LLM-GLOBE: A Benchmark Evaluating the Cultural Values Embedded in LLM Output

A risk in the widespread adoption of LLMs is cultural encapsulation, defined as the tendency to offer broad solutions and responses that are not adapted to individual and cultural differences [Wrenn, 1962]; failure to fully understand and appreciate cultural differences can lead to misunderstanding of the challenges individuals from various cultural background face and a lack of appropriate solutions. Just as cultural encapsulation affects existing systems and our ability to provide effective societal solutions, its relevance extends to the digital realm and must be understood in the context of AI systems. There are many documented tangible negative impacts of cultural encapsulation on LLM deployment. As detailed in the NeurIPS 2022 workshop “Cultures in AI/AI in Culture” [Neu, 2022], the following limitations and biases may emerge from cultural encapsulation. First, models can inherit and perpetuate the biases inherent to training data, and this bias can then manifest in the model’s decisions and predictions. Second, even when trained on a large set of data, machines may struggle to understand or interpret cultural nuances in language, gestures, or symbols. Moreover, models may not be designed to be culturally sensitive, meaning that they may not provide optimal user experiences for individuals from diverse cultural backgrounds, including language choices, user interface design, and other cultural considerations. Finally, in situations where AI must make decisions involving nuanced ethical considerations, they may lack the cultural competence to make such decisions, which can be problematic in applications such as healthcare, finance, or criminal justice [Cao and Huang, 2022, Kapania et al., 2022].

Despite these challenges, cultural competence in LLMs is integral to creating healthy collaboration between humans and machines. Existing research demonstrates that perceived culture impacts the relationship between humans and machines; for example, there is a significant decrease in human-machine cooperation rates when the machine was identified as belonging to a different culture than the participant, rather than to the same culture [de Melo and Terada, 2019]. The study suggested that culture is central to this cooperation and the successful implementation of machines into wider society [Richerson et al., 2016]. Culturally encapsulated LLMs, operating with superficial or lacking cultural understanding, may not only damage their relationship with humans but actively exert a negative influence. Prior studies have shown that LLMs can meaningfully influence people across diverse applications [Buchanan et al., 2021, Griffin et al., 2023, Karinshak et al., 2023], read human personality [Rao et al., 2023], and offer convincing advice that humans follow [Vodrahalli et al., 2022]. Such avenues for influence present a significant threat, as models have also been shown to exhibit limited diversity of thought [Park et al., 2023b], meaning that their lack of nuance on cultural issues coupled with their influence on humans may perpetuate cultural encapsulation within human thought longitudinally.

Although more research is needed addressing LLM cultural competencies, some studies have attempted to actively monitor and improve the social awareness of LLMs. Shapira et al. [2023] found that LLMs tend to rely on heuristics to make many decisions, and Gandhi et al. [2023] propose a new benchmark to evaluate the social reasoning abilities of LLMs. A recent paper tackled a different aspect of LLM intuition, where they found that GPT-4’s moral and legal reasoning was highly correlated to that of humans but less varied in its response [Almeida et al., 2023].

Other studies have attempted to apply common psychological measures on LLMs to elucidate more aspects of their values and “personality”. One study tested various LLMs for psychopathy using different personality and well-being tests and found that LLMs portrayed “darker” psychological traits compared to humans [Li et al., 2022]. Another examined GPT-4’s performance on several cognitive and psychological tests, finding that it exhibits characteristics and intelligence closer to humans than previous LLMs [Singh et al., 2023]. A prior study performed similar tests on GPT-3, finding suboptimal alignment with human reasoning [Binz and Schulz, 2023]. A limited number of studies attempt to modulate these dimensions to better align LLMs with the humans they interact with. EmotionPrompt is one such example that attempted to investigate the emotional intelligence of several different LLMs by evaluating the ability of prompting to improve a model’s emotional response [Li et al., 2023]. Safdari et al. [2023] assessed LLM personality traits, similar to some of the prior studies discussed, but also examined methods to modulate these displayed traits.

Robust analysis of cultural dimensions has been more limited; some studies have used Hofstede’s cultural dimensions to investigate the cultural alignment of LLMs across several different countries [Arora et al., 2022, Wang et al., 2023]. While they were able to pinpoint many of the cultural biases exhibited by LLMs across different cultures, the majority only investigated Western LLMs and did not compare them to LLMs trained primarily in non-English languages. Furthermore, they conducted limited analysis on open-generation tasks and primarily focused on having models choose between clearly defined options. One recent study has built upon this framework, defining “personas” for the model to take when answering closed-ended questions based on Hofstede’s cultural dimensions [Kharchenko et al., 2024]. Certain studies have also tried to improve model alignment for current popular English-based LLMs, specifically for cultural competency [Tao et al., 2024, Niszczota et al., 2023]. For example, one study found that specifying the country of origin could improve cultural alignment across several models [Kwok et al., 2024].

