Large language models (LLMs) offer vast computational power yet consistently overlook cultural nuance. While often praised for bridging language gaps, recent research highlights a deeper issue: LLMs largely reflect Anglophone and Western European cultural values, which are embedded in the English-dominated data that shapes them. This review synthesizes findings from 2020 to 2025 to assess the implications of this technological spread for vulnerable groups, particularly immigrants, refugees, and international students, who must navigate adaptation within English-centric and Western communication norms.

 

Using insights from cultural cognition and identity-protective cognition theories, the analysis identifies five central forms of bias: (1) representational bias that undermines non-Western perspectives; (2) linguistic inequity that amplifies challenges for low-resource languages; (3) authenticity failures, with stereotypes substituting for real cultural understanding; (4) identity erosion as users’ voices are homogenized; and (5) reliance on LLMs that may hinder independent language skill development. This “equity paradox” means that the very systems marketed as democratizing global communication can actually deepen exclusion and sameness among those who are most reliant on them.

 

Ultimately, the review concludes that current governance and policy efforts are insufficient to address the underlying power dynamics that shape LLM development. Authentic cross-cultural communication, the evidence suggests, depends on human qualities absent in LLMs: presence, vulnerability, and the openness to change that underpins accurate understanding. In an AI-mediated world, recognizing the limits of these tools is not a matter of nostalgia, but rather necessary wisdom.