People Are Not Just Their Countries. Disentangling Social Determinants of LLM Value Alignment Across Europe

LLMs show unequal value alignment across socio-demographic groups: The research found that Large Language Models are not equally aligned with the values of different populations, with particularly significant disparities based on education level, income, occupation, and religion—meaning some groups' values are systematically better represented in these AI systems than others.

Country alone doesn't explain the full picture: While a person's country of residence does explain a substantial portion of differences in LLM alignment (comparable to all socio-demographic factors combined), this finding reveals that treating populations as monolithic national units masks important internal diversity within countries.

Multiple factors work together to shape alignment gaps: The study found that country-level factors and socio-demographic characteristics work together in complementary ways to explain value alignment patterns, suggesting that responsible AI development requires attention to both geographic and within-country social divisions.

Current LLM evaluation methods may be insufficient: By focusing primarily on country-level or cultural comparisons, existing research on LLM value alignment has overlooked how socio-demographic divides create disparities—indicating that more granular analysis is needed to identify whose values are being left out.

Education, income, and religion are key equity concerns: The research specifically highlights education, income, occupation, and religion as socio-demographic dimensions where LLM alignment varies significantly, suggesting these are priority areas for ensuring equitable AI systems.