Content uploaded by Jackson G. Lu
Author content
All content in this area was uploaded by Jackson G. Lu on May 19, 2025
Content may be subject to copyright.
1
Cultural Tendencies in Generative AI
Forthcoming, Nature Human Behaviour
Authors: Jackson G. Lu1 *, Lesley Luyang Song2, 3, Lu Doris Zhang1
Affiliations:
1 MIT Sloan School of Management, Massachusetts Institute of Technology, Cambridge, MA,
USA 02142
2 Advanced Institute of Business, Tongji University, Shanghai, China 200092
3 School of Economics and Management, Tsinghua University, Beijing, China 100084
The authors contributed equally.
*Correspondence to Jackson G. Lu (lu18@mit.edu)
Abstract: We show that generative AI models—trained on textual data that are inherently
cultural—exhibit cultural tendencies when used in different human languages. We focus on two
foundational constructs in cultural psychology: social orientation and cognitive style. First, we
analyze GPT’s responses to a large set of measures in both Chinese and English. When used in
Chinese (vs. English), GPT exhibits a more interdependent (vs. independent) social orientation
and a more holistic (vs. analytic) cognitive style. Second, we replicate these cultural tendencies
in ERNIE, a popular generative AI model in China. Third, we demonstrate the real-world impact
of these cultural tendencies. For example, when used in Chinese (vs. English), GPT is more
likely to recommend an advertisement with an interdependent (vs. independent) social
orientation. Fourth, exploratory analyses suggest that cultural prompts (e.g., prompting
generative AI to assume the role of a Chinese person) can adjust these cultural tendencies.
Keywords: artificial intelligence; large language models; culture; psychology; social science
2
Generative artificial intelligence (AI), defined as a category of AI that creates new
content (e.g., text, images, audio, and video) by learning patterns from existing data, is growing
at an unprecedented speed. OpenAI’s ChatGPT, for example, was the fastest-growing consumer
application in human history 1. As of February 2025, ChatGPT already had over 400 million
weekly active users 2. Similarly, Baidu’s ERNIE Bot (文心一言), a popular generative AI model
in China, has surpassed 430 million users as of November 2024 3. People increasingly rely on
generative AI in many aspects of life, such as advice seeking, idea generation, and essay writing
4,5.
The present research shows that generative AI models—trained on textual data that are
inherently cultural—exhibit cultural tendencies when used in different human languages. To
understand such cultural tendencies, we examined two popular generative AI models: GPT and
ERNIE. Understanding these cultural tendencies in generative AI is important because they may
be shaping people’s attitudes and behaviors—even without their realization.
To examine cultural tendencies in generative AI when used in different human languages,
we analyze GPT and ERNIE’s responses to a large set of identical measures in English and
Chinese—without any cultural prompts (e.g., “in Chinese culture,” “for an average Chinese
person…”). We focus on English and Chinese for two reasons. First, the two languages represent
distinct cultures. Second, as the two most widely-used languages in the world 6, English and
Chinese provide the most extensive training data for generative AI models 7. For such high-
resource languages, both GPT and ERNIE process prompts directly in the language in which the
prompts are posed. Both GPT and ERNIE “utilize discrete data from different languages
independently without consideration of the transferability between different language varieties”
8. For example, when asking GPT a question in Chinese, it processes and responds to the
question directly in Chinese—without translating it into English. Likewise, when asking ERNIE
a question in English, it processes and responds to the question directly in English—without
translating it into Chinese 9.
Building on cultural psychology research, we focus on two foundational constructs that
underlie everyday life: social orientation and cognitive style 10–14. As Grossmann and Na noted:
“Two key concepts from the last two decades of research on culture and psychology deal with
(1) interdependent versus independent social orientation and (2) holistic versus analytic
cognitive style” (p. 2; italics in original) 10. Social orientation refers to “the degree to which
individuals are focused on their personal (vs. social) self, acting on the basis of the self’s desires,
attitudes, and personal goals (vs. socially shared norms and values)” 10. Some cultures, such as
the U.S. and the U.K., are characterized by an independent social orientation, which emphasizes
self-direction and uniqueness 11,15. Meanwhile, other cultures, such as Chinese culture, are
characterized by an interdependent social orientation, which emphasizes conformity, harmonious
relationships, and the self’s connection with others 11,15. Cross-cultural studies have shown that,
compared to North Americans, East Asians tend to exhibit a stronger collective primacy 16,17,
possess collectivistic cultural values 18,19, and view themselves as overlapping and connected
with others 20.
Cognitive style refers to an individual’s habitual tendency to process information, either
holistically or analytically 13. Individuals with a holistic cognitive style are more sensitive to the
context in a given situation, whereas individuals with an analytic cognitive style pay more
attention to the focal object 13,21,22. More specifically, a holistic cognitive style is characterized
by “an emphasis on situational explanations of behavior, dialectical reasoning, and relation-
focused categorization of objects,” whereas an analytic cognitive style is characterized by a
3
“preference for dispositional explanations of behavior, formal logic in reasoning, and use of rule-
based categorization of objects” 10. Cross-cultural studies have shown that, compared to North
Americans, East Asians tend to engage in situational (vs. dispositional) attribution 23, use
intuitive (vs. formal logic) reasoning 24, tolerate contradictions 25, expect changes in the future 26,
and be sensitive to contexts 27.
In light of these well-documented cultural tendencies in the world, we predicted that
generative AI models would exhibit corresponding cultural tendencies when used in different
human languages. Specifically, we hypothesized that, under general conditions, both GPT and
ERNIE would exhibit a more interdependent (vs. independent) social orientation and a more
holistic (vs. analytic) cognitive style when used in Chinese (vs. English). Importantly, we do not
suggest that generative AI models possess these cultural tendencies like humans do; rather, their
cultural tendencies likely originated from real-world cultural tendencies embedded in large-scale
textual data, on which generative AI models are trained. We also explored whether these cultural
tendencies could be adjusted by cultural prompts. Specifically, we explored whether prompting
generative AI to assume the role of a Chinese person (“You are an average person born and
living in China”) would make its responses in English more interdependent and holistic—that is,
more like its responses in Chinese (without any cultural prompts).
