Language models reduce news variety and national outlets' visibility for users in Latin America

Language models are becoming a major gateway to news consumption; however, this increase risks amplifying existing biases that favour narratives from wealthier countries over those from the Global South, something particularly relevant in Latin America, since the region presents the largest share of AI usage in the Global South. To study the inequality of information availability, we construct a dataset of 1,901 current events across nine Latin American countries and query four language models and three search engines under four scenarios varying prompt language (national vs English) and query location (in-country vs United States), with a total of 31,443 queries and 40,382 cited sources. We find that language models reduce news-source diversity, retrieve domestic news less often than search engines when users search in their own language and region, and under-cite domestic sources relative to what they actually retrieve. Together, these results show that AI-mediated search amplifies existing global information hierarchies, in which countries in the Global South have lower visibility for their own sources and narratives, while often relying on international narratives because technical systems mediate knowledge circulation.