[논문 리뷰] AI Suggestions Homogenize Writing Toward Western Styles and Diminish Cultural Nuances

D. P. Agarwal, Mor Naaman외 1명|arXiv (Cornell University)|2024. 09. 17.

Computational and Text Analysis Methods인용 수 11

한 줄 요약

교차문화 연구는 서구 중심의 AI 글쓰기 제안이 미국인들에게 더 큰 생산성 증가를 가져오고 인도인들의 글쓰기를 서구 스타일로 균질화하여 문화적 뉘앙스를 감소시킨다.

ABSTRACT

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.

연구 동기 및 목표

• 서 Western-centric AI writing suggestions가 다른 문화(인도 vs. 미국) 사용자의 영향을 어떻게 미치는지 평가한다.
• 문화 간 AI 제안에 대한 생산성 증가 및 참여를 측정한다.
• 비서구권 글쓰기가 서구 스타일로 동질화되는지 examined.
• AI에 서구 규범을 내재화하는 문화적 해를 논의하고 완화 전략을 제안한다.

제안 방법

• 2x2 between-subjects experiment with 118 participants (60 Indian, 58 American).
• Conditions: AI suggestions via GPT-4o vs No AI, crossed with participant culture (Indian vs American).
• Four writing tasks designed from Hofstede’s Cultural Onion to elicit explicit and implicit cultural aspects.
• Inline autocomplete suggestions from GPT-4o shown after typing pauses (100 ms); TAB accepts, ESC rejects; logging of interactions.
• Metrics include AI reliance, suggestion acceptance rate, suggestion modification, and writing productivity; NLP measures include Type-Token Ratio and cosine similarity of embeddings.
• Prompts and data collection approved by IRB; demographic survey included to confirm cultural differences (SSVS).

Figure 1. A representation of the potential cultural homogenization from Western-centric AI models

실험 결과

연구 질문

• RQ1 RQ1: Does writing with a Western-centric AI provide greater benefits to users from Western cultures compared to those from non-Western cultures?
• RQ2 RQ2: Does writing with a Western-centric AI homogenize the writing styles of non-Western users toward Western styles?

주요 결과

• AI boosts productivity for both Indian and American participants, with larger gains for Americans.
• AI causes Indian participants to write more like Americans, indicating Western-style homogenization.
• Indian participants showed higher engagement with AI suggestions (AI reliance) and accepted a larger share of suggestions than Americans.
• A total of 12,015 AI suggestions were shown, with 1,476 accepted (12.3%); 0.6% of rejections were explicit via Escape, while most rejections occurred during flow.
• Top autocomplete suggestions reveal a Western bias, especially for food and festival tasks (Table 4).
• Eight of ten Schwartz values differed significantly between Indian and American participants, supporting cultural distance.