How Culture Shapes What People Want From AI

• U.S. portrayals often cast AI as separate from people and threatening, whereas other cultural contexts may view people and the environment as connected and reciprocal.The paper relates these differences to independent and interdependent models of self and environment.Source: §1 · p. 1
• The authors argue that cultural models shape how people imagine ideal AI, while culture remains underexamined in AI theory and design.They suggest that varied cultural ideas could broaden imagined AI categories and their potential societal and environmental benefits.Source: §1 · p. 2
• The authors use two exploratory surveys of European American, African American, and Chinese groups to examine culturally shaped AI preferences without treating those groups as monolithic.Source: §1 · p. 2
• European Americans preferred more control and AI with fewer capacities to influence than Chinese respondents, who preferred more connection; African American preferences reflected both models.The capacities included autonomy, spontaneity, emotion, and whether AI needed care from people.Source: §1 · p. 2
• The work aims to develop empirical approaches to cultural AI preferences and expand HCI models by surfacing their implicit assumptions about people.Source: §1 · p. 2

2 HUMAN-COMPUTER INTERACTION RESEARCH ON CULTURE

• Between 2010 and 2020, culture was explicit in only 1.9% of reviewed HCI publications and 1.6% of CHI publications.Earlier, only 28 of 3,286 papers (less than 1%) focused on culture between 1990 and 2006.Source: §2 · p. 2
• The review groups cultural HCI research into globalization, situated understanding, design practices, and equity; most work falls under globalization.Source: §2 · pp. 2–3

2.1 Limitations

• Many HCI studies do not draw on behavioral-science research on culture, instead treating culture as a usability or interface-design consideration.Source: §2.1 · p. 3
• Most studies examine technology’s effects across countries, while almost no research asks how culture shapes AI design practices or is reflected in AI products.Source: §2.1 · p. 3
• This research examines how culture shapes AI preferences, with potential implications for culturally sensitive and relevant AI systems.Source: §2.1 · p. 3

3.1 Conception of Culture and Cultural Models

• Cultural models are tacit assumptions and meanings encoded in practices, relationships, institutions, and artifacts that orient attention, feelings, and behavior.Source: §3.1 · p. 3
• The paper defines cultural models of self and environment as ideas and practices about how people understand themselves, their surroundings, and their relationship to them.Source: §3.1 · p. 3

3.2 Cultural Models in European American, African American, and Chinese Contexts

• The paper examines European American, Chinese, and African American groups, whose prior research profiles tend toward independent, interdependent, and mixed models, respectively.Source: §3.2 · p. 3
• The independent model emphasizes distinct individuals acting on their environments, while the interdependent model emphasizes connection to people and surroundings and responsiveness to them.Source: §3.2 · pp. 3–4
• These models are contextual rather than inherent, and African American cultural contexts can combine independent and interdependent tendencies.Source: §3.2 · p. 4
• Few studies examine how cultural models guide relationships with non-human systems such as AI, motivating the authors’ pilot study across the three groups.Source: §3.2 · p. 4

4.1 Cultural Models Predict Ideal Human-AI Interactions

• Independent models suggest hierarchical control over AI, while interdependent models allow more equal, connecting relationships; African Americans were expected to seek both control and connection.The authors expected Chinese respondents to favor connection more than European Americans, and African American responses to fall between the groups.Source: §4.1 · p. 4

4.2 Cultural Models Shape Preferences Regarding AI’s Capacities to Influence

• The authors propose that European Americans will prefer fewer AI capacities to influence than Chinese respondents, with African American preferences falling between them.They define capacities to influence as qualities that make AI an active force and derive these predictions from cultural models of the self and environment.Source: §4.2 · pp. 4–5

