Source-linked AI summary
AI Suggestions Homogenize Writing Toward Western Styles and Diminish Cultural Nuances
Dhruv Agarwal, Mor Naaman, Aditya Vashistha
TL;DR
The paper examines whether Western-centric embedded writing suggestions serve users across cultures differently and whether they shift non-Western writing toward Western styles. In a controlled experiment with Indian and US participants, AI yielded greater productivity gains for Americans and homogenized Indian writing toward Western norms, diminishing cultural nuance.
Problem
Prior work had not examined cultural clashes in embedded AI applications, whose models can prioritize Western norms while users cannot prompt or fine-tune them to fit cultural preferences.
Source: §1 · p. 1Method
A controlled experiment compared culturally grounded writing by 118 Indian and US participants randomly assigned to GPT-4o autocomplete or no AI.
Source: §1 · p. 2, §6 · p. 15Results
AI suggestions gave Americans greater productivity gains and homogenized Indian writing toward Western styles, diminishing cultural nuance in content and expression.
Source: §1 · p. 2, §5 · p. 14, §6 · p. 15Takeaways & Limitations
The authors warn that accepted biased suggestions could create a feedback loop in which Western-conforming online content further perpetuates model biases.
Source: §5.3 · p. 15The study’s scale was limited by the small pool of Indian Prolific participants, and its findings need testing across more cultures and larger participant pools.
Source: §5.4 · p. 15
Abstract
from arXiv · showhide
Large language models (LLMs) are being increasingly integrated into everyday products and services, such as coding tools and writing assistants. As these embedded AI applications are deployed globally, there is a growing concern that the AI models underlying these applications prioritize Western values. This paper investigates what happens when a Western-centric AI model provides writing suggestions to users from a different cultural background. We conducted a cross-cultural controlled experiment with 118 participants from India and the United States who completed culturally grounded writing tasks with and without AI suggestions. Our analysis reveals that AI provided greater efficiency gains for Americans compared to Indians. Moreover, AI suggestions led Indian participants to adopt Western writing styles, altering not just what is written but also how it is written. These findings show that Western-centric AI models homogenize writing toward Western norms, diminishing nuances that differentiate cultural expression.
1 INTRODUCTION
This section motivates studying cultural clashes in embedded AI writing tools, then reports that the tested suggestions benefited Americans more and shifted Indian writing toward Western styles. It describes a cross-cultural experiment designed to examine those differences.
1 INTRODUCTION
- Prior research documents Western-prioritizing models and cultural harms, but had not examined cultural clashes in embedded AI applications.Unlike chat-based tools, embedded applications do not let users prompt or fine-tune models to suit their cultural preferences.Source: §1 · pp. 1–2
- The study asks whether a Western-centric AI gives greater benefits to Western users and whether it homogenizes non-Western users’ writing toward Western styles.The paper also discusses potential harms and proposes strategies to address cultural imperialism and linguistic singularity.Source: §1 · p. 2
- The experiment compared writing by 118 Indian and US participants randomly assigned to GPT-4o autocomplete or a no-AI condition.Prompts based on Hofstede’s Cultural Onion framework elicited practices ranging from explicit food to implicit rituals.Source: §1 · p. 2
- AI suggestions brought greater productivity gains for Americans and shifted Indian writing toward American styles, diminishing cultural nuance in both content and expression.The reported examples include Indian participants describing their own food and festivals from a Western gaze.Source: §1 · p. 2
2 RELATED WORK
The related work reviews cultural bias in AI and writing-support tools. It positions this study as an examination of cultural effects from embedded writing suggestions, beyond prior work focused on other biases or language proficiency.
2.1 Cultural Bias in AI Models
- Prior scholarship finds generative AI centers Western norms, raising concerns about cultural erasure, stereotypes, and systems that inadequately serve non-Western regions.The section connects these patterns to data colonization and AI colonialism, and notes the challenge of mitigating bias when embedded models are obscured.Source: §2.1 · p. 2
2.2 AI-Based Writing Support
- Writing support has evolved from efficiency tools to active collaborators, while studies report context-dependent efficiency gains and possible costs to satisfaction and text ownership.Suggestions can trade efficiency against ideation: short suggestions may help efficiency, while longer ones may inspire but distract.Source: §2.2 · p. 3
- This study extends biased-language-model writing research with a cross-cultural analysis, unlike prior comparisons that did not control for cultural differences.The closest cited work compared native and non-native English speakers using a coarse-grained proficiency distinction.Source: §2.2 · p. 3
3 METHODOLOGY
The study compared writing by Indian and American participants with and without AI suggestions, using culturally grounded tasks and measures of writing behavior and essay content. Its methods combined interaction logs, text analysis, participant surveys, and content analysis.
