🤖 AI

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

Using data from the European Social Survey, this study reveals that while socio-demographic factors significantly influence Large Language Model value alignment across Europe, a respondent's country of residence remains a distinct and equally powerful predictor, with both dimensions offering complementary insights into alignment disparities.

Original authors: Maria-Louisa Wightman, Guillaume Bied, Tijl De Bie

Published 2026-08-10

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Maria-Louisa Wightman, Guillaume Bied, Tijl De Bie

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a world where you have a super-smart, all-knowing robot friend who can answer any question, tell you a joke, or help you write a story. This is what Large Language Models (LLMs) are like today. They are the engines behind the chatbots and AI tools we use every day. But here's the catch: these robots aren't born with their own opinions. They learn everything by reading a massive library of books, websites, and conversations written by humans. Because of this, they start to sound like the people who wrote those words.

Now, think about "values" as the invisible rules we all live by—like what we think is fair, how we treat our neighbors, or what we believe is important in life. If a robot learns mostly from people in one specific type of neighborhood, it might think that neighborhood's rules are the only rules. This is a big deal because we are starting to ask these robots for advice on everything from school projects to serious life choices. If the robot's "moral compass" is tilted toward only one group of people, it might accidentally ignore or misunderstand everyone else. Scientists are trying to figure out: Does this robot friend actually understand you, or does it only understand a specific version of "you" based on where you live or how much money your family makes?

The Great Robot Identity Crisis

In this study, a team of researchers decided to put 10 of the most popular AI chatbots to the test. They didn't just ask the robots to guess; they gave them a giant, real-world survey called the European Social Survey. This survey was filled with questions about what people in 29 European countries actually think and feel about life, money, religion, and politics. The researchers then compared the robots' answers to the answers of over 50,000 real humans.

Think of it like a massive game of "Guess Who?" played across a whole continent. The researchers wanted to see if the robots' answers matched the humans' answers, and if so, which humans they matched best. Did the robots sound more like a rich, university-educated person living in a big city in Sweden? Or did they sound more like a retired farmer in a small village in Bulgaria?

The Robots Have a Favorite Audience

The results were pretty clear: the robots definitely have a favorite audience. The study found that the AI models were much better at "aligning" with (meaning, agreeing with) people who were richer, more educated, and lived in Western European countries.

Imagine a party where the DJ (the AI) is playing music. The DJ seems to know exactly what the wealthy, well-educated guests in the VIP section want to hear. But when the DJ tries to play songs for the guests in the back of the room—people with lower incomes, less formal education, or those from Eastern or Southern Europe—the music feels a bit off. The robots were consistently less accurate when trying to understand the values of these groups.

The study also looked at religion. It turned out that the robots were much better at understanding people who weren't religious or who belonged to certain Christian groups. However, they struggled significantly with people who identified as Muslim or Eastern Orthodox. It's as if the robot's "cultural dictionary" was missing a few key pages for those specific communities.

It's Not Just About Where You Live

Here is the twist that surprised the researchers. For a long time, people assumed that if you just looked at the country someone lived in, you could guess what the robot would think about them. You might have thought, "Oh, if I'm from France, the robot will think like a French person."

But the researchers dug deeper. They asked: Is the robot just matching the country, or is it matching the type of person living there? To find out, they used a clever statistical trick (called "inverse propensity weighting") to pretend that every country had the exact same mix of rich and poor, educated and less educated people. If the country was the only thing that mattered, the differences between countries should have disappeared after this trick.

But they didn't. Even when the researchers made the countries look identical on paper, the robots still treated people from different countries differently. This means that where you live matters just as much as who you are. The robot isn't just looking at your income or your job; it's also looking at your passport. The country you live in and your personal background work together like a double-lock system to determine how well the robot understands you.

The "WEIRD" Problem

The study confirms a pattern scientists call "WEIRD," which stands for Western, Educated, Industrialized, Rich, and Democratic. The robots seem to be built by and for people who fit this description. Even within Europe, the robots were most accurate for people who fit the "WEIRD" mold.

The researchers also tested two different types of questions. One set was about broad, everyday opinions (like "Do you trust your government?"). The other set was about deep, abstract human values (like "Is it important to be humble?"). They found that for the broad questions, the country you lived in explained a huge chunk of the difference in how the robot answered. But for the deep, abstract questions, the country mattered less, and your personal background (like your education) mattered more. This suggests that how we ask the questions changes what the robot learns.

What This Means for You

The big takeaway isn't that the robots are "broken" or that they are bad. It's that they are mirrors. They reflect the world they were trained on. If that world is mostly made up of voices from wealthy, Western, educated backgrounds, then the robot will sound like that.

The researchers found that if you are a student with a master's degree living in a big city in Germany, the robot will likely understand you perfectly. But if you are a factory worker in a rural village in a different part of Europe, the robot might miss the mark. This isn't just a small glitch; it means that as we rely more on AI for advice, we might be accidentally reinforcing the values of the people who are already the most powerful, while leaving everyone else behind.

The paper doesn't say we should stop using AI. Instead, it suggests that we need to be much more careful about how we build and test these tools. We can't just say, "The robot is aligned with Europe." We have to ask, "Aligned with which Europeans?" Because, as this study shows, people are not just their countries, and robots shouldn't treat them as if they are.