4.3 Other Related Research

• “Capacities to influence” includes sourcehood, action, and behavior, extending beyond mind perception’s self-control emphasis to interdependent agency such as harmony and flexibility.The authors use this broader construct because mind perception does not capture aspects of agency relevant to both independent and interdependent cultural models.Source: §4.3 · p. 5
• Anthropomorphism may miss cases where people treat AI as active or intelligent without likening it to humans, using other schemas to understand influential elements of the environment.The authors note that anthropomorphism assumptions were largely derived from independent cultural models and an anthropocentric view of influence.Source: §4.3 · p. 5

5 CURRENT STUDIES

• The pilot study assessed whether the three groups differed in cultural models of self and environment before the main study tested hypotheses about ideal human-AI interaction.The surveys were administered in English to European and African American participants and Simplified Chinese to Chinese participants, with attention and comprehension checks.Source: §5 · p. 5, Figure 2 · p. 6
• Both studies received Stanford Institutional Review Board approval, and participants consented before taking part.Data and surveys were made available through the Open Science Framework.Source: §5 · pp. 5–6
• Group means were compared using ANOVA or MANOVA, with Tukey’s HSD post hoc tests.Main-study mediator and regression analyses explored potential mediators and controlled for other sources of variation.Source: §5 · p. 5

6.1.1 Participants.

• After attention-check exclusions, the pilot analyzed 131 European Americans, 121 African Americans, and 121 Chinese participants recruited through online survey platforms.Recruitment targeted at least 100 participants per group; the Chinese sample was recruited in Beijing and included participants from varied regions of China.Source: §6.1.1 · p. 6

6.1.2 Procedure.

• Participants reported age, gender, and gross household income after answering questions about their ideal models of self and environment.Source: §6.1.2 · p. 6

6.1.3 Measures.

• Participants selected an illustration representing their ideal relationship between person and environment, using a seven-picture measure that varied the relative influence shown by arrows.The environment included external factors such as physical objects, other people, social norms, and cultural ideas.Source: §6.1.3 · p. 6, Figure 3 · p. 6

6.2.2 Additional analyses.

• After controlling for age, gender, and income, cultural-group effects remained significant, but the regression model fit was not strong.The authors recommend further research on how demographic-background differences may contribute to the observed effects.Source: §6.2.2 · p. 6

6.2.1 Cultural differences in the ideal level and direction of influence between the self and the environment.

• Chinese respondents more strongly preferred environmental influence over the person than European and African Americans; African American responses resembled European Americans and differed from Chinese responses.The Chinese mean was 3.62, compared with 4.47 for European Americans and 4.20 for African Americans; the European–African American difference was not significant.Source: §6.2.1 · p. 6, Figure 4 · p. 7, §6.3 · p. 7

6.3 Pilot Study Summary and Discussion

• The pilot supported the theoretical expectation of distinct cultural models: European Americans favored people influencing their surroundings, while Chinese favored the environment influencing people.African American responses resembled European Americans and were statistically different from Chinese; these findings preceded the main study of ideal AI.Source: §6.3 · p. 7

7.1.1 Participants.

• After attention-check exclusions, the main study analyzed 127 European Americans, 105 African Americans, and 116 Chinese participants.The study recruited through the pilot’s platforms, criteria, and procedures; its planned minimum sample was based on 80% power to detect a small-to-medium effect.Source: §7.1.1 · p. 7

7.1.3 Measures.

• Participants read one of six randomly assigned scenarios representing common, non-exhaustive AI applications before answering questions about their ideal AI.They also reported AI familiarity and demographic information.Source: §7.1.3 · p. 7, Table 4 · p. 8
• The study measured the importance of control and connection, and a nine-item two-option measure of AI influence capacities based on independent and interdependent cultural models.The influence-capacity items covered characteristics including autonomy, emotionality, physicality, social presence, and relationality; the measure had internal consistency α = .86.Source: §7.1.3 · p. 7, Table 5 · p. 8

7.2.1 Overall group differences.

• Culture had a significant multivariate effect across control, connection, and AI influence capacities, Pillai’s Trace = .34, F(6,345) = 23.73, p <.001, partial 𝜂p2 = .17.Univariate effects were significant for all three variables.Source: §7.2.1 · pp. 8–9