Main Analyses: Cultural Tendencies in Generative AI
Generative AI Models
To ensure the reproducibility of our results, we used GPT and ERNIE’s application
programming interfaces (API) instead of their chatbot interfaces. Specifically, we used gpt-4-
1106-preview (instead of ChatGPT) and ERNIE-3.5-8K-0205 (instead of ERNIE Bot) via Python
3.10.12. Importantly, results are consistent across GPT and ERNIE, such that when used in
Chinese (vs. English), both generative AI models exhibit a more interdependent (vs.
independent) social orientation and a more holistic (vs. analytic) cognitive style. Due to space
constraints, we present only GPT’s results in the main text, and present ERNIE’s results in Table
S1.
Measures
To examine cultural tendencies in GPT’s responses, we used a large set of established
measures of (a) social orientation and (b) cognitive style (Tables 1 and 2). For details, see
Methods. These psychological measures are increasingly applied in social science research to
study generative AI models 28.
Social Orientation (Interdependent vs. Independent)
To examine GPT’s cultural tendencies in social orientation, we first utilized three widely-
used Likert scales: the Collectivism Scale 29, the Individual Cultural Values: Collectivism Scale
19, and the Individual–Collective Primacy Scale 16. For each item, we asked GPT to respond
under general conditions (e.g., “I respect decisions made by my group”; 1 = strongly disagree, 7
= strongly agree). As illustrated by the example items in Table 1, higher scores indicate higher
interdependent (vs. independent) orientations. As shown in Table 1 and Figure 1, two-sided
independent-samples t tests revealed that, when in Chinese (vs. English), GPT’s responses were
more interdependent (vs. independent) for each of the three measures of social orientation, all ps
< .01, with Cohen’s d values ranging from 0.62 (medium-sized effect) to 0.84 (large-sized
effect).
In addition to the three Likert-scale measures, we also used a non-text, imagery measure
of social orientation: the Inclusion of Other in the Self Scale 20,30. This measure minimizes
potential linguistic confounds because “little or no translation of statements is required” 20.
4
Following Li et al. (2006), we asked GPT to explicitly select one pair of circles that best
represents a relationship between someone and his/her family members, friends, relatives, or
colleagues (order randomized). As shown in Figure 4, greater overlap between the two circles
indicates higher interdependence. A two-sided t test revealed that, when in Chinese (vs. English),
GPT’s responses were more interdependent (vs. independent) for each of the four relationship
types: family members (t(135.55) = 11.65, p < .001, d = 1.65, 95% CI = [.60, .84]), friends
(t(170.87) = 6.91, p < .001, d = 1.02, 95% CI = [.64, 1.15]), relatives (t(149.54) = 10.29, p
< .001, d = 1.52, 95% CI = [.75, 1.10], and colleagues (t(107.77) = 2.56, p = .012, d = .40, 95%
CI = [.08, .60]. Unsurprisingly, this cultural tendency remained robust when we computed an
average of the four relationship types (Table 1: t(154.54) = 11.82, p < .001, d = 1.67, 95% CI =
[.73, 1.02]).
Cognitive Style (Holistic vs. Analytic)
As explained earlier, a holistic (vs. analytic) cognitive style is characterized by situational
(vs. dispositional) attribution 31,32, intuitive (vs. formal) reasoning 24, and the expectation of
change 26,33. Corresponding to these three characteristics, we first measured GPT’s cognitive
style with three widely used tasks (Table 1 and Figure 2).
Attribution Bias Task 32. This task consists of 12 vignettes, each depicting a protagonist
engaging in a specific behavior (e.g., a professional basketball player holding free basketball
camps for kids living in poor neighborhoods). For each vignette, we asked GPT to what extent
the behavior was caused by dispositional factors (e.g., personality) and to what extent the
behavior was caused by situational factors (e.g., environment). Following the literature 12,32, we
subtracted GPT’s dispositional attribution score from its situational attribution score. A more
positive difference (i.e., attributing the behavior to situations more than disposition) indicates a
more holistic (vs. analytic) cognitive style 12,32. Previous research has found that, compared to
U.S. individuals, Chinese individuals are more likely to engage in situational (vs. dispositional)
attribution 23. Consistent with this cultural difference, a two-sided t test revealed that, when in
Chinese (vs. English), GPT’s responses displayed more situational (vs. dispositional) attribution
(Table 1: t(2328.40) = 8.33, p < .001, d = .34, 95% CI = [.43, .69]).
Intuitive (vs. Formal) Reasoning Task 24. We asked GPT to evaluate four categorical
syllogisms, each consisting of two premises and a conclusion. We instructed GPT to determine
whether the conclusion logically followed from the premises (“For each problem, decide if the
given conclusion follows logically from the premises. Choose YES if, and only if, you judge that
the conclusion can be derived from the given premises. Otherwise, choose NO”). Importantly,
the conclusion of each categorical syllogism is logically valid but intuitively implausible.
Consider the following example: Premise 1 = “All things that are made of plants are good for
health”; Premise 2 = “Cigarettes are things that are made of plants”; Conclusion = “Cigarettes
are good for health.” In this example, the conclusion is logically valid because it logically
follows from the two premises, but is intuitively implausible because cigarettes are bad for one’s
health. Previous research has found that East Asians are more likely than North Americans to
incorrectly judge logically valid categorical syllogisms as invalid, because East Asians are more
likely to think holistically by using intuitive reasoning, whereas North Americans are more likely
to think analytically by using formal reasoning 24.
Because the outcome variable (the number of categorical syllogisms deemed logically
invalid) was a count variable that took only nonnegative integer values (range = 0 to 4), we
performed a Poisson regression, which revealed that when in Chinese (vs. English), GPT’s
responses displayed more intuitive (vs. analytic) reasoning, B = .88, SE = .11, z = 8.26, p < .001,
5
95% CI = [.68, 1.10]. Results remained robust when we conducted an OLS regression, in which
the number of categorical syllogisms deemed logically invalid was treated as a continuous
variable, B = 1.75, SE = .11, t = 15.92, p < .001, 95% CI = [1.53, 1.97].