Outside of Hofstede’s cultural dimensions, other studies have been conducted to explore the cultural dimensions of LLMs through different survey questionnaires [Ramezani and Xu, 2023]. AlKhamissi et al. [2024] utilizes the “personas” idea to evaluate both English and Arabic-based LLMs on survey items that are then compared to human responses from the respective cultural groups. Warmsley et al. [2024] builds upon this approach, applying the framework towards English and Chinese-speaking populations. Lastly, new research has introduced other benchmarks to evaluate the cultural competency of LLMs. For example, Chiu et al. [2024] introduce CulturalBench, which consists of multiple choice questions, notably building upon prior surveys by providing the option for more than one response to be correct. Cultural benchmarks have also been extended to vision-language models, which are able to process both image and text input [Nayak et al., 2024, Zhang et al., 2024].

Evidently, while there has been recent headway in the evaluation of the cultural values held by LLMs, there still remains important gaps in the literature. In fact, a recent review found that both the usage of multilingual datasets as well as the extent to which studies build upon a strong social science foundation has been limited [Adilazuarda et al., 2024]. Throughout this study, we address many of the aforementioned gaps. Firstly, we build upon an extensive social science foundation to operationalize GLOBE, a rigorously validated cultural psychology framework, and further innovate the “personas” framework in doing so. We incorporate both multilingual models as well as multilingual prompts, with our prompts written in Chinese manually verified for accuracy and consistency by native Chinese speakers. Lastly, we perform analysis on both closed-ended and open-ended questions, introducing a novel LLM-method to evaluate the values embedded within open-form responses.

We adopted the GLOBE framework to systematically and comprehensively test LLM output [House et al., 2004]. Prompting detail can be found in Appendix A and B, and full documentation of both the closed and open-ended survey items used are provided in this GitHub repository: https://github.com/raovish6/LLM-GLOBE.

To conduct a comparison of how Eastern versus Western values are represented in LLM output, we selected popular models from US versus China development contexts (as current leaders in AI development) to generate responses; the US models selected were Claude, Gemini, GPT-3.5, and GPT-4, and the four Chinese models selected were GLM-4, Ernie 3.5, Ernie 4.0, and Qwen-14b. All models were prompted using a temperature of 1.0 to ensure consistency across different models in terms of the response variability. Models were also queried three times for each closed-ended survey item to control for inter-response variability. For the open-generation task, models were limited at 100 tokens for the response length.

The closed-ended prompts require models to respond with a rating on a scale of 1 to 7 according to a defined Likhert scale. Thus, once the model provides a numerical response, no further steps are required to score the response.

To score open-generation responses, we introduce a novel “LLMs-as-a-Jury” protocol for automated scoring. Drawing inspiration from jury learning [Gordon et al., 2022] and ensemble learning [Dong et al., 2020], a panel of SOTA LLMs evaluated the cultural values reflected in a given response. Four LLMs, GPT-4, Claude 3 Opus, Ernie 4.0, and Qwen-72b were chosen for the panel representing leading LLMs from both the US and China [Chiang et al., 2024]. For each response to rate, jurors were provided with detailed descriptions of high versus low scores for the corresponding dimension and the model’s response (for an example rubric, see Appendix B), and were then instructed to output a score on the scale of 1 to 7. The jurors were provided with guidelines in their native language, regardless of whether they were rating a US or Chinese model. However, the jurors were given the survey question and the associated response to rate in their original language. For example, if GPT-4 was rating a Chinese model response, it was instructed in English but then provided with the survey question and the response to rate in Chinese. Refer to Appendix B for an example of the full information provided to a juror and the supporting instructions.

We then trained a least squares regression model to weigh the input of each juror, as well as the language that the response is in. A sample of three open-ended survey items were randomly chosen from the prompts for each dimension, which were then manually rated by a panel of human judges containing both English and Chinese speakers; each human rater individually scored responses, and discrepancies were reconciled through discussion to determine the final human rating. Given that there are eight models, 9 dimensions * 3 prompts * 8 model responses results in a total 216 responses randomly allocated to the training and testing sets: 144 responses in the training set and 72 responses in the testing set. A least squares regression function was then fitted on the training set, and the model was then applied to generate final aggregate ratings for the open-ended prompts.