3.1 Experiment Design
- The 2 × 2 experiment assigned 118 participants—60 from India and 58 from the US—to write with or without AI suggestions.The researchers treated the US as representing a smaller cultural distance from the Western-aligned model and India as a larger one.Source: §3.1 · p. 3, §3 · p. 3, §3.4 · p. 5
3.2 Writing Tasks
- Four writing tasks used Hofstede’s Cultural Onion to elicit cultural practices and values, with one task for each layer.Prompts drew on representative cultural artifacts and a digital interaction designed to evoke underlying values; an attention check was also included.Source: §3.2 · pp. 3–4, Table 2 · p. 4
3.3 Study Procedure
- Participants completed the tasks sequentially in a study portal after consent; AI-condition participants received a tutorial on accepting and rejecting suggestions.Each task required at least 50 words, and participants completed a demographic survey after the writing tasks.Source: §3.3.1 · p. 4, §3.3 · p. 4
3.3.2 Retrieving Autocomplete Suggestions.
- GPT-4o suggestions appeared inline after a typing pause, while participants could accept, reject, or ignore them.The researchers used a neutral prompt because preliminary tests found Western defaults but adaptation to Indian context when the leading text supplied it.Source: §3.3.1 · p. 4, §3.3.2 · pp. 4–5
3.3.3 Demographic Survey.
- The survey collected demographic and cultural information; Indian and American participants differed significantly on eight of ten Short Schwartz Value Survey values.Participants were included only if born in and resident throughout their lives in the same country; the survey helped triangulate cultural background.Source: §3.3.3 · p. 5, §3.4 · pp. 5–6
3.5.1 Interaction Logs.
- Interaction logs captured task time and suggestion use, including acceptance rate and AI reliance based on retained AI-suggested essay characters.Longest-common-subsequence matching identified retained suggestion text in the final essay.Source: §3.5.1 · p. 6
3.5.2 NLP Metrics.
- Essay analysis measured lexical diversity with Type-Token Ratio and compared essays using embedding-based cosine similarity, alongside manual coding of cultural content.The manual coding examined cultural symbols and idols, exoticization, misrepresentation, and Western gaze; similar essay lengths were expected to limit TTR length bias.Source: §3.5.2 · pp. 6–7
3.6 Ethics and Positionality
- The researchers obtained institutional review approval and adapted prompt wording to cultural context, including using “festivals” alongside “holidays” for Indian participants.The team reported members with experience living in both India and the US as context for study design and interpretation.Source: §3.6 · p. 7
4 FINDINGS
Among participants who received AI suggestions, 12.3% were accepted, and acceptance increased as tasks progressed. The paper also reports a Western bias in suggestions for food and festival tasks.
4 FINDINGS
- 472 essays came from 118 participants who each completed four tasks; 65 participants received AI-generated writing suggestions.The suggestion group included 36 participants in India and 29 in the US.Source: §4 · p. 7
- 12.3% of 12,015 AI suggestions were accepted; 68.8% of rejected suggestions were dismissed within 500 ms, while 0.6% were explicitly rejected with the escape key.The remaining rejected suggestions were dismissed by continuing to type.Source: §4 · p. 7
- The most common suggestions showed a noticeable Western bias in the food and festival tasks.The paper points to Table 4 for examples of the model's most common suggestions.Source: §4 · p. 7
- Acceptance rose from 9.2% in the first task third to 12.1% in the second and 15.9% in the final third.The authors suggest that suggestions may become more useful as the model receives more context.Source: §4 · p. 7
4.1 Does AI impact Indians and Americans differently?
AI helped both Indian and American participants write faster, but Americans derived more productivity per suggestion. Indians engaged with suggestions more and modified them more often.
4.1.1 Engagement with AI Suggestions.
- Indians showed higher AI reliance and acceptance than Americans: mean reliance was 0.53 versus 0.42, and acceptance was 25% versus 19%.Both differences were statistically significant, with small effect sizes.Source: §4.1.1 · p. 7
4.1.2 Suggestion Modification Behavior.
- Indians modified suggestions in 63.5% of tasks, compared with 59.4% for Americans, indicating that accepted suggestions more often needed adjustment.The authors note that suggestions showed Western bias for food and festival prompts.Source: §4.1.2 · p. 7, Table 4 · p. 8
4.1.3 Productivity Derived from AI Suggestions.
- Both cohorts wrote faster with AI, but Indians derived significantly less productivity per suggestion than Americans, despite higher measured productivity gains.The per-suggestion comparison was limited to essays with up to seven accepted suggestions because few Americans accepted more than seven.Source: §4.1.3 · p. 8, Figure 5 · p. 9
4.2 Does AI homogenize writing within cultures?
AI increased writing similarity within both cultural cohorts. The increase was statistically similar across cohorts, on top of a slightly higher baseline similarity among Indian participants.