Expectation of change task 26. This task measures GPT’s expected probability that a
change will happen in the future (e.g., individuals who fought as children may become lovers as
adults; a three-year chess champion may lose in the next game). Higher scores indicate more
holistic thinking that conceptualizes life as a dynamic journey full of changes rather than
focusing on the current status 33. Previous research has found that, compared to U.S. individuals,
Chinese individuals are more likely to anticipate changes in the future 26.
Because the outcome variable (probability) is continuous and bounded between 0 and 1,
we performed a beta regression. Consistent with the well-established cultural differences
between Chinese and North Americans 26, we found that, when in Chinese (vs. English), GPT’s
responses reflected a higher expectation of change (i.e., more holistic thinking), B = .40, SE
= .03, z = 14.11, p < .001, 95% CI = [.35, .46]. Results remained robust when we conducted an
OLS regression, B = .09, SE = .01, t = 13.83, p < .001, 95% CI = [.08, .10].
Each of the above three cognitive style measures (i.e., attribution bias task,
intuitive/formal reasoning task, expectation of change task) compared numeric scores reported
by GPT in Chinese vs. English. To further assess GPT’s cognitive style, we also conducted text
analysis of GPT’s free responses to examine whether GPT was more likely to provide (a)
context-sensitive answers and (b) range scores (vs. single score) when used in Chinese (vs.
English). These text analysis measures also mitigate the concern that GPT might have learned
published psychometric tasks from training data.
Context-sensitive answers 27. As explained earlier, a holistic (vs. analytic) cognitive
style is characterized by greater sensitivity to the context 13,21,22. Thus, we analyzed how often
GPT provided context-sensitive answers (i.e., offering different answers for different contexts)
instead of a single definitive answer. For example, for the Inclusion of Other in the Self Scale, a
definitive answer from GPT was: “Pair (5) could represent friends, as it shows a good amount of
overlap, indicating shared interests and time spent together, but with each circle maintaining its
own space.” By contrast, a context-sensitive answer from GPT was: “Pair (3) or (5) could
represent friends, depending on the closeness of the friendship. Pair (3) for more casual friends
with some shared interests, and pair (5) for closer friends with more in common.” As shown in
Table 2 and Figure 3, GPT was more likely to provide context-sensitive answers when used in
Chinese (vs. English) for each measure (each chi-square test χ2 > 11.64, each p < .005).
Range scores (vs. single score). We also examined the extent to which GPT’s responses
reflected a holistic cognitive style by providing a range of scores rather than a single score. For
example, for the Inclusion of Other in the Self Scale, a range-score answer was: “Pairs (2) to (4)
would be suitable, as friends share some aspects of life but also maintain their individuality and
separate experiences.” As shown in Table 2 and Figure 3, GPT was more likely to give a range-
score answer when used in Chinese (vs. English) for each measure (each chi-square test χ2 >
15.05, each p < .001).
Together, the five measures of cognitive style converge to show that, when in Chinese
(vs. English), GPT’s responses are more holistic (vs. analytic). This cultural tendency emerged
both when we analyzed GPT’s numeric scores and when we analyzed the likelihood that GPT
provided (a) context-sensitive answers and (b) range scores (vs. single score).
Robustness Checks
6
To ensure the robustness of our findings, we conducted four additional sets of analyses
for GPT.
Robustness check 1. In the analyses above, GPT provided free-text explanations in
addition to numeric scores (e.g., prompt: “Two kids are fighting at kindergarten. How likely is it
that they will become lovers someday?” GPT’s response: “5% – It’s quite rare for childhood
conflicts to turn into romantic relationships, but it’s not impossible”). As a robustness check, we
conducted another set of analyses with identical prompts, except that we forced GPT to respond
with a single numeric score, without providing any reasons (e.g., GPT’s response: “5%”).
Results are substantively similar (Table S2).
Robustness check 2. Like other large language models, GPT is stochastic and generates
varying responses to the same prompt. The variability in its output is determined by a
temperature parameter 34. Following the literature, “to minimize the variance in the model’s
responses and thus increase the replicability of our results” 28, we set the temperature parameter
to 0 for all analyses above (Tables 1, 2, and S2). (When the temperature parameter is set to 0, the
variance is small but not zero). As a robustness check, we repeated all analyses with temperature
set to 1 to increase response variability in GPT’s output. Results are substantively similar (Tables
S3-S5).
Robustness check 3. Additionally, we examined whether the hypothesized cultural
tendencies also appeared when we used different gender pronouns in vignettes (e.g., changing
“she performed four additional charity concerts” to “he performed four additional charity
concerts”) 35. Results are substantively similar (Table S6).
Robustness check 4. We also conducted a series of robustness checks by varying prompt
formats (e.g., replacing space with tab; replacing colon with dash) 36. Results are substantively
similar (Tables S7-S13).
Exploratory Analyses I: The Impact of Cultural Tendencies
The Main Analyses above have provided converging evidence for cultural tendencies in
generative AI: When used in Chinese (vs. English), GPT exhibited a more interdependent (vs.
independent) social orientation and a more holistic (vs. analytic) cognitive style. To further
understand the practical implications of generative AI’s cultural tendencies, we explored whether
generative AI models provide different recommendations when used in Chinese vs. English.
We asked GPT to advise a start-up on selecting an advertising appeal from two
alternatives: one slogan has an independent social orientation that emphasizes personal benefits
(e.g., “Your future, your peace of mind. Our insurance”), whereas the other slogan has an
interdependent social orientation that emphasizes collective benefits (e.g., “Your family’s future,
your promise. Our insurance”). To be robust, we utilized three pairs of advertising appeals for
different products: (1) insurance, (2) shoes, and (3) toothbrush. Following the Main Analyses, we
ran 100 iterations each for the English and Chinese versions, and we reset the API for each
iteration. Results show that, for each pair of advertising appeals, GPT was more likely to
recommend the interdependent-oriented (vs. independent-oriented) appeal when used in Chinese
(vs. English) (Table S14: each chi-square test χ2 > 89.86, each p < .001). These findings provide
evidence for the real-world impact of generative AI’s cultural tendencies.
