences, emphasizing the necessity for ongoing refinement of training methodologies. The study demonstrates the importance of integrating deeper cultural insights and ethical frameworks into model development, contributing valuable knowledge to the field of AI ethics and cultural competence.