4.2.1 Similarity Analysis.
- Without AI, Indian participants had slightly higher within-culture essay similarity than Americans, though the effect size was negligible.The difference was statistically significant (t(2726) = 4.04, p < 0.001; Cohen’s d = 0.16).Source: §4.2.1 · p. 9
- AI significantly increased within-culture similarity for both cohorts, with the increase statistically similar for Indians and Americans.The authors suggest that common phrasing, structure, or content may contribute to this homogenizing effect.Source: §4.2.1 · p. 9
4.3 Does AI homogenize writing across cultures?
AI increased similarity between Indian and American writing, with evidence that Indian writing moved more toward American styles. Classification and lexical-diversity analyses suggest the shifts extended beyond explicit cultural references or grammar correction.
4.3.1 Similarity Analysis.
- 0.48 without AI versus 0.54 with AI mean cross-culture similarity; convergence was stronger for Indians toward American styles (d = −0.22) than in reverse (d = −0.12).The cross-culture increase was larger than the within-culture increases; the authors connect the asymmetric direction to Western model biases.Source: §4.3.1 · pp. 9–10, Figure 6 · p. 10
4.3.2 Predictive Modeling.
- Accuracy fell 7.1 percentage points (90.6% to 83.5%) and macro F1 fell 7 points (89.9% to 82.9%) when classification used AI-written essays.The model found Indian versus American authorship harder to distinguish in the AI condition.Source: §4.3.2 · pp. 10–11, Figure 7 · p. 11
4.3.3 Did AI cause deep writing changes or merely omit explicit
- Classification declines persisted across embedding dimensions and were larger on the implicit email task, supporting deeper style shifts beyond omitted explicit cultural markers.For the email task, accuracy fell from 82.9% without AI to 60% with AI, and F1 fell from 81.7% to 49.2%.Source: §4.3.3 · p. 11
4.3.4 Did AI homogenize Indian writing solely by correcting gram-
- AI reduced grammatical errors in both cohorts, but the reduction was significant only for American essays, so grammar correction alone did not explain homogenization.Source: §4.3.4 · p. 11, Figure 8 · p. 12
4.3.5 A Concrete Example: Altering Lexical Diversity.
- AI raised Indian essays’ lexical diversity toward American levels, eliminating the no-AI difference; the figure reports no change for Americans.Lexical diversity was measured using type-token ratio.Source: Figure 9 · p. 12, §4.3.5 · p. 12
4.4 Impact of AI on Cultural Writing
The cultural-writing analysis found that AI suggestions often replaced specific Indian cultural expression with generic descriptions and Western defaults. It also examined how suggestions related to participants’ artifact choices, though the study could not observe what AI-condition participants would have written without AI.
4.4 Impact of AI on Cultural Writing
- The researchers manually coded Indian essays on food, festivals, and public figures, comparing the same festivals and foods across AI and No AI conditions.They examined cultural symbols and dimensions including exoticization, misrepresentation, and Western gaze.Source: §4.4 · p. 12, §4.4.1 · p. 12
4.4.1 Changing Cultural Expression.
- AI-assisted Diwali essays omitted specific rituals, while Biryani descriptions lost regional details and used generic, sometimes exoticizing tropes.The authors note these descriptions were not incorrect or overtly misrepresentative; the public-figure comparison was inconclusive because few figures overlapped across conditions.Source: §4.4.1 · pp. 12–13
4.4.2 Cultural Incongruence in AI Suggestions.
- Indian participants’ first celebrity suggestions were almost exclusively Western and never matched the figure ultimately named; food and festival defaults were pizza or sushi and Christmas.Culturally incongruent celebrity suggestions sometimes persisted as participants typed their intended figure.Source: §4.4.2 · p. 13
4.4.3 Did AI suggestions influence participants’ choice of artifacts?
- 54% of Indian participants chose Indian public figures without AI, compared with 44% with AI; Christmas essays rose from 12.5% to 13.9%.The authors suggest these observations may indicate changing cultural preferences, but note small sample size and that participants often bypassed first suggestions.Source: §4.4.3 · pp. 13–14
5 DISCUSSION
The discussion argues that AI suggestions bring greater productivity gains to Americans while homogenizing Indian writing toward Western styles, with both explicit and subtle cultural effects. It calls for culturally plural design and notes limits to generalizing from this study’s country categories, participant pool, and between-subjects design.
