VSM13 International Survey Insights

• The survey operationalizes Hofstede’s cultural dimensions using 24 Likert-scale items to produce normalized scores on Power Distance, Individualism, Masculinity, Uncertainty Avoidance, Long-Term Orientation, and Indulgence.
• It employs a weighted algorithm with normalization constants to enable robust cross-country comparisons and establish reliable international benchmarks.
• Recent adaptations extend its application to benchmarking AI cultural alignment, helping to diagnose biases in large language models across global cultural profiles.

The Values Survey Module 2013 (VSM13) International Survey is a standardized instrument developed for quantitatively comparing national cultures along core value dimensions identified by Geert Hofstede. Originally designed for cross-country analysis in organizational and societal contexts, VSM13 operationalizes culture through a 24-item survey, producing six primary dimension scores—Power Distance, Individualism, Masculinity, Uncertainty Avoidance, Long-Term Orientation, and Indulgence—on a normalized 0–100 scale. The module is recognized for its broad theoretical impact in cross-cultural psychology, international business, and, more recently, in the benchmarking of AI cultural alignment. VSM13 is referenced as a methodological standard in both human and machine alignment studies, notably for diagnosing cultural biases in LLMs and evaluating cross-national user experiences in technology and media contexts (Luther et al., 10 Dec 2025).

1. Theoretical Foundation and Dimensions

VSM13 formalizes Hofstede’s theory of national cultural value systems, which posits that enduring cultural dimensions underlie behavioral, attitudinal, and organizational differences across societies. The principal dimensions encoded in VSM13 (2013 revision) are:

1. Power Distance Index (PDI): The extent to which power differentials and hierarchical order are accepted by less powerful members of organizations.
2. Individualism vs. Collectivism (IDV): Preference for self-reliance and autonomy versus integration into cohesive in-groups.
3. Masculinity vs. Femininity (MAS): Societal emphasis on competition, achievement, and assertiveness (masculinity) versus cooperation, modesty, and care for the weak (femininity).
4. Uncertainty Avoidance Index (UAI): Degree of discomfort with ambiguity and preference for predictability and codified norms.
5. Long-Term vs. Short-Term Orientation (LTO): Focus on future-oriented virtues (perseverance, thrift) versus historical continuity and immediate compliance with tradition.
6. Indulgence vs. Restraint (IVR): Tendency to allow the gratification of basic human drives versus suppressing them through strict social norms.
7. Pragmatism vs. Normative (PRA): (Occasionally included) Contrasts pragmatic adaptation with strict adherence to norms.

These constructs are empirically defined via question batteries and validated with repeated international fieldings every 5–10 years (Luther et al., 10 Dec 2025).

2. Survey Design, Itemization, and Scoring

VSM13 consists of 24 five-point Likert-scale questions, segmented by thematic blocks mapping onto the six (or seven) value dimensions. Example items include frequency of subordinate dissent (PDI) and the valuation of independence (IDV).

Scoring is algorithmic, using a weighted sum of mean responses ($m_i$) to specified question pairs/subsets, with post-hoc normalization via dimension-specific constants to standardize outcomes over [0,100]:

$$\mathrm{PDI}=35\,(m_{07}-m_{02}) + 25\,(m_{20}-m_{23}) + C_{PDI}$$

$$\mathrm{IDV}=35\,(m_{04}-m_{01}) + 35\,(m_{09}-m_{06}) + C_{IDV}$$

$$\mathrm{MAS}=35\,(m_{05}-m_{03}) + 35\,(m_{08}-m_{10}) + C_{MAS}$$

$$\mathrm{UAI}=40\,(m_{18}-m_{15}) + 25\,(m_{21}-m_{24}) + C_{UAI}$$

$$\mathrm{LTO}=40\,(m_{13}-m_{14}) + 25\,(m_{19}-m_{22}) + C_{LTO}$$

$$\mathrm{IVR}=35\,(m_{12}-m_{11}) + 40\,(m_{17}-m_{16}) + C_{IVR}$$

where normalization constants are, for example: $C_{PDI}=15$, $C_{IDV}=11.5$, and so forth (Luther et al., 10 Dec 2025).