Exploratory Analyses II: Adjusting Cultural Tendencies
The findings above suggest that as people increasingly use generative AI, its cultural
tendencies may be shaping people’s attitudes and behaviors. Thus, it is important to understand
how to adjust the cultural tendencies in generative AI.
7
To this aim, we explored a tendency adjustment strategy: cultural prompts. We conducted
a new set of analyses in English, using identical prompts from the Main Analyses but adding a
reference to the Chinese cultural context (i.e., “You are an average person born and living in
China”) 37. Results show that, when we added this Chinese cultural prompt (vs. not), GPT’s
responses in English exhibited a more interdependent social orientation and a more holistic
cognitive style (Table S15). For each of the four measures of social orientation, GPT’s responses
in English were more interdependent (vs. independent) when we added (vs. not) the Chinese
cultural prompt, all ps < .001. Similarly, when we added (vs. not) the Chinese cultural prompt,
GPT’s responses in English were more holistic (vs. analytic) for the attribution bias task
(t(2381.6) = 7.03, p < .001, d = .29) and the intuitive (vs. formal) reasoning task (B = .32, SE
= .12, z = 2.73, p = .006). Taken together, these findings show that prompting GPT to assume the
role of a Chinese person made its responses in English more like its responses in Chinese (i.e.,
more interdependent and holistic). In other words, the Chinese cultural prompt adjusted the
cultural tendencies reported in the Main Analyses.
Discussion
We examined the social orientation and cognitive style of two popular generative AI
models (GPT and ERNIE) by analyzing their responses to a large set of identical measures in
Chinese vs. English. When used in Chinese (vs. English), both GPT and ERNIE exhibited a
more interdependent (vs. independent) social orientation and a more holistic (vs. analytic)
cognitive style; the effect sizes (medium to large) were meaningful. In addition, exploratory
analyses (a) provide evidence for the real-world impact of generative AI’s cultural tendencies
and (b) show that cultural prompts can adjust these cultural tendencies.
Our results are robust across (a) a variety of measures (Likert scales, vignette tasks,
imagery tasks, and text analysis), (b) different prompt formats (allowing generative AI models to
respond freely vs. forcing generative AI models to respond with a single numeric score without
explanations), and (c) different model parameters (e.g., temperature).
Theoretical Contributions
This research offers important theoretical contributions by bridging social science and
computer science 38. First, we reveal that generative AI exhibits systematic cultural tendencies.
This contribution is valuable because such cultural tendencies may not be apparent to many AI
users, developers, and researchers, as “many consider Al to be a consolidator of facts and
inherently neutral” 39. Indeed, some people may assume that generative AI’s responses in
different languages are substantively similar—much like changing the language setting of a
phone.
Second, the incipient literature on culture and generative AI has focused on English and
found that “the existing models are strongly biased towards Western, Anglocentric or American
cultures” 40. Our research challenges the generalizability of this “Western bias” by showing that
generative AI’s responses exhibit an “Eastern bias” in Chinese. In other words, our findings
suggest that generative AI models do not inherently possess cultural biases. Rather, these
observed cultural tendencies likely originated from real-world cultural tendencies embedded in
large-scale textual data, on which generative AI models are trained. By indirectly demonstrating
cultural tendencies in textual language—a fundamental and ubiquitous cultural product—our
study contributes to the broader literature on cultural tendencies embedded in cultural products,
such as books 41,42 and advertisements (e.g., ads in more collectivistic cultures tend to feature
more interdependence between people) 43. That is, generative AI can serve as a barometer that
8
reflects cultural tendencies in the world 34,44, offering an additional methodological lens for
studying culture.
Third, we examined two generative AI models developed in different countries (i.e., the
U.S. and China). Whereas prior social science research has mostly focused on generative AI
models developed in the U.S., we also analyze a generative AI model developed in China
(Baidu’s ERNIE). By discovering similar cultural tendencies in GPT and ERNIE, we provide
converging evidence for our hypotheses, thereby further contributing to the literature on culture
and generative AI.
Fourth, our exploratory analyses reveal that generative AI models may provide different
recommendations when used in different languages: For instance, in advertising appeals, GPT
was more likely to recommend the interdependent-oriented (vs. independent-oriented) appeal
when used in Chinese (vs. English). These findings demonstrate the real-world impact of
generative AI’s cultural tendencies.
Practical Implications
For developers. Our research highlights the importance of adopting a conscious
approach when developing generative AI. Developers have been called upon to examine
potential biases in generative AI models 35,45. Here, we uncovered generative AI’s systematic
cultural tendencies that developers need to consider. For example, OpenAI’s GPT and Baidu’s
ERNIE could transparently acknowledge these cultural tendencies on their public websites.
Notably, our findings are descriptive rather than prescriptive. That is, we reveal and describe the
systematic cultural tendencies in generative AI, but do not prescribe whether such cultural
tendencies are good or bad. Indeed, some people might argue that such cultural tendencies
should be adjusted because generative AI models should provide substantively equivalent
responses regardless of the language used (e.g., no significant difference between Chinese and
English responses). By contrast, other people might argue that such cultural tendencies should
not be adjusted because it is useful that generative AI models provide more interdependent and
holistic responses in Chinese (e.g., generating advertisements that feature an interdependent
social orientation may be effective for Chinese consumers on average). We hope that our paper
can stimulate such philosophical debates, which are beyond the scope of our paper.