Technical Summary: People Are Not Just Their Countries

Problem Statement

As Large Language Models (LLMs) become primary sources of information and advice, understanding their value alignment with human populations is critical. Existing literature on LLM value alignment predominantly relies on large-scale surveys (e.g., World Values Survey) and focuses on cross-national or cross-cultural comparisons. This approach often treats countries as monolithic units, neglecting the role of socio-demographic divides (such as education, income, and religion) in creating value alignment disparities. Furthermore, the repeated use of established surveys like the World Values Survey (WVS) raises concerns regarding data contamination, as LLMs may have memorized these specific datasets. There is a lack of research disentangling the contributions of country of residence versus individual-level socio-demographics in explaining value alignment patterns, particularly within a European context.

Methodology

The study addresses these gaps by leveraging the 11th wave of the European Social Survey (ESS), conducted between 2023 and 2024 across 29 European countries and Israel.

Data Source: The analysis utilizes 50,116 respondents from the ESS, incorporating 47 value-related survey questions (excluding country-specific items). A subset of 21 questions corresponds to the Portrait Value Questionnaire (PVQ) based on Schwartz's theory of basic human values.
Models: The study evaluates 10 prominent commercial LLMs from four providers: OpenAI (GPT-5 variants), Anthropic (Claude Opus/Sonnet variants), Mistral (various reasoning efforts), and DeepSeek (V3/V4).
Prompting Strategy: Models were prompted 20 times per question in English using a standardized schema. Responses were mapped to Likert scale categories. To ensure robustness, answers were aggregated via majority vote across model calls.
Alignment Score Definition: An alignment score ( $A_{p,m,q}$ ) is calculated for each person $p$ and model $m$ on question $q$ as the normalized distance between the human response and the model's majority vote on the Likert scale, ranging from 0 to 1.
Analytical Approach:
1. Aggregate Deviation: Calculation of cross-model mean deviations ( $d_{G,M,Q}$ ) for socio-demographic groups and countries relative to the population mean.
2. Inverse Propensity Weighting (IPW): Reweighting country samples to match the pooled ESS socio-demographic distribution to test if country-level differences are merely artifacts of demographic composition.
3. Variance Decomposition: Using Ordinary Least Squares (OLS) and Gradient Boosted Models (XGBoost) to predict individual alignment scores. The study compares the $R^2$ variance explained by: (i) country of residence only, (ii) socio-demographics only, and (iii) the combination of both.

Key Results

1. Socio-Demographic Disparities

LLMs exhibit unequal alignment across socio-demographic groups. The analysis reveals that alignment is significantly higher for groups with:

Higher Socio-Economic Status: Individuals with higher education levels, higher income deciles, and those in white-collar occupations show better alignment.
Western Backgrounds: Respondents with Western migration backgrounds align better than those with non-Western backgrounds.
Secularism: Non-religious individuals and those with lower religiosity scores align better than religious individuals, particularly Muslims and Eastern Orthodox respondents, who show the most negative deviations.
Political Interest: Higher political interest correlates with higher alignment.
Generational Trends: A U-shaped pattern emerges where the youngest and oldest cohorts align less well than middle-aged groups, though this varies by model.

2. Country-Level Effects

Geographic Variation: Significant disparities exist between countries. Scandinavian and Central European countries show the highest alignment, while Balkan and Baltic countries show the lowest. The difference between the highest and lowest aligning countries (Sweden vs. Bulgaria) is 0.0896, exceeding the variation found within any single socio-demographic factor.
Independence from Composition: Inverse Propensity Weighting analysis demonstrates that reweighting countries to have identical socio-demographic compositions does not significantly reduce the variance in country-level alignment scores. This indicates that country of residence explains alignment disparities independently of the socio-demographic makeup of the population.

3. Variance Decomposition

Complementary Factors: Both country of residence and socio-demographics are complementary in explaining value alignment. The combination of both factors yields the highest explanatory power.
Relative Weight: For the full set of questions, country of residence alone explains a substantial amount of variance (ranging from ~7% to ~28% depending on the model), which is on par with the full set of 15 socio-demographic variables.
Question Set Sensitivity: The explanatory power of country of residence varies by question type. For the broader question set ( $Q$ ), country is a strong predictor. However, for the abstract PVQ questions, the variance explained by country shrinks notably for most models, suggesting that abstract values are more driven by individual demographics, while broader stances are more influenced by national context (political climate/media).

Significance and Contributions

The paper claims three primary contributions:

First Cross-National Socio-Demographic Evaluation: It provides the first evaluation of value alignment across socio-demographics using the ESS, offering an alternative to the potentially contaminated WVS.
Evidence of WEIRD Reproduction: The study confirms that LLMs reproduce global patterns of alignment to "WEIRD" (Western, Educated, Industrialized, Rich, Democratic) populations. Specifically, within Europe, LLMs are better aligned with richer, more educated, and more Western individuals, potentially reinforcing the values of already privileged populations.
Disentangling Geography and Demographics: The research establishes that country of residence is a distinct and critical variable in alignment research. It cannot be replaced by socio-demographics alone, nor can it be fully explained by them. The two factors are complementary, with their relative importance shifting based on the specific definition of "values" being measured (abstract vs. practical).

The authors conclude that while socio-demographics are crucial, ignoring the specific role of national context leads to an incomplete understanding of value alignment disparities. They emphasize the need for value-sensitive design that accounts for these intersecting layers of inequality.