3. Standard and Novel Applications

VSM13 is methodologically mandated for international comparisons, requiring a minimum of 20 respondents per country and cross-national sampling of at least 10 countries to ensure statistical reliability. It has been systematized for organizational climate assessment, the study of technology acceptance, design evaluation, and—since 2025—benchmarking the cultural alignment of LLMs.

Recent research used VSM13 to quantify the “cultural fit” of LLM-generated responses, reconstructing survey protocols in both English and Simplified Chinese, and introducing cultural priming instructions to steer models toward U.S. or Chinese perspectives. This adaptation confirms the operational flexibility of the VSM13 instrument as both a human and human-mimetic measurement framework (Luther et al., 10 Dec 2025).

4. Implementation in AI Cultural Alignment Studies

The novel adaptation of VSM13 for LLM benchmarking involves prompt engineering: survey questions are reframed from personal introspection (“Do you …”) to descriptive proxies (“Does the average person …”). System prompts are introduced to simulate cultural context (“You are an average person from [country] answering a survey question”), and question banks are translated into the target language (e.g., Simplified Chinese).

Each LLM is subjected to the full 24-item battery across multiple experimental conditions:

• English/Chinese prompt language
• U.S./China/no cultural priming via system message

Multiple independent runs (e.g., 20 per model per condition at high temperature) are aggregated to compute model-specific dimension means. Alignment is then defined as the sum of absolute deviations from country benchmarks across all relevant dimensions. A misfit threshold (e.g., $\Delta \le 90$) demarcates “strong” alignment (Luther et al., 10 Dec 2025).

5. Empirical Benchmarks and Interpretation

Global means for the most referenced countries are as follows:

Dimension	United States	China
PDI	40	80
IDV	91	20
MAS	62	66
UAI	46	30
LTO	26	87
IVR	68	24

Empirical LLM alignments under optimal prompting reveal persistent WEIRD (Western, Educated, Industrialized, Rich, Democratic) profile biases in flagship models such as DeepSeek-V3, DeepSeek-V3.1, and GPT-5, with strong alignment to U.S. values but weak or no alignment to Chinese value profiles. Lower-cost models (GPT-4o, GPT-4.1) exhibit greater response malleability and can be moved toward target cultural profiles with prompt language and cultural prompting. However, even with these interventions, entrenched deviations remain in key dimensions, especially for models trained on large-scale Western-centric corpora (Luther et al., 10 Dec 2025).

Prompt language selection (English vs. Chinese) and explicit cultural cues improve proximity to benchmarks by 15–30% depending on the model, but the efficacy of these levers is model-dependent and bounded by pretraining biases.

6. Limitations, Controversies, and Analytical Caveats

While VSM13 is widely accepted for cross-cultural research, it is not immune to critiques regarding representativeness, stability of constructs across translation, and the risk that psychometric dimensions may not capture deeper or emergent cultural dynamics. In AI benchmarking, alignment as measured by VSM13 aggregates pointwise simulated response averages, but does not reflect deeper generative or inferential biases unless the models are structurally steered or explicitly fine-tuned on relevant cultural corpora.

This suggests that while VSM13 provides a standardized quantitative yardstick for both human societies and artificial agents, full cultural adaptation of LLMs will require interventions at the architectural and training-corpus levels, rather than solely superficial prompt alignment. (Luther et al., 10 Dec 2025)

7. Significance and Future Directions

VSM13 remains the reference instrument for rigorous, reproducible quantification of national culture in human and AI populations. Its application to LLM alignment exposes persistent Western value dominance in state-of-the-art models, elucidates the boundaries of prompt-based cultural adaptation, and foregrounds the need for more representative, globalized training paradigms as LLMs are deployed internationally.

A plausible implication is that ongoing cross-disciplinary research will increasingly treat survey-based culture modeling as an upstream component in the design, auditing, and regulation of both computational and organizational systems, leveraging instruments such as VSM13 for systematic calibration and compliance in global contexts (Luther et al., 10 Dec 2025).