TLDR: A study using the 20 Questions game reveals that Large Language Models (LLMs) exhibit significant implicit geographic biases, performing much better at identifying entities from the Global North and West compared to the Global South and East. These disparities are not fully explained by entity popularity or language of interaction, suggesting biases are deeply embedded in their reasoning processes.
Large Language Models, or LLMs, have become incredibly sophisticated, but a recent study sheds light on a subtle yet significant issue: implicit geographic biases embedded within their reasoning processes. While much effort has gone into making these models explicitly fair and unbiased, this research suggests that deeper, hidden prejudices can still influence how they understand and deduce information about the world.
The paper, titled “The World According to LLMs: How Geographic Origin Influences LLMs’ Entity Deduction Capabilities,” introduces a novel approach to uncover these biases. Instead of directly asking LLMs questions that might trigger their built-in safeguards, the researchers observed how models behave when they proactively ask questions themselves, much like in the classic 20 Questions game.
A New Way to Test LLM Biases
To conduct their study, authors Harsh Nishant Lalai, Raj Sanjay Shah, Jiaxin Pei, Sashank Varma, Yi-Chia Wang, and Ali Emami developed a new dataset called Geo20Q+. This dataset is unique because it includes a wide range of notable people and culturally significant objects (like foods, landmarks, and animals) from diverse regions across the globe. They then set up a game where one LLM acts as a “guesser” trying to identify an unknown entity by asking a series of yes/no/maybe questions, and another LLM acts as a “judge” providing answers based on its knowledge of the target entity.
The researchers tested popular LLMs, including GPT-4o-mini, Gemini-2.0-Flash, and Llama-3.3-70B-Instruct, under two game configurations: a traditional 20-question limit and an unlimited-turn setting. They also evaluated the models in seven different languages: English, Hindi, Mandarin, Japanese, French, Spanish, and Turkish, to see if language played a role in these biases.
Key Findings: A World Divided
The results revealed a clear and consistent pattern of geographic disparity. LLMs were significantly more successful at deducing entities originating from the Global North (economically developed regions like North America and Europe) compared to the Global South (less economically dominant regions across Africa, Latin America, South, and Southeast Asia). Similarly, they performed better on entities from the Global West (regions with predominantly Western cultural traditions) than from the Global East (regions with distinct non-Western cultural traditions).
For instance, models found it much easier to deduce entities like the Eiffel Tower or LeBron James, while struggling with counterparts such as the Taj Mahal or Jack Ma. This suggests a systemic favoritism towards more economically and epistemically dominant regions in the models’ reasoning processes.
Interestingly, the study found that these disparities could not be fully explained by how popular an entity is (measured by Wikipedia pageviews) or how frequently it appeared in the models’ pre-training data. While these factors had a mild correlation, they didn’t account for the substantial performance gaps. This indicates that the biases are more fundamental than just data availability.
Another significant finding was that the language in which the game was played had minimal impact on these performance gaps. Even when playing in a language associated with a specific region, models did not show a systematic advantage for entities from that region. For example, Hindi, primarily spoken in Asia, did not improve accuracy for Asian entities compared to English.
The research also noted that LLMs generally performed better when deducing “Notable people” than “Things,” and that allowing unlimited turns significantly improved success rates across all conditions, albeit requiring more questions.
Also Read:
- AI’s Hidden Hand: How Language Models Show Gender Bias in Academic Citations
- Local Truths for Global AI: The Concept of Geo-Alignment
Implications for AI Fairness
This research highlights the value of creative, free-form evaluation frameworks that go beyond simply examining model outputs. By analyzing how LLMs initiate and pursue reasoning goals over multiple turns, the study uncovered subtle geographic and cultural disparities embedded in their reasoning processes that might remain hidden in standard prompting setups.
The findings suggest that even with extensive alignment efforts, implicit biases can persist in LLMs, influencing how they perceive and process information about different parts of the world. This calls for continued research into understanding the sources of these biases and developing more inclusive and robust AI systems. You can read the full research paper here.
