Original title: Aligned but Flattened: Analyzing the Trade-off between Cultural Alignment and Diversity in LLMs
Jingshen Zhang, Shaoyang Xu, Wenxuan Zhang · September 2026
Teaching large language models to respect different cultures seems like a great idea. Usually, engineers train these systems to match cultural viewpoints by targeting a single alignment score. However, this approach creates a major problem. Researchers tested six mainstream models using the World Values Survey to see if they truly understand diverse people. Instead of capturing real cultural depth, the models simply clung to the most dominant majority opinions. Trying to maximize alignment wiped out the wide variety of beliefs found inside actual human communities. The models collapsed into flat, repetitive answers that ignored minority voices. A deeper look inside the systems suggests that this loss of diversity is tied directly to how neural network math naturally optimizes itself. In short, today's training methods flatten cultures instead of reflecting them.
Imagine a group of friends deciding where to eat. Each friend has different tastes, but one friend always picks the most popular restaurant. Soon the group forgets the quieter preferences and always ends up at the same place. In this paper, the AI similarly learns to favor the majority opinion in a culture, losing the variety of viewpoints that exists among real people.
When AI learns to represent a culture, it usually just mirrors the most common opinions, losing the wide range of real human views. This matters for anyone using AI for education, journalism, or any task where understanding nuance is key. If AI flattens cultural diversity, it risks silencing minority voices and giving a false picture of what people actually think.
The evaluation covers only six LLMs and the World Values Survey, so it may not capture all cultural dimensions. The mechanistic link to low-rank bias is suggested through analysis, not proven causally. The findings apply specifically to the cultural fine-tuning paradigm studied, not necessarily to other alignment methods.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Aligned but Flattened: Analyzing the Trade-off between Cultural Alignment and Diversity in LLMs — Jingshen Zhang, Shaoyang Xu, Wenxuan Zhang
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.