AI systems with multiple languages tend to perpetuate existing biases within their programming and interaction design.
At the 2025 Annual Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics, a groundbreaking study was presented, shedding light on the language bias in AI tools. The research team, led by Nikhil Sharma, a PhD student in the Whiting School of Engineering’s Department of Computer Science, Kenton Murray from the Human Language Technology Center of Excellence, and Ziang Xiao, an assistant professor of computer science, highlighted the substantial risks of concentrated power over AI technologies.
The study revealed that the dominance of English in large language model tools (LLMs) can lead to linguistic imperialism, with English being prioritised over other languages. This dominance can create a digital language divide, as AI tools like ChatGPT prioritise information in the language used to query them.
In a hypothetical scenario, a Hindi-speaking user and a Chinese-speaking user would receive answers shaped by their respective sources when asking about the longstanding India-China border dispute. An Arabic-speaking user, on the other hand, would get answers from the American English perspective due to the lack of Arabic documents about the conflict.
The researchers found that when there is no article in the language of the query (for example, Sanskrit), LLMs will generate answers based on information found in higher-resource languages, such as English, ignoring other perspectives. This bias was also evident in the coverage of international conflicts, with AI-generated information being biased. For instance, if an English article states that a political figure is bad, but a Hindi article states they are good, the model will say they are bad if the question is in English, but good if the question is in Hindi.
The team created two sets of fake articles with "truthful" and "alternative," conflicting information, written in high-resource languages like English, Chinese, and German, as well as lower-resource languages like Hindi and Arabic. They used these articles to study the impact of language bias on AI tools.
The researchers label current multilingual LLMs "faux polyglots" that fail to break language barriers, keeping users trapped in language-based filter bubbles. To address this issue, the team plans to build a dynamic benchmark and datasets to help guide future model development to mitigate the information disparity in LLMs.
The researchers encourage the larger research community to look at the effects of different model training strategies, data mixtures, and retrieval-augmented generation architectures. They also recommend collecting diverse perspectives from multiple languages, issuing warnings to users who may be falling into confirmatory query-response behavior, and developing programs to increase information literacy around conversational search to reduce over-trust in and over-reliance on LLMs.
The Leibniz ScienceCampus DiTraRe, which conducted research on artificial intelligence in language processing and studied the impact on information dissemination in different languages, including the dominance of English and other widely spoken languages versus minority languages, supports the findings of this study. The researchers emphasise the importance of users getting the same information regardless of their language and background.
In conclusion, the study underscores the need for a more equitable approach to AI technology, ensuring that all users, regardless of their language, receive accurate and unbiased information. The research team's recommendations provide a roadmap for the AI community to work towards this goal.