5 DISCUSSION
- AI suggestions yield greater benefits for Americans and homogenize Indian writing toward American styles, affecting cultural preferences, lexical diversity, exoticization, and Westernization.The discussion notes that these changes range from explicit to subtle and may be harder to identify.Source: §5 · p. 14
5.1 Cultural Differences in AI Attitudes
- Indian participants engaged more with AI suggestions, while the authors frame cross-cultural differences in trust and reliance as part of human-AI interaction, not merely confounders.They warn that embedded applications offering diminished benefits may inadvertently push non-Western users toward overreliance.Source: §5.1 · p. 14
5.2 Harms of Cultural Homogenization
- Homogenization may reinforce Western cultural hegemony and feed future models with more Western-conforming content; the authors also identify risks to storytelling and culturally distinct formality.They describe these as potential broader harms and call for further interdisciplinary study of subtle shifts.Source: §5.2 · pp. 14–15, §5.3 · p. 15
5.3 Call for Mitigation Strategies
- The authors recommend cultural plurality through model-level mitigation and application-level intentional friction, such as tailoring suggestions to users’ writing samples or sociocultural context.They argue that user control and configuration can help avoid generalizing to Western styles.Source: §5.3 · p. 15
5.4 Limitations and Future Work
- Generalizability is limited: India and the US are broad geographic proxies, the Indian Prolific pool was small, and the between-subjects design supports group-level rather than individual-level analysis.The authors call for studies across more cultures and larger samples, and for within-subjects validation and evaluation of other productivity dimensions.Source: §5.4 · p. 15
6 CONCLUSION
In a controlled experiment with 118 Indian and American participants, the authors examined culturally grounded writing with and without inline AI suggestions. They found greater productivity gains for Americans and Western-style homogenization in Indian writing, which they describe as reinforcing Western cultural hegemony.
6 CONCLUSION
- The study randomly assigned 118 Indian and American participants to culturally grounded writing tasks with inline AI suggestions or no AI support.The authors analyzed writing logs and essays.Source: §6 · p. 15
- AI suggestions offered greater productivity gains to American users, leaving non-Western users to put in more effort to achieve similar benefits.Source: §6 · p. 15
- AI homogenized writing toward Western norms, subtly influencing Indians to adopt Western writing patterns and reinforcing Western cultural hegemony.The authors argue biases should be addressed before culturally biased content proliferates into future models’ training data.Source: §6 · p. 15
A PARTICIPANT DEMOGRAPHIC DETAILS
The section directs readers to Table 6 for participant demographic details.
A PARTICIPANT DEMOGRAPHIC DETAILS
- The section points readers to Table 6 for participant demographic details.Source: A PARTICIPANT DEMOGRAPHIC DETAILS · p. 18
B FAVORITE CELEBRITIES SELECTED BY PARTICIPANTS
The passage directs readers to Table 7.
B FAVORITE CELEBRITIES SELECTED BY PARTICIPANTS
- The passage directs readers to Table 7.Source: B FAVORITE CELEBRITIES SELECTED BY… · p. 18
C CHANGING CULTURAL EXPRESSION: ADDITIONAL EXAMPLES
The section presents examples comparing writing with and without AI, including descriptions of dosa and other cultural artifacts.
C CHANGING CULTURAL EXPRESSION: ADDITIONAL EXAMPLES
- Table 6 lists participants’ self-disclosed languages and occupations.Source: Table 6 · p. 19
- Table 7 lists public figures chosen by Indian participants across the AI and no-AI conditions.Blue cells indicate Indian public figures, while yellow cells represent non-Indian public figures.Source: Table 7 · p. 20, C CHANGING CULTURAL EXPRESSION:… · p. 20
- The without-AI dosa description names ghee masala dosa and its accompaniments, while the with-AI example describes dosa more generally.The with-AI example mentions fillings, chutneys, and sambar; the accompanying legend marks some cultural details as missing and identifies generic or Westernized descriptions.Source: C CHANGING CULTURAL EXPRESSION:… · p. 21
- Table 8 provides additional comparisons of cultural artifacts described by Indian participants with and without AI suggestions.Its highlights mark cultural details; the legend also identifies details missing with AI, generic AI phrasing, and Westernized descriptions.Source: Table 8 · p. 21, C CHANGING CULTURAL EXPRESSION:… · p. 21