Do AI writing assistants push non-Western writers toward Western styles?

The study's central claim is that an embedded, Western-centric writing assistant does two things at once. It gives Americans larger productivity gains than Indians, and it moves Indian writers' essays toward Western styles. The abstract puts the second half most directly: "Western-centric AI models homogenize writing toward Western norms, diminishing nuances that differentiate cultural expression." The first half is reported as "the gains are higher for American participants," which the authors read as a quality-of-service harm, since non-Western users "need to put in more effort to achieve similar benefits." The second half reaches past content into form. AI "influences not just what is written (e.g., shifting preferences toward Western cultural artifacts such as food items), but also more ingrained elements of how it's written." The excerpt's example is that Indians "describe their own food and festivals from a Western gaze."

The design carries the claim. The authors treat cultural distance as the manipulated variable: Americans stand in for a smaller distance from a model they describe as often aligned with Western values, and Indians for a larger one. Participants recruited through Prolific were randomly assigned to inline GPT-4o autocomplete or to writing without it, across four tasks drawn from Hofstede's cultural onion, moving from explicit symbols and rituals down to implicit values. Comparing the four groups gives two contrasts: AI against no AI within each country, and Indians against Americans within each condition. The excerpt does not say how the essays were scored. The discussion names "lexical diversity, exoticization, Westernization" as kinds of change observed, and the conclusion cites "writing logs and essays" as its data, but no measure is described.

Against the nearest notes, the excerpt is a cultural-scale case of narrowing that other notes describe in other settings. The metaphor experiment in Does AI assistance homogenize or preserve creative diversity? found that AI ideation shrank the pool of human diversity. This study finds that suggestions narrow the cultural range of phrasing, though the tasks and measures differ, so neither result confirms the other. The gap inventory in Does linguistic alignment work the same way across cultures? names Western-sample dominance as what keeps alignment claims local. This excerpt supplies a non-Western sample, though from two countries only. The sharpest contrast is with Can AI systems learn social norms without embodied experience?. That note treats accurate norm prediction as evidence of cultural competence. This excerpt shows a model that can be accurate about a culture while still pulling a writer's text toward its own defaults. The preference result in Can user preference guide AI writing tool alignment? already argues that preference is a poor target; this excerpt adds that the drift is cultural as well as personal, though it does not test preference.

What the excerpt does not establish is the limit of the claim. It gives no effect sizes, group means, test statistics or sample breakdown beyond the 60 and 58 split, so the size of either gap cannot be judged from it. It does not test GPT-4o's cultural skew. The premise that the model is Western-aligned is taken from prior work. Generalization is limited by the authors' own caveat that "India" and "the US" are proxies for broader cultures, and that "future work is required to determine if our results generalize to other countries and sub-cultures." The conclusion's reading, "concrete evidence of AI colonialism," is the authors' interpretation and goes beyond the experiment's data. What the excerpt supports is narrower. With random assignment, one model, one English writing task set and two national groups, access to AI suggestions was followed by a Western shift in Indian participants' writing and by unequal gains. That is enough to treat cultural drift as a measurable property of embedded writing tools. It is not enough to call it a settled general harm.