Cultural Consensus Theory reveals LLMs homogenize norms across 10 countries

A new paper from Krishna Pothugunta and John P. Lalor, accepted to ACL Findings 2026 and posted on arXiv (2608.09937), tackles a growing problem in AI alignment: LLMs often represent culture as a single, normalized average rather than the diverse, consensus-driven structures seen in real societies. The researchers brought in Cultural Consensus Theory (CCT) from cultural anthropology to model how beliefs are shared within and across groups. They applied CCT to the World Values Survey across 10 countries and 12 domains, revealing a consistent failure mode: models either failed to produce a coherent cultural consensus, treating values as scattered and contradictory, or they over-regularized consensus, smoothing away meaningful intra-group differences until everyone appears to hold identical views.

Both failure modes have practical consequences. When LLMs over-regularize, they amplify algorithmic homogenization—making minority or subcultural perspectives invisible, even when those perspectives are statistically significant in the underlying data. When they fail to form consensus, the model becomes unreliable for tasks requiring culturally grounded reasoning, such as multilingual customer support, policy analysis, or personalized content generation. The authors argue that CCT provides a principled way to expose these issues: by explicitly representing intra-group variance, researchers can generate diagnostics that show where an LLM's cultural output diverges from human diversity. This shifts the evaluation focus from crude country-level averages to the actual structure of cultural beliefs. The paper is currently available on arXiv with a DOI from DataCite, and the authors have linked to code and data via Hugging Face, Replicate, and related tools, making the framework directly usable by NLP researchers and AI engineers.