Grafy News
Try Grafy free

AI search engines favor global narratives over Latin American news

Research shows LLM-based search engines rely less on local and diverse news from Latin America than traditional search engines, risking regional bias.

Illustration for the news story: AI search engines favor global narratives over Latin American newsSource: Grafy Research Lab

Large language model (LLM) search engines may be narrowing the window of information available to users in the Global South by prioritizing global narratives over local news. A recent study suggests that as AI platforms become primary gateways for current events, they risk amplifying geographical biases that favor wealthier nations at the expense of regional diversity.

How does AI search compare to traditional search engines?

According to research conducted by Francisco W. Kerche, Franziska Sofia Hafner, Felix M. Simon, and Luc Rocher, there is a measurable difference in how information is retrieved when moving from conventional search engines to LLM-based search. The researchers focused on Latin America, a region that represents the largest share of AI usage within the Global South.

To test this, the team constructed a dataset containing 1,901 current events spanning nine different Latin American countries. They then submitted realistic queries to three different search engines to observe how the resulting information differed. The findings indicate that LLM-based search engines tend to utilize less local and diverse news sources compared to traditional search engines, which may lead to a more homogenized view of regional events.

Why is this bias a concern for news consumption?

As AI-driven interfaces become the preferred method for many to stay informed, the selection of sources used to ground these models becomes critical. The researchers suggest that if LLMs rely heavily on a subset of mainstream or international news, they may overlook the nuances and specificities of local reporting.

This shift could create a feedback loop where narratives from more prominent, wealthier countries dominate the AI's output. For users in Latin America, this means the information provided by an AI agent might not fully reflect the breadth of local perspectives or the specificities of regional developments, potentially leading to a distorted understanding of local reality.

What were the specifics of the research methodology?

The study employed a rigorous approach to quantify this information gap. By selecting 1,901 distinct current events across nine countries, the authors were able to create a representative sample of the news landscape in the region.

By crafting queries that mimicked how actual users interact with search tools, the researchers could compare the "information density" and the diversity of the sources cited by traditional search engines versus those synthesized by LLMs. The study's goal was to determine if the transition to AI-mediated search inherently alters the geographical distribution of the information being presented to the user.

How can researchers mitigate information bias in AI?

The research highlights a significant challenge for the development of retrieval-augmented generation (RAG) and search-integrated AI. To prevent the marginalization of local news, models must be able to effectively identify and prioritize high-quality, regional sources that might otherwise be drowned out by global media giants.

Addressing this requires not just better models, but a more diverse training and retrieval set that explicitly includes local journalistic outputs. Without these interventions, the convenience of AI search may come at the cost of a more diverse and accurate global information ecosystem.

How can users explore these research trends?

For researchers and students looking to dive deeper into the complexities of AI bias and information retrieval, tools that allow for direct interrogation of academic papers can be invaluable. Using Reader, you can upload complex studies like this one to search for specific data points, ask questions about the methodology, and synthesize findings across multiple documents. This allows for a more critical and granular examination of how AI models handle diverse datasets.

Frequently asked questions

Do AI search engines use less local news than traditional engines?

Yes, according to the research, LLM-based search engines tend to rely on a less diverse and less local set of news sources when compared to conventional search engines. This trend is particularly evident when searching for current events within the Latin American region.

Why does Latin America matter in this study?

Latin America is a critical focal point because it accounts for the largest share of AI usage in the Global South. Because the region is a major consumer of these technologies, the potential for geographical bias in AI-driven news consumption is particularly significant there.

What is the risk of using LLMs for news retrieval?

The primary risk is the amplification of existing geographical biases. By favoring narratives from wealthier countries and overlooking local, diverse news, AI platforms may provide users with a limited or skewed perspective of events occurring in the Global South.


Source: Grafy Research Lab — Web search in Latin America: language models use less local and diverse news than search engines