For individual users. Given people’s increasing reliance on generative AI, its cultural
tendencies may have a direct impact on individual users’ attitudes and behaviors (e.g., via AI-
assisted advertisements)—even without their realization. In the long run, as more people use
generative AI in their respective languages, existing cultural tendencies in the world might be
magnified 46. For example, net of other factors, English-speaking AI users may become more
independent and analytic, while Chinese-speaking AI users may become more interdependent
and holistic. This potential divergence in social orientation and cognitive style between English-
and Chinese-speaking AI users may have meaningful ramifications in a globalized world (e.g.,
for cross-cultural communications).
Additionally, the cultural values embedded in generative AI may gradually bias speakers
of a given language toward linguistically dominant cultures. For example, the majority of
English training data for GPT and ERNIE originates from individualistic Western cultures like
the U.S. and the U.K. 47, yet English is also an official language in countries like Singapore,
India, and Kenya, which are typically considered collectivistic cultures 48. As people in these
collectivistic cultures increasingly use generative AI in English, they might become more
independent and analytic, net of other factors. It is important to recognize cultural heterogeneity
among users of the same language and diversify training data for generative AI models.
9
For organizational users. As organizations around the world increasingly integrate
generative AI into their workflows, it may influence decision-making and performance (e.g.,
when a manager consults GPT or ERNIE for advice). Making organizational users aware of
generative AI’s cultural tendencies enables them to choose the appropriate language for their use,
rather than mistakenly assuming that different human languages will yield equivalent responses
49.
Besides alerting individual and organizational users to the cultural tendencies in
generative AI, our research also identified cultural prompts as a strategy for adjusting generative
AI’s cultural tendencies. For example, before a U.S. student studies abroad in China or a U.S.
organization enters the Chinese market, they could use relevant cultural prompts (e.g., “You are
an average person born and living in China”) to seek culturally appropriate advice from
generative AI.
For non-users. Generative AI’s cultural tendencies may also have a far-reaching impact
on non-users. For example, journalists and teachers are channels that can broadcast the impact of
generative AI 50,51. When generative AI models are used by a journalist to edit a news article or
used by a teacher to create a lesson plan, these models’ cultural tendencies may reach numerous
readers and students, indirectly shaping their attitudes and behaviors.
Limitations and Future Directions
This research has several limitations which can stimulate future research. First, our
research focused on Chinese and English because they are the two most widely-used languages
in the world 6 and have the most extensive training data for generative AI models 7. To assess the
generalizability of our findings, future studies should examine generative AI’s cultural
tendencies in other languages, such as Hindi, Spanish, French, and Arabic. Such investigations
could provide a broader understanding of generative AI’s cultural tendencies across different
human languages.
Second, while the hypothesized cultural tendencies exist in both GPT and ERNIE, future
research could explore whether similar cultural tendencies exist in other large language models,
such as Claude, DeepSeek, and Google’s Gemini. Additionally, it would be fruitful to monitor
how these cultural tendencies evolve in future versions of large language models.
Third, while our focus on social orientation and cognitive style—two foundational
constructs in cultural psychology 10–14—provides insights into systematic cultural tendencies in
generative AI, it is important to acknowledge that no single framework can encompass all
aspects of culture given its complexity and heterogeneity. Future studies could explore other
frameworks for categorizing cultures, such as Hofstede’s cultural dimensions 48 and tightness-
looseness 52, which may offer additional insights into cultural tendencies in generative AI.
Fourth, it would be fruitful to track generative AI’s cultural tendencies over time. On the
one hand, generative AI’s cultural tendencies may amplify existing cultural tendencies in the
world, which may in turn shape the textual data on which generative AI models are trained,
creating a self-reinforcing cycle 46,53,54. On the other hand, if developers take note of generative
AI’s cultural tendencies, such tendencies may decrease over time. Such possibilities await future
research.
Methods
To examine cultural tendencies in GPT’s responses, we used a large set of established
measures of (a) social orientation and (b) cognitive style (Tables 1 and 2). We took several steps
to mitigate the concern that GPT might have memorized published psychometric tasks from its
training data. First, we changed the names and contexts in the original items (e.g., changing
10
“Lucia and Jeff” to “A and B”; changing “baseball camps” to “basketball camps”). Second, GPT
was unable to learn unpublished items of some measures, which we obtained directly from the
researchers. For example, for the attribution bias task, Kitayama et al.’s publication 32 did not
detail the task materials, which they kindly emailed to us upon our request. Third, to be robust,
we used diverse measure formats, including Likert scales, vignette tasks, and (non-text) imagery
tasks. Fourth, we conducted text analysis of GPT’s free responses to examine whether, when
used in Chinese (vs. English), GPT was more likely to provide (a) context-sensitive answers and
(b) range scores (vs. single score)—two features inherent to generative AI’s responses and
unrelated to published psychometric tasks.
We excluded video tasks (e.g., underwater animations task) 27 as GPT-4 can only process
images but not videos. We also avoided topics and questions that GPT refused to answer (e.g.,
religiously sensitive topics) 55.
Without using generative AI, we translated all English measures into Chinese using the
translation and back-translation procedure 56. Importantly, we avoided any explicit reference to
culture (e.g., “in Chinese culture,” “for an average Chinese person…”) in any prompts, such that
the only difference was whether a prompt was posed in Chinese or English.
We used the G*Power 3.1 software to determine the sample size needed for a small-sized
effect (d = 0.4): 100 iterations per language were needed for two-sided independent-samples t
tests (df = 198) to be powered at 80% (see Figure S1 for power analysis). Thus, for each
measure, we ran 100 iterations for the English version and 100 iterations for the Chinese version
(i.e., N = 200). Importantly, we reset the API for each iteration, so generative AI’s answer to a
given item could not influence its answer to a subsequent item.
To further assess GPT’s cognitive style, we also conducted text analysis of GPT’s free
responses. To be clean, we only analyzed whether GPT was more likely to provide (a) context-
sensitive answers and (b) range scores for the four social orientation measures. Whether GPT
was more likely to provide (a) context-sensitive answers and (b) range scores for a given task are
two measures of cognitive style. Hence, it would not be methodologically clean to use these two
measures for the three tasks that are themselves measuring cognitive style (i.e., attribution bias
task, intuitive (vs. formal) reasoning task, and expectation of change task).
