Is It Bad to Work All the Time? Cross-Cultural Evaluation of Social Norm Biases in GPT-4

Zhuozhuo Joy Liu; Farhan Samir; Mehar Bhatia; Laura K. Nelson; Vered Shwartz

Source-linked AI summary

Is It Bad to Work All the Time? Cross-Cultural Evaluation of Social Norm Biases in GPT-4

Zhuozhuo Joy Liu, Farhan Samir, Mehar Bhatia, Laura K. Nelson, Vered Shwartz

arXiv:2505.18322v1May 23, 2025cs.CLcs.AI

TL;DR

LLM cultural alignment has often been measured through direct value surveys, but such responses may not reflect behavior in situated narratives. This paper instead evaluates GPT-4 on social norms elicited from culturally grounded movie plots and finds generic norms alongside stereotypes that remain recoverable despite overt avoidance.

Problem
Survey-based evaluations of LLM cultural values provide limited evidence about how models apply those values in real-world scenarios.
Source: §1 · pp. 1–2
Method
The study prompts GPT-4 and annotators from China, India, Iran, and the United States to reason about social norms in culturally grounded movie narratives.
Source: §1 · p. 2
Results
GPT-4 generates norms that are significantly less culture-specific than human-written norms, while stereotypical associations can be recovered through agreement-prediction prompts.
Source: §1 · p. 2, §4.1 · p. 6, §4.3 · pp. 7–8
Takeaways & Limitations
Culturally competent LLMs face a tension between representing culture-specific knowledge and avoiding stereotypes about the same cultures.
Source: §5 · p. 8
The study uses countries as proxies for cultures, evaluates only GPT-4 across four countries, and has relatively few human annotators per culture.
Source: Limitations · p. 9

Abstract

from arXiv · show

LLMs have been demonstrated to align with the values of Western or North American cultures. Prior work predominantly showed this effect through leveraging surveys that directly ask (originally people and now also LLMs) about their values. However, it is hard to believe that LLMs would consistently apply those values in real-world scenarios. To address that, we take a bottom-up approach, asking LLMs to reason about cultural norms in narratives from different cultures. We find that GPT-4 tends to generate norms that, while not necessarily incorrect, are significantly less culture-specific. In addition, while it avoids overtly generating stereotypes, the stereotypical representations of certain cultures are merely hidden rather than suppressed in the model, and such stereotypes can be easily recovered. Addressing these challenges is a crucial step towards developing LLMs that fairly serve their diverse user base.

1 Introduction

The paper examines whether GPT-4’s apparent Western cultural alignment transfers to reasoning about social norms in culturally grounded narratives. It argues that bottom-up evaluation reveals generic norms, latent stereotypes, and tensions between cultural specificity and stereotype avoidance.

Motivation

Western-dominated training data, development choices, and guardrails can embed Western values and biases into LLMs.The authors connect English-heavy web data and typically Western developer choices to concerns about stereotyping and representational bias.Source: §1 · p. 1
Prior survey-based studies find strong North American alignment but may not show whether LLMs apply reported values consistently in real-world situations.LLMs lack a stable persona and are optimized to generate human-like responses, motivating a behavioral evaluation.Source: §1 · pp. 1–2

Approach

The study uses culturally grounded movie narratives from China, India, Iran, and the United States, asking annotators and GPT-4 to identify and evaluate social norms as Rules of Thumb.Annotators assess the resulting norms for accuracy, culture-specificity, and stereotypicality.Source: §1 · p. 2

Contributions

GPT-4 generates norms that are significantly more generic than human-written norms while remaining not necessarily incorrect, and stereotypes avoided during generation can resurface through agreement-prediction prompts.The study presents this pattern as evidence of default representational biases rather than fully suppressed stereotypes.Source: §1 · p. 2

Scope and risks

The work includes potentially offensive examples involving stereotypes, associations, and other harms.This content warning identifies a reader-facing limitation of the material.Source: §1 · p. 2

2 Background

The background distinguishes the paper’s bottom-up norm evaluation from survey-based studies of cultural values and reviews evidence that LLMs align more strongly with Western cultures. It also frames stereotype avoidance as potentially superficial, motivating tests of whether stereotypes remain recoverable.

