Can Persona-Prompted LLMs Emulate Subgroup Values? An Empirical Analysis of Generalisability and Fairness in Cultural Alignment

A research team from Singapore has published a groundbreaking study examining whether Large Language Models (LLMs) can accurately emulate the cultural values of specific demographic subgroups. Using Singapore as a case study and data from the World Values Survey, the researchers constructed a dataset of over 20,000 samples to test models including GPT-4.1. Their findings reveal that even state-of-the-art models achieve only 57.4% accuracy in predicting subgroup modal preferences, highlighting a significant gap in fine-grained cultural alignment.

While simple fine-tuning on structured numerical preferences yielded substantial gains—improving accuracy on unseen subgroups by an average of 17.4%—the research uncovered troubling biases. Models demonstrated significantly better performance when emulating young, male, Chinese, and Christian personas compared to other demographic groups. Furthermore, while fine-tuning improved average performance metrics, it actually widened disparities between subgroups when measured using distance-aware fairness metrics, creating a fairness-performance tradeoff.

The study provides crucial insights into the limits of current cultural alignment techniques and their real-world implications. As AI systems are increasingly deployed globally, their inability to accurately represent diverse cultural perspectives—and their tendency to amplify existing biases—poses serious challenges for equitable AI development. This research establishes important benchmarks for evaluating cultural alignment and highlights the need for more sophisticated approaches to ensure AI systems respect and represent global diversity.