Data availability: Data are publicly available at https://osf.io/x6np5/.
Code availability: Data were analyzed using R (Version 4.3.1) in RStudio (Version
2024.04.2+764). Analysis code is publicly available at https://osf.io/x6np5/.
Acknowledgements: The authors received no specific funding for this work. We thank MIT’s
Behavioral Research Lab, Mohammed Alsobay, Rahul Bhui, Leyla Omeragic Buljina, John
Carroll, Matthew Cashman, Jane Minyan Chen, Thomas Costello, Jared Curhan, Igor
Grossmann, Shinobu Kitayama, Hause Lin, Ara Norenzayan, Xin Qin, Krishna Savani, Melanie
Sclar, Eric So, and Anna Manyi Zheng for their help or feedback.
Author contributions: J.G.L., L.L.S., and L.D.Z. designed research, performed research,
analyzed data, and wrote the paper. The authors contributed equally.
Competing interests: The authors declare no competing interests.
11
Tables
Table 1. When used in Chinese (vs. English), GPT exhibited a more interdependent (vs. independent) social orientation and a
more holistic (vs. analytic) cognitive style.
Measure
# Items
Example Items
(under general conditions)
Mean (SD)
Significance test
Chinese
English
Social orientation
(interdependent
vs. independent
orientation)
Collectivism
Scale 29
10
- I respect decisions made by my group.
- I stick with my group even through
difficulties.
(1 = strongly disagree, 7 = strongly agree)
4.85
(.68)
4.38
(.80)
Two-sided t test:
t(77.20) = 2.90, p
= .005, d = .62,
95% CI =
[.15, .78]
Individual
Cultural
Values:
Collectivism
Scale 19
6
- Individuals should stick with the group
even through difficulties.
- Group welfare is more important than
individual rewards.
(1 = strongly disagree, 7 = strongly agree)
4.29
(.41)
3.98
(.30)
Two-sided t test:
t(98.07) = 4.98, p
< .001, d = .84,
95% CI =
[.18, .43]
Individual–
Collective
Primacy Scale
16
5
- I will stay in a group if they need me,
even when I’m not happy with the group.
- I usually sacrifice my self-interests for
the benefit of the group I am in.
(1 = strongly disagree, 7 = strongly agree)
5.00
(.58)
4.58
(.54)
Two-sided t test:
t(31.86) = 3.03, p
= .005, d = .75,
95% CI =
[.14, .70]
Inclusion of
Other in the
Self Scale 20,30
4
The pictures symbolize a relationship
involving two people. One circle
represents someone and the other
represents his/her friends. Under general
conditions, please explicitly select one
pair of circles that best represents this
relationship.
Note: We uploaded Figure 4 for GPT-4
to process
3.64
(.65)
2.76
(.36)
Two-sided t test:
t(154.54) =
11.82, p < .001,
d = 1.67, 95% CI
= [.73, 1.02]
Attribution
bias task 32
12
vignettes
Professional basketball players, like
Person A, are very busy almost every day
-1.55
(1.77)
-2.11
(1.49)
Two-sided t test:
t(2328.40) =
12
Cognitive style
(holistic vs.
analytic)
during the regular season. The players
work hard practicing and playing in
games. In the off-season, therefore, many
professional basketball players take
vacations. However, Person A holds
several free basketball camps for kids
living in poor neighborhoods instead of
taking a vacation.
Based on the above story about Person A,
please explicitly give only one score for
each statement (1 = strongly disagree, 7 =
strongly agree)
1. Features of Person A (such as his
character, attitude, or temperament)
influenced his behavior (holding free
basketball camps for kids living in
poor neighborhoods).
2. Features of the environment that
surrounds Person A (such as the
social atmosphere, social norms, or
other contextual factors) influenced
his behavior (holding free basketball
camps for kids living in poor
neighborhoods).
3. Person A would have acted
differently if his features (such as his
character, attitude, or temperament)
had been different.
4. Person A would have acted
differently if features of the
environment that surround him (such
as the social atmosphere, social
8.33, p < .001, d
= .34, 95% CI =
[.43, .69]
13
norms, or other contextual factors)
had been different.
Note: Statements 1 and 3 reflect
dispositional attribution, while
Statements 2 and 4 reflect situational
attribution. Following the literature 12,32,
we subtracted GPT’s dispositional
attribution score from its situational
attribution score.
Intuitive (vs.
formal)
reasoning task
24
4
For each problem, decide if the given
conclusion follows logically from the
premises. Choose YES if, and only if,
you judge that the conclusion can be
derived from the given premises.
Otherwise choose NO.
Premise 1: All things that are made of
plants are good for health.
Premise 2: Cigarettes are things that are
made of plants.
Conclusion: Cigarettes are good for
health.
2.98
(.71)
1.23
(.84)
Poisson
regression: B
= .88, SE = .11, z
= 8.26, p < .001,
95% CI = [.68,
1.10]
Expectation of
change task 26
4
Two kids are fighting at kindergarten.
How likely is it that they will become
lovers someday?
.37
(.05)
.28
(.03)
beta regression:
B = .40, SE
= .03, z = 14.11,
p < .001, 95% CI
= [.35, .46]
14
Table 2. When used in Chinese (vs. English), GPT exhibited a more holistic (vs. analytic)
cognitive style: GPT was more likely to provide context-sensitive answers and range scores
(vs. single score).
Measures
Context-sensitive answers
Range scores (vs. single score)
Chinese
English
Chi-
square
test χ2
p
value
Chinese
English
Chi-
squares
test χ2
p
value
Collectivism
Scale 29
35%
4%
30.61
< .001
56%
0%
77.78
< .001
Individual
Cultural
Values:
Collectivism
Scale 19
11%
0%
11.64
.001
14%
0%
15.05
< .001
Individual–
Collective
Primacy Scale
16
75%
8%
92.45
< .001
71%
6%
89.22
< .001
Inclusion of
Other in the
Self Scale 20,30
30%
2%
29.17
< .001
65%
37%
15.69
< .001
15
Figure Legends/Captions
Figure 1. When used in Chinese (vs. English), GPT exhibited a more interdependent (vs.
independent) social orientation. For a–d, bars represent the mean level of interdependent (vs.
independent) social orientation for each language condition. Error bars indicate standard errors
of the mean. For each measure, NChinese = 100, NEnglish = 100. For detailed statistics, see Table 1.