2 Background

The paper focuses on social norms, cultural alignment, and stereotype reinforcement as dimensions of LLM cultural evaluation.It largely sets aside the distinction between norms and values for this study.Source: §2 · p. 2, §2.1 · p. 2

2.1 Values, Norms, and Morals

Prior norm resources and cultural evaluations have predominantly centered Western norms, while survey-based studies find strong North American alignment and weaker alignment with other Western countries.English prompts without cultural context typically lead models to assume Western or US culture.Source: §2.1 · p. 2
Survey instruments such as HCS, WVS, and PEW directly ask models about values, but such answers may not predict behavior because LLMs lack a consistent persona.The paper therefore prompts GPT-4 to infer norms from existing narratives rather than answer value questions directly.Source: §2.1 · pp. 2–3

2.2 Stereotypes and Cultural Bias

LLMs learn societal biases from web data, while instruction tuning, preference tuning, and guardrails improve avoidance of harmful language without necessarily removing underlying biases.Prior work reports subtle stereotypes and superficial refusal behavior in some prompting setups.Source: §2.2 · p. 3
This study tests whether GPT-4 avoids stereotypes when generating cultural norms but predicts agreement with them when simulating people from the relevant country.The contrast directly targets whether stereotype avoidance reflects suppression or merely masking.Source: §2.2 · p. 3

3 Data

The study builds a four-country movie-plot dataset and collects culturally grounded Rules of Thumb from human annotators and GPT-4. Humans are culturally primed and rate norms, while GPT-4 is tested with default and cultural prompts, with culture leakage acknowledged in the default condition.

3.1 Culturally-Grounded Narratives

The dataset uses English Wikipedia movie plots from the United States, China, India, and Iran as culturally grounded narratives for eliciting social norms.The authors retain 20 moderately sized plots per country and report broad thematic variety across family, romance, rural life, politics, and law enforcement.Source: §3 · p. 3, §3.1 · pp. 3–4

3 Data

Figure 2 shows that each plot is given to human annotators and GPT-4, producing human, default-prompt, and cultural-prompt RoTs.Each human annotator writes 3–5 RoTs, while GPT-4 is prompted under two settings, yielding 19–25 RoTs per movie across sources.Source: Figure 2 · p. 4

3.2 Human-written RoTs

Human annotators are recruited from four cultural groups, culturally primed with images, and recruited based on substantial residence in the respective country.Because CloudConnect annotators reside in English-speaking countries, the study uses bicultural annotators and cultural activation before annotation.Source: §3.2 · pp. 4–5

3.3 GPT-generated RoTs

GPT-4 generates five RoTs per movie with few-shot prompting, either without cultural context to estimate default values or with an explicit country-background instruction.The authors acknowledge that plot content can leak cultural information, so default prompting is less than fully culture-agnostic.Source: §3.3 · p. 5

4 Results

The study evaluates GPT-4’s reasoning about culturally grounded social norms, finding that its norms are generally accurate but less culture-specific, its default stance aligns most with the United States, and stereotypes remain recoverable despite surface avoidance.

4 Results

The study compares GPT-4-generated and human-written rules of thumb across accuracy, culture-specificity, and stereotypicality, using narratives from four countries.Source: §4 · p. 5, §4.1 · p. 5

4.1 Which Cultures does GPT-4 Know about?

GPT-4-generated norms were generally as accurate as human-written norms, with no clear advantage from cultural prompting.Accuracy was at least 3.8 across countries; human-written norms were significantly better only for China, by 0.2 points.Source: §4.1 · pp. 5–6, Figure 3 · p. 6
GPT-4-generated norms were less culture-specific than human-written norms, suggesting that its accuracy partly reflects generic norms acceptable across cultures.Human-written norms received higher culture-specificity ratings (β = 0.17, p = .037), and cultural prompting provided no significant advantage.Source: §4.1 · p. 6, Figure 3 · p. 6

4.2 Which Cultures is GPT-4 Aligned with?

GPT-4’s default ratings aligned most closely with the United States and deviated most from China, with average JSD values of 0.12 and 0.33, respectively.Divergence was 0.20 for Iran and 0.24 for India.Source: §4.2 · pp. 6–7, Figure 4 · p. 6

4.3 Does GPT-4 Reinforce Stereotypes?

Human-written norms were more stereotypical than GPT-4-generated norms, but culture-specificity and stereotypicality were positively correlated.The correlation was r = 0.462, p < .001; the authors discuss in-group tolerance and GPT-4’s limited culture-specific knowledge as possible explanations.Source: §4.3 · p. 7, Figure 5 · p. 8

5 Conclusion

The study finds that GPT-4 exhibits default cultural biases when reasoning about culturally grounded social norms. Its representations align most closely with the US, while stereotypes avoided in generated language can still be recovered implicitly.

