Whose Values Prevail? Bias in Large Language Model Value Alignment

As large language models (LLMs) are increasingly integrated into our lives, concerns have been raised about whether they are biased towards the values of particular cultures. We show that while LLMs were biased toward the values of WEIRD populations, some non-Western populations, including East Asia and Russia, were also represented relatively well. Notably, the Rich dimension was the strongest predictor of LLM's alignment instead of the most discussed Western dimension. This suggests the need to attend to less prosperous populations instead of focusing only on easily accessible populations. We also found that one source of this bias could be unbalanced training data as approximated by an Internet Freedom measure, and that prompting the model to act as individuals from different populations reduced the bias but could not eliminate it. These findings raise the importance of training process disclosure and the consideration of culture-specific models to ensure ethical usage of LLMs.