Liu, Kaneko & Chu (2026) find LLMs' unprompted opinion distributions align most closely with the most recent WVS wave, among countries pre-selected as well-aligned

llm-evaluationcultural-awarenessworld-values-surveypromptingarxiv

Liu, Kaneko & Chu, "On the Alignment of Large Language Models with Global Human Opinion" (AAAI 2026), test a model's unprompted — no year, no survey date given — opinion distribution against World Values Survey waves 5–7, restricted to a small set of countries pre-selected because the model is already well aligned with them ("to ensure that we can filter out at least five countries that are well aligned by LLMs"). Within that filtered set, the correlation between a model's unprompted distribution and each WVS wave is strongest for the most recent wave: "We can see that LLMs most align with the latest wave of human opinions."

This is a third, independently-run data point converging with ChronoCulture and Daryani, Bogen & Daepp on the same general shape: left to their own devices, these models default to a present-skewed representation of a country's values rather than the specific historical moment a user might intend. But the method never prompts a model with an explicit older survey year, and it does not stratify by the size of the real shift between waves — the two features that define ChronoCulture's older-year-naming asymmetry. It therefore does not test, confirm, or disconfirm that specific conditional effect; see question-corroborate-chronoculture-older-year-naming-asymmetry. The country pre-selection (already-aligned countries only) is also a scope limit on how far "LLMs default to the present" generalizes from this paper alone.