GPT-4’s latent cultural representation aligns most closely with the US and least with China, with India and Iran between them.Source: §5 · p. 8
Although GPT-4 tends to avoid overt stereotypes, stereotypical representations remain implicitly ingrained and can be resurfaced.Source: §5 · p. 8
The findings expose a design tension between providing culture-specific knowledge and avoiding the perpetuation of cultural stereotypes.Source: §5 · p. 8

Limitations

The study’s conclusions are constrained by its use of countries as cultural proxies, four cultures, one model, and limited annotator diversity. Its English, movie-based setup and bicultural annotators also introduce additional scope boundaries.

The study uses countries as proxies for cultures, although cultural identity does not map neatly onto national boundaries and varies individually.Source: Limitations · p. 9
The evaluation covers only four countries and GPT-4, so broader conclusions require more cultures and models.Source: Limitations · p. 9
English movie plots may reflect Western editorial perspectives or unusual-event reporting, while bicultural annotators may have influenced judgments.Source: Limitations · p. 9

Ethical Considerations

The study describes safeguards for participant safety and data integrity while acknowledging that country-based grouping and cultural priming can simplify or reinforce cultural stereotypes.

Annotators were compensated above local minimum-wage standards and instructed not to include personally identifiable information.Source: Ethical Considerations · p. 9
The study screened movie plots for harmful content before human evaluation to support participant safety and data integrity.Source: Ethical Considerations · p. 9
The study uses country as a cultural proxy despite individual variation in cultural identity and experience.Source: Ethical Considerations · p. 9
Culturally relevant images and culturally grounded norms may inadvertently reflect or reinforce stereotypes.Source: Ethical Considerations · p. 9

A.1 Annotators Demographics

The appendix reports annotator demographics for the RoT collection and evaluation tasks and identifies the interfaces used for cultural activation, collection, and evaluation.

A.1 Annotators Demographics

RoT collection included 88 annotators across four groups: US 44, India 23, Iran 13, and China 10.Source: §A.1 · p. 13
Collection annotators ranged from 18 to 62 years, while evaluation participants ranged from 18 to 66 years.Source: §A.1 · p. 13
RoT evaluation included 56 participants across four groups: US 25, China 14, Iran 10, and India 7.Source: §A.1 · p. 13

A.2 Annotation Interface

Figures 6–8 show the interfaces for cultural activation, RoT collection, and RoT evaluation using Iran as the example culture.Source: §A.2 · p. 13

B LLM Prompts

The appendix describes interfaces and prompts for collecting culturally grounded Rules-of-Thumb and evaluating their accuracy, cultural specificity, and stereotypicality. The workflow combines cultural priming, culture-specific narratives, structured norm generation, and culturally grounded ratings.

B.1 RoT Collection

The RoT collection prompt operationalizes the task through a structured instruction for producing culturally grounded norms.Source: Table 6 · p. 13
Annotators first complete a cultural activation task using five culturally relevant images before writing social norms.Source: Figure 6 · p. 14
After cultural activation, annotators generate 3–5 Rules-of-Thumb from a movie plot associated with their culture.Source: Figure 7 · p. 15
The collection interface highlights story action verbs and anchors each norm statement to a specific action.Source: Figure 7 · p. 15
Collection instructions require cultural alignment, verb specificity, and a clear judgment-action format for norm statements.Source: Table 6 · p. 13, Figure 7 · p. 15

B.2 RoT Evaluation

For evaluation, annotators rate each Rule-of-Thumb on accuracy, cultural specificity, and stereotypicality.Source: Figure 8 · p. 16, Table 7 · p. 17
Evaluators repeat the cultural priming task before rating norms from each cultural background.Source: Figure 8 · p. 16
The evaluation prompt varies the country or cultural background through an explicit country-specific placeholder.Source: Table 7 · p. 17

C Example RoTs

The appendix presents selected stereotypical Rules-of-Thumb for each culture alongside human evaluations. These examples are assessed for accuracy, cultural specificity, and stereotypicality.

The examples are evaluated across accuracy, cultural specificity, and stereotypicality.Source: C Example RoTs · p. 17, Table 8 · p. 18
Table 8 links representative stereotypical norms to the corresponding human assessment criteria.Source: C Example RoTs · p. 17
Selected stereotypical Rules-of-Thumb are presented for each culture with individual annotator ratings.Source: Table 8 · p. 18