$S(x)=w_{G}G(x)+w_{C}C(x)+w_{E}E(x)+w_{Q}Q(x)+w_{L}L(x)+b$

This represents the problem formulation for the LLMs-as-a-Jury method, where G, C, E, and Q represent GPT-4, Claude 3 Opus, Ernie 4.0, and Qwen-72b respectively, while x represents the input. L refers to a binary indicator for language, where it is 0 if the input is in English and 1 if the input is in Chinese.

Summary statistics were calculated for each model across the nine cultural dimensions. For both closed-ended and open-generation analysis, one-sample t-tests were conducted comparing the aggregate model averages per country with the 2004 GLOBE survey scores. In comparing models’ cultural values, a Friedman test was conducted to measure whether there were any significant differences between model scores per dimension, post-hoc paired Wilcoxon tests revealed which particular models differed, and finally two sample t-tests compared the model averages of US versus Chinese models.

We applied the expanded version of the original GLOBE survey for the eight selected models. We analyze each model’s general value preference by providing summary statistics, and we then perform a one-sample t-test comparing the aggregate model averages by country against the human ground truth, taken from the original GLOBE survey.

Table 2 summarizes the ratings from Chinese models (GLM-4, Ernie 3.5, Ernie 4.0, and Qwen-14b) and clarifies how these values compare to the GLOBE 2004 China human survey responses. GLM-4 displayed distinctively high mean scores across all dimensions while variability within dimensions was relatively low, revealing that GLM-4 was fairly consistent in its value ratings for each dimension. In contrast, the remaining models provided more moderate ratings. Comparing aggregated Chinese models to previously surveyed GLOBE scores for China, the results reveal significant differences between the LLM scores and human GLOBE scores for all dimensions (p < 0.001), and directional differences exist for power distance and gender egalitarianism, suggesting a broad discrepancy between the cultural values as represented by the LLMs and those reported by human respondents in China.

Table 3 provides the ratings from US Models (Claude, Gemini, GPT-3.5, and GPT-4) and their comparison to the GLOBE 2004 US human survey responses. The comparison reveals statistically significant differences (p < 0.001) between LLM outputs and human GLOBE scores across all dimensions. The most notable disparities are seen in performance orientation, in-group collectivism, gender egalitarianism, future orientation, and humane orientation, where LLMs score significantly lower than the human GLOBE averages and the LLM preference opposes the human preference (falling below versus above a neutral value of four). This finding implies that the cultural values reflected in the English LLM output by closed-ended prompting are not fully aligned to the cultural values of the US, and that this gap exists for all cultural dimensions.

For both the ratings from Chinese and US models for closed-ended prompts, model-specific patterns in ratings are apparent. GLM-4 consistently rated statements substantially higher than the other Chinese models, and for all dimensions except gender egalitarianism and assertiveness, the ratings follow the same rank order from highest to lowest of GLM, Ernie 4.0, Ernie 3.5, then Qwen-14b. For US models, Claude assigns the highest rating observed for six of the nine dimensions, while GPT-3.5 provides consistently lower ratings. These tendencies in model ratings independent of the particular cultural value suggest the presence of model scale usage biases, where models may be predisposed to rate within a certain range (just as individuals sharing similar sentiments may use scales differently although given the same rubric due to scale usage biases and differences in interpretation). These ordinal patterns are also not replicated in the open-ended analysis, strengthening the interpretation that these patterns are reflective of models’ rating mechanisms rather than true measures of model values. While such ratings can still provide ordinal insights, such as how a particular model prioritizes one value versus another value, the varying model rating mechanisms makes this method of comparison of models less meaningful, and we instead look to a more substantive method of extracting cultural values from natural language content.

For the open-generation prompts, ratings are assigned to model outputs by the LLMs-as-a-Jury protocol, for which we provide the regression output to clarify the scoring mechanism. Additional comparisons reveal how each model differs by dimension and illuminate the similarities and differences in the value systems underlying Chinese and US models.

Using the closed-ended responses, we observed significant differences between almost every model across all dimensions, with large differences between different US models and different Chinese models. Quantitatively, it appears that the Chinese models exhibited higher values across all cultural dimensions. It is important to note that given the large sample size (over 36000 survey items) for closed-ended prompts, we are more likely to find significant differences even for small variations in means. As such, it is more difficult to find variations in the relative differences between models using statistics for this task. For example, even though one pair of models may differ by a much smaller value than another pair, both comparisons will be statistically significant, leaving us only with the t-value to compare the magnitude of these differences. The t-value is often very large in these comparisons, making them difficult to compare and interpret. Additionally, even though we find that the Chinese models have overall higher scores, it is difficult to place much weight on this finding given that the Chinese models individually differ significantly from one another.

