Who Anchors AI Overviews in Health? Baidu, Google, and the Geography of Authority
This study audits Google and Baidu's AI Overviews across 12 countries and four languages, revealing that both platforms prioritize their own ecosystems over diverse sources, exhibit geographic and linguistic biases in local citations, and apply health disclaimers inconsistently, thereby highlighting the urgent need for culturally aware oversight of generative search in health contexts.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Every day, billions of people turn to search engines to find answers, but when those answers concern health, the stakes are life and death. For decades, search results have looked like a list of links, requiring a user to click through to find the truth. Recently, however, a new kind of search has emerged, powered by artificial intelligence. Instead of a list, these systems generate a single, synthesized answer right at the top of the screen, pulling information from across the web to explain a topic in plain language. This shift promises to save time, but it also concentrates immense power in the hands of the companies that build these tools. If an artificial intelligence decides which facts to highlight and which to ignore, it can subtly shape how people understand their bodies, their treatments, and their choices. The question is no longer just about finding information, but about who gets to decide what that information looks like.
A team of researchers at Dartmouth College set out to investigate how these new AI summaries handle health questions, specifically looking at the two giants of the internet: Google, which dominates most of the world, and Baidu, the primary search engine for China. They wanted to see if the answers a person receives depend on which company they use, where they live, or what language they speak. To do this, they ran nearly two thousand health queries across twelve different countries and four languages. They asked questions about everything from common conditions like diabetes to more complex or culturally specific topics like Traditional Chinese Medicine. By comparing the results side-by-side, they mapped out where the information came from and how the systems decided what to show.
The researchers found that both Google and Baidu tend to route users toward their own corporate ecosystems rather than a diverse mix of independent sources. On Google, the most frequently cited source for health answers was YouTube, a video platform owned by Google's parent company. This was surprising because established medical authorities like the Mayo Clinic or the World Health Organization appeared less often in these summaries. On Baidu, the top sources were its own encyclopedia and health platforms, along with a specific independent medical site called Bohe. In both cases, the AI was not pulling from a wide, neutral pool of the internet's best medical advice. Instead, it was leaning heavily on content that lived within the company's own digital walls. This creates a situation where the authority of the answer depends less on the medical credentials of the source and more on the algorithms that decide which videos or articles to surface.
Geography played a massive role in what information people received. When the researchers asked the same health questions in English from different countries, the results varied wildly based on location. Users in large, established markets like the United States, the United Kingdom, Australia, and Canada received answers that cited local health institutions. However, people in smaller or less localized markets, such as Nigeria, India, or Singapore, were far more likely to receive answers citing American or international organizations, even when they were asking about local health issues. The study showed that smaller countries received significantly fewer references from their own domestic health systems. It appears that for these regions, the AI defaults to a global standard, potentially overlooking local experts, regulations, or cultural nuances that matter to the people living there.
Language proved to be an even stronger switch than geography. When the researchers asked the same questions in a country's official language instead of English, the answers changed dramatically. In countries like Japan, South Korea, and Malaysia, switching from English to the local language increased the share of locally sourced citations by a factor of three to thirteen. This means that a person searching for health advice in their native tongue is much more likely to get information from their own country's health system, while searching in English often routes them away from local resources. This suggests that the language a person chooses to type can inadvertently determine whether they see advice from their local doctor or a distant international authority.
The study also looked at how these systems handled sensitive or culturally specific topics, using Traditional Chinese Medicine as a test case. Here, the two platforms diverged in their approach. Google was less likely to generate an AI summary for these queries at all, perhaps showing caution about topics where medical consensus is debated. When it did answer, it provided more citations. Baidu, on the other hand, was more willing to generate an answer and was more supportive of Traditional Chinese Medicine practices, yet it also included clearer safety warnings. This indicates that the AI's "caution" is not a single setting but a complex mix of tone, the number of sources cited, and the willingness to generate an answer in the first place. The researchers noted that while Baidu was more supportive of these cultural practices, it also took the time to warn users about safety, showing that endorsement and caution can coexist in these systems.
Ultimately, this research reveals that the artificial intelligence guiding our health decisions is not a neutral observer. It is shaped by the company that built it, the country it is serving, and the language used to ask the question. The sources it trusts are often the ones it owns or the ones that are most visible in English. For millions of people, this means that the path to health information is not the same for everyone. As these tools become more common, the study suggests that we need to be aware of these hidden biases. The way a search engine is built can determine whether a person sees a local expert or a distant corporation, and in matters of health, that difference can be profound.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.