In our increasingly interconnected world, artificial intelligence tools such as ChatGPT have been celebrated for their potential to democratize access to information. However, a recent study by computer scientists at Johns Hopkins University reveals a growing digital language divide that might inadvertently amplify bias. Large language models (LLMs), the backbone of many AI tools, appear to reinforce the dominance of high-resource languages like English, unintentionally marginalizing minority languages and perspectives. Without addressing these disparities, existing global knowledge access gaps may widen.
Presented at the 2025 Annual Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics, the study poses a critical question: Are multilingual LLMs truly dismantling language barriers? Researchers, including Nikhil Sharma, Kenton Murray, and Ziang Xiao, sought answers by analyzing coverage of international conflicts through news articles in both high- and low-resource languages. They found that LLMs typically favor information in the language of the query, often defaulting to English when lower-resource languages lack sufficient data.
This preference for high-resource languages becomes apparent when disparities arise between articles in different languages. For example, if an English article depicts a political figure negatively while a Hindi article portrays the same figure positively, the model’s response tends to align with the language of inquiry. Such issues are exacerbated when articles are nonexistent in the query language, like Sanskrit, causing LLMs to default to English or other high-resource languages, ignoring the original linguistic perspective.
The bias raises concerns of linguistic imperialism, where dominant languages and their associated perspectives overshadow others. A scenario presented by the researchers depicted three users querying the India-China border dispute in different languages, resulting in distinct narratives influenced by the data available in each language.
Addressing this language divide, researcher Chen suggests several measures. Developing more dynamic benchmarks and datasets for LLMs is crucial. Integrating diverse perspectives from multiple languages in the training of models and informing users about potential biases can reduce reliance on a singular viewpoint. Additionally, enhancing conversational search literacy can mitigate the risks of overtrusting AI outputs.
In summary, the study points out a significant challenge in AI development: ensuring equitable access to global knowledge across linguistic boundaries. Failure to correct the inherent bias in multilingual AI tools may lead to a few high-resource languages dominating the flow of information, skewing perceptions and decision-making. As Sharma emphasizes, crafting AI systems that deliver an unbiased and diverse array of perspectives is essential for fostering an informed global citizenry. These efforts are key to creating a digital information landscape that genuinely respects and represents our world’s linguistic diversity.