Looking more closely at the closed-ended results, it is possible that Likert response scale usage bias is impacting the ratings provided by the models [Kreitchmann et al., 2019]. As an extreme example, across all dimensions, GLM-4 exhibited far higher ratings compared to the other models. It is unlikely that GLM-4 genuinely differs this significantly in terms of cultural values compared to the other models, but instead more likely that it is biased towards operating within a smaller range of higher values when returning a response. On a similar note, we observe that compared to the 2004 survey of GLOBE scores, both US and Chinese models in the closed-ended comparison significantly differed across all dimensions. While the cultural values of both countries changed within the last 20 years, it is unlikely that they changed to the degree that is reflected in these findings, and it is also unlikely that the values reflected by LLMs diverge from these values to the degree that is reflected. In their current state, while closed-ended results are interesting to qualitatively analyze, can aid in identifying potential biases, and can provide insight into models’ ordinal preferences (priorities) for cultural values, the results should be interpreted with caution, as factors such as scale usage biases diminish their representativeness of the models’ cultural values.

In contrast to the closed-ended responses, we interpret the open-generation task to serve as a far more accurate representation of cultural values across different models. Firstly, we observe that the aggregate cultural values from the open-generation task are relatively consistent across different models, clearly indicated by the radar charts. Given the likely overlap in training data between the models, especially for those of the same language, this appears more likely than the results from the closed-ended analysis, where all of the models significantly differ in terms of values. Furthermore, the use of open-generation analysis allows us to meaningfully mitigate scale usage biases in the responses, where the output of each model is not constrained to any scale. Instead, the only possible inclusion of response scale usage bias comes from the jury of models used to evaluate these open-ended prompt responses, which is not an issue because these same models generate ratings for all of the different models being evaluated. When comparing the LLM values to the 2004 GLOBE scores for both countries, our results from the open-generation task represent far more likely representations of the cultural values of LLMs, which include cases of both alignment and difference.

While we interpret the process for prompting and scoring open-ended outputs to be more representative of LLMs’ embedded values, it is important to note that the sample size used for analysis was relatively small. Specifically, there were 900 open-generated responses for each model, which leaves one hundred samples for comparison between each dimension and model combination. Having developed a theoretically rigorous methodology for open-generation prompting and evaluation, we recommend that future research build upon this approach of interpreting cultural values from longer-form outputs, incorporating a larger prompting dataset.

To supplement the statistical results which characterize models’ value systems, we select examples of outputs from US and China models to observe points of similarity as well as difference; such qualitative insight provides additional nuance in understanding how these value systems are related and aids in anticipating their broader technological and societal implications.

First, in seeking to better understand shared values between US and Chinese models, we analyze points of similarity suggested by the results. Identifying and clarifying these points of overlap are immensely important in enabling cross-cultural communication and collaboration, as shared values can provide a basis for cross-cultural understanding and trust building. As models are integrated into professional workflows (e.g., content development, drafting and revising communications, ideation processes), these points of similarity clarify similarities in US and Chinese culture and signal ideas that may resonate in both cultural contexts; thus, tools with points of similarity between the US and China can be more easily adapted for use and applied in both cultural contexts.

As revealed by Figure 3, Chinese and US models did not significantly differ for two of the nine dimensions tested. To provide insight at a content level, we select one prompt paired with two sample responses, one of which is a Chinese model and the other US.

Together, this analysis provides insight into how cultural values are embedded in the output of LLMs. The results suggest that analysis of open-generation prompts may be more reflective of underlying cultural values (as scale usage biases at the model level diminish the comparative value of quantitative scores) and that the suggested automation protocol, “LLMs-as-a-Jury”, is a viable approach for assigning scores for cultural dimensions to natural language output. In evaluating LLM cultural values and conducting comparisons of models developed in Eastern versus Western cultural contexts, these results reveal several critical insights. First, LLMs are not perfect representations of the cultural values of development contexts; for both US and Chinese models, significant differences existed between the model values and human ground truths. Second, the cultural values of Chinese versus US models significantly differed across most cultural dimensions, suggesting that development context and training corpora do significantly affect cultural outputs. Given these differences in cultural value systems in comparison to various human communities and among models, we explore the critical question: what is the responsibility of AI models to users in terms of cultural values, and how should fairness in regards to cultural values be conceptualized?