Reproduce “AI Suggestions Homogenize Writing Toward Western Styles and Diminish Cultural Nuances” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Dhruv Agarwal, Mor Naaman, Aditya Vashistha

The authors tested GPT-4o inline autocomplete with 118 Indian and American participants writing culturally grounded English essays. Suggestions followed a typing pause and could be accepted or dismissed. Both groups finished faster while cross-cultural essay similarity rose from 0.48 without AI to 0.54 with it. The authors interpret this as a Westernward pull limited to this model, sample, and task.

Key insights

  • AI assistance was useful for both groups, but Indian participants accepted about 25% of suggestions versus 19% for Americans, relied on more generated text, and modified suggestions more frequently; productivity gains per accepted suggestion were nevertheless lower.
  • Cross-cultural essay similarity increased from 0.48 to 0.54 with AI, while an authorship classifier’s accuracy fell from 90.6% on unaided writing to 83.5% on AI-assisted writing, including in the implicit workplace-email task.
  • Indian descriptions assisted by AI more often replaced specific practices, ingredients, and histories with generic language; early suggestions also favored culturally incongruent items such as pizza, sushi, and Christmas.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.