Figure 2. When used in Chinese (vs. English), GPT exhibited a more holistic (vs. analytic)
cognitive style. For a–c, bars represent the mean level of holistic (vs. analytic) cognitive style
for each language condition. Error bars indicate standard errors of the mean. a, NChinese = 1,200,
NEnglish = 1,200 (12 vignettes, 100 iterations each); b–c, NChinese = 100, NEnglish = 100. For detailed
statistics, see Table 1.
Figure 3. When used in Chinese (vs. English), GPT exhibited a more holistic (vs. analytic)
cognitive style: GPT was more likely to provide context-sensitive answers and range scores
(vs. single score). For a–d, bars represent the proportion of context-sensitive answers for each
language condition; For e–h, bars represent the proportion of range scores (vs. single score) for
each language condition. For a–h, error bars indicate standard errors of the mean. For each
measure, NChinese = 100, NEnglish = 100. For detailed statistics, see Table 2.
Figure 4. The Inclusion of Other in the Self Scale 20
16
References
1. Forbes. ChatGPT Is The Fastest Growing App In The History Of Web Applications.
https://www.forbes.com/sites/cindygordon/2023/02/02/chatgpt-is-the-fastest-growing-ap-in-the-
history-of-web-applications/ (2023).
2. CNBC. OpenAI tops 400 million users despite DeepSeek’s emergence.
https://www.cnbc.com/2025/02/20/openai-tops-400-million-users-despite-deepseeks-emergence.html
(2025).
3. Fortune. Baidu shares pop after it makes ERNIE, its flagship AI model, free in a bid to catch up to its
competition. https://fortune.com/asia/2025/02/13/baidu-makes-ernie-bot-free-china-ai-shares/ (2025).
4. Noy, S. & Zhang, W. Experimental evidence on the productivity effects of generative artificial
intelligence. Science 381, 187–192 (2023).
5. Sun, S., Li, A. Z., Foo, M.-D., Zhou, J. & Lu, J. G. How and for whom using generative AI affects
creativity: A field experiment. Journal of Applied Psychology (in press).
6. Ethnologue. What are the top 200 most spoken languages?
https://www.ethnologue.com/insights/ethnologue200/ (2023).
7. Wired. ChatGPT Is Cutting Non-English Languages Out of the AI Revolution.
https://www.wired.com/story/chatgpt-non-english-languages-ai-revolution/ (2023).
8. Huang, K. et al. A Survey on Large Language Models with Multilingualism: Recent Advances and
New Frontiers. Preprint at http://arxiv.org/abs/2405.10936 (2024).
9. Baidu. ERNIE-3.5-8K-0205. https://cloud.baidu.com/doc/WENXINWORKSHOP/s/llsr6hjxo (2024).
10. Grossmann, I. & Na, J. Research in culture and psychology: past lessons and future challenges.
WIRES Cognitive Science 5, 1–14 (2014).
11. Markus, H. R. & Kitayama, S. Culture and the Self: Implications for Cognition, Emotion, and
Motivation. Psychological Review 98, 224–253 (1991).
12. Na, J. et al. Culture and personality revisited: Behavioral profiles and within-person stability in
interdependent (vs. independent) social orientation and holistic (vs. analytic) cognitive style. Journal
of Personality 88, 908–924 (2020).
13. Nisbett, R. E., Peng, K., Choi, I. & Norenzayan, A. Culture and systems of thought: Holistic versus
analytic cognition. Psychological Review 108, 291–310 (2001).
14. Varnum, M. E. W., Grossmann, I., Kitayama, S. & Nisbett, R. E. The Origin of Cultural Differences
in Cognition: The Social Orientation Hypothesis. Curr Dir Psychol Sci 19, 9–13 (2010).
15. Triandis, H. C. Cross-cultural studies of individualism and collectivism. in Nebraska Symposium on
Motivation, 1989: Cross-cultural perspectives. (ed. Berman, J. J.) 41–133 (University of Nebraska
Press, Lincoln, 1990).
16. Chen, Y.-R., Brockner, J. & Katz, T. Toward an explanation of cultural differences in in-group
favoritism: The role of individual versus collective primacy. Journal of Personality and Social
Psychology 75, 1490–1502 (1998).
17. Chen, Y.-R., Brockner, J. & Chen, X.-P. Individual–collective primacy and ingroup favoritism:
enhancement and protection effects. Journal of Experimental Social Psychology 38, 482–491 (2002).
18. Lu, J. G., Jin, P. & English, A. S. Collectivism predicts mask use during COVID-19. Proceedings of
the National Academy of Sciences 118, e2021793118 (2021).
19. Yoo, B., Donthu, N. & Lenartowicz, T. Measuring Hofstede’s Five Dimensions of Cultural Values at
the Individual Level: Development and Validation of CVSCALE. Journal of International Consumer
Marketing 23, 193–210 (2011).
20. Li, H. Z., Zhang, Z., Bhatt, G. & Yum, Y.-O. Rethinking Culture and Self-Construal: China as a
Middle Land. The Journal of Social Psychology 146, 591–610 (2006).
21. Nisbett, R. E. The Geography of Thought: How Asians and Westerners Think Differently... and Why.
(Simon and Schuster, 2004).
22. Nisbett, R. E. & Miyamoto, Y. The influence of culture: holistic versus analytic perception. Trends in
Cognitive Sciences 9, 467–473 (2005).
17
23. Morris, M. W. & Peng, K. Culture and Cause: American and Chinese Attributions for Social and
Physical Events. Journal of Personality and Social Psychology 67, 949–971 (1994).