7.2.2 Group differences in importance of having control over AI and connecting with AI.

• Chinese rated control as less important (M = 3.72) than European Americans (M = 4.21) and African Americans (M = 4.08), but connection as more important (M = 3.81) than both groups.African Americans valued connection more than European Americans, while their control ratings did not significantly differ.Source: §7.2.2 · p. 9, Figure 5 · p. 9

7.2.3 Group differences in preferences for ideal AI’s capacities to influence.

• Chinese preferred higher AI influence capacities (M = 1.68) than European Americans (M = 1.24), while African Americans (M = 1.36) fell between them.The African American preference was significantly higher than the European American preference and lower than the Chinese preference.Source: §7.2.3 · p. 9, Figure 6 · p. 10

7.2.4 Exploratory mediation analyses.

• Control and connection were associated with significant indirect effects on cultural differences in AI influence preferences, but the direct effect remained significant at .71.The indirect effects were .17 for control and .39 for connection; Figure 7 depicts the mediation results.Source: §7.2.4 · p. 10, Figure 7 · p. 10

7.2.5 Additional analyses.

• The samples differed in gender, age, income, and AI familiarity, with Chinese participants younger and higher on the income ladder than the other groups.African Americans reported the highest AI familiarity; only the African American–Chinese difference was significant.Source: §7.2.5 · p. 10
• Cultural differences remained significant after controlling for gender, age, income, and AI familiarity, and additional regressions found the African American–European American differences robust to those controls.The control-importance model had a lower adjusted R-squared than the connection and AI-influence models.Source: §7.2.5 · p. 10

7.3 Study Summary and Discussion

• Chinese regarded control over AI as less important and connection as more important than European Americans, and preferred greater AI influence capacities.African Americans were similar to European Americans on control, intermediate on connection and AI influence, and still expressed a desire for control.Source: §7.3 · pp. 10–11

8 GENERAL DISCUSSION

• Independent and interdependent models of self and environment frame different preferences: control and low AI influence versus connection and greater AI influence.The paper links these preferences to contexts that promote each model.Source: §8 · p. 11
• The Pilot Study measured cultural differences in ideal relationships with the environment, and the Main Study measured differences in desired AI interactions across three cultural groups.Both studies used new measures.Source: §8 · p. 11

8.1.1 Rethinking control-based relationships in designing human-AI interaction.

• A connection-focused relationship is presented as an alternative to control-focused AI design and may fit values across a broader range of the world population.The authors also suggest it may address some societal problems associated with current AI use.Source: §8.1.1 · p. 11

8.1.2 Broadening the space of the imaginary for conceptualizing AI’s characteristics.

• The framework and findings challenge the default view of AI as impersonal systems that optimize, analyze, and repeat.The authors propose experimenting with AI designed with “needs” and foregrounded in social situations.Source: §8.1.2 · p. 11

8.1.3 Re-evaluating how human-centeredness is conceived and im-

• Because values and behaviors vary across societies, human-centered AI may need to look different across cultural contexts, with culture considered from the initial design phase.The paper also calls for reflection on organizational barriers to designing for culturally variable human needs.Source: §8.1.3 · p. 12

8.2 Limitations and Future Directions

• The findings are constrained by relatively small, demographically different samples and new measures whose reliability and validity need further establishment.The studies used general AI preferences rather than specific products, measured reported preferences rather than actual interactions, and call for evidence from other culturally diverse samples.Source: §8.2 · p. 12

• The paper uses different cultural models of the self and environment to begin expanding current visions of AI.Source: §9 · p. 12
• Theoretical frameworks and empirical tools are offered for examining how culture shapes what people want in AI.Source: §9 · p. 12
• Preliminary findings contribute to understanding culturally responsive and relevant human–AI interactions.Source: §9 · p. 12
• The authors call for more basic research on technology’s sociocultural construction and applied research to diversify AI design, aiming to serve a broader range of the world population.Source: §9 · p. 12