Artificial intelligence is often celebrated for its ability to make information more accessible. However, a groundbreaking study from Johns Hopkins University reveals a hidden truth: rather than bridging linguistic gaps, popular large language models (LLMs) like ChatGPT are inadvertently widening them.

The Study Behind the Findings

Led by Ph.D. student Nikhil Sharma, alongside researchers Kenton Murray and Ziang Xiao, the team set out to investigate whether multilingual LLMs are truly breaking language barriers. The research was presented earlier this year at the 2025 Annual Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics.

Their approach was meticulous. The team analyzed coverage of global conflicts, such as the Israel–Gaza and Russia–Ukraine wars, identifying key types of information—common knowledge, contradictory assertions, unique facts, and perspective-driven narratives. They then created two sets of fictional news articles: one containing truthful information and another with conflicting details. These articles were written in both high-resource languages like English, Chinese, and German, and lower-resource languages such as Hindi and Arabic.

Testing the AI Giants

Next, the researchers tested well-known LLMs from developers like OpenAI, Cohere, Voyage AI, and Anthropic. They posed queries that:

• Asked the models to resolve contradictions between articles in different languages.
• Queried about facts present in only one language.
• Explored biased or perspective-driven reporting.

The results were