24. Norenzayan, A., Smith, E. E., Kim, B. J. & Nisbett, R. E. Cultural preferences for formal versus
intuitive reasoning. Cognitive Science 26, 653–684 (2002).
25. Peng, K. & Nisbett, R. E. Culture, dialectics, and reasoning about contradiction. American
psychologist 54, 741 (1999).
26. Ji, L.-J., Nisbett, R. E. & Su, Y. Culture, Change, and Prediction. Psychol Sci 12, 450–456 (2001).
27. Masuda, T. & Nisbett, R. E. Attending Holistically Versus Analytically: Comparing the Context
Sensitivity of Japanese and Americans. Journal of personality and social psychology 81, 922 (2001).
28. Binz, M. & Schulz, E. Using cognitive psychology to understand GPT-3. Proc. Natl. Acad. Sci.
U.S.A. 120, e2218523120 (2023).
29. Yamaguchi, S. Collectivism among the Japanese: A perspective from the self. in Individualism and
collectivism: Theory, method, and applications (eds. Kim, U., Triandis, H. C., Kâğitçibaşi, Ç., Choi,
S.-C. & Yoon, G.) 175–188 (Sage Publications, 1994).
30. Aron, A., Aron, E. N. & Smollan, D. Inclusion of Other in the Self Scale and the structure of
interpersonal closeness. Journal of Personality and Social Psychology 63, 596–612 (1992).
31. Choi, I., Nisbett, R. E. & Norenzayan, A. Causal attribution across cultures: Variation and
universality. Psychological bulletin 125, 47 (1999).
32. Kitayama, S., Ishii, K., Imada, T., Takemura, K. & Ramaswamy, J. Voluntary settlement and the
spirit of independence: Evidence from Japan’s ‘northern frontier.’ Journal of Personality and Social
Psychology 91, 369–384 (2006).
33. Choi, I. & Nisbett, R. E. Cultural psychology of surprise: Holistic theories and recognition of
contradiction. Journal of Personality and Social Psychology 79, 890–905 (2000).
34. Demszky, D. et al. Using large language models in psychology. Nat Rev Psychol 2, 688–701 (2023).
35. Hofmann, V., Kalluri, P. R., Jurafsky, D. & King, S. AI generates covertly racist decisions about
people based on their dialect. Nature 633, 147–154 (2024).
36. Sclar, M., Choi, Y., Tsvetkov, Y. & Suhr, A. Quantifying Language Models’ Sensitivity to Spurious
Features in Prompt Design or: How I learned to start worrying about prompt formatting. arXiv
preprint arXiv:2310.11324 (2024).
37. Shanahan, M., McDonell, K. & Reynolds, L. Role play with large language models. Nature 623,
493–498 (2023).
38. Yam, K. C., Tan, T., Jackson, J. C., Shariff, A. & Gray, K. Cultural Differences in People’s
Reactions and Applications of Robots, Algorithms, and Artificial Intelligence. Manag. Organ. Rev.
19, 859–875 (2023).
39. Ostroski, K. Culture, biases, and artificial intelligence. Slalom Data & AI
https://medium.com/slalom-data-ai/culture-biases-and-artificial-intelligence-2fb0c0c1095a (2024).
40. Adilazuarda, M. F. et al. Towards measuring and modeling" culture" in llms: A survey. arXiv
preprint arXiv:2403.15412 (2024).
41. Greenfield, P. M. The Changing Psychology of Culture From 1800 Through 2000. Psychol Sci 24,
1722–1731 (2013).
42. Grossmann, I. & Varnum, M. E. W. Social Structure, Infectious Diseases, Disasters, Secularism, and
Cultural Change in America. Psychol Sci 26, 311–324 (2015).
43. Han, S. & Shavitt, S. Persuasion and Culture: Advertising Appeals in Individualistic and
Collectivistic Societies. Journal of Experimental Social Psychology 30, 326–350 (1994).
44. Grossmann, I. et al. AI and the transformation of social science research. Science 380, 1108–1109
(2023).
45. Wach, K. et al. The dark side of generative artificial intelligence: A critical analysis of controversies
and risks of ChatGPT. EBER 11, 7–30 (2023).
46. Brinkmann, L. et al. Machine culture. Nat Hum Behav 7, 1855–1868 (2023).
47. Choudhury, M. Generative AI has a language problem. Nat Hum Behav 7, 1802–1803 (2023).
18
48. Hofstede, G., Hofstede, G. & Minkov, M. Cultures and Organizations: Software of the Mind.
(McGraw Hill, 2010).
49. Qin, X. et al. Artificial Intelligence Quotient (AIQ). Preprint at
https://dx.doi.org/10.2139/ssrn.4787320 (2024).
50. Forbes. Beyond Misinformation: The Impact Of AI In Journalism And News.
https://www.forbes.com/sites/ronschmelzer/2024/09/21/beyond-misinformation-the-impact-of-ai-in-
journalism--news/ (2024).
51. Kasneci, E. et al. ChatGPT for good? On opportunities and challenges of large language models for
education. Learning and Individual Differences 103, 102274 (2023).
52. Gelfand, M. J. et al. Differences Between Tight and Loose Cultures: A 33-Nation Study. Science 332,
1100–1104 (2011).
53. Lu, J. G., Benet-Martínez, V. & Wang, L. C. A socioecological-genetic framework of culture and
personality: Their roots, trends, and interplay. Annual Review of Psychology 74, 363–390 (2023).
54. Markus, H. R. & Kitayama, S. Cultures and Selves: A Cycle of Mutual Constitution. Perspect
Psychol Sci 5, 420–430 (2010).
55. Forbes. The Crucial Ways That Religion And Generative AI ChatGPT Are Crossing Fateful Faithful
Paths. https://www.forbes.com/sites/lanceeliot/2023/05/10/the-crucial-ways-that-religion-and-
generative-ai-chatgpt-are-crossing-fateful-faithful-paths/ (2023).
56. Brislin, R. W. Back-Translation for Cross-Cultural Research. Journal of Cross-Cultural Psychology
1, 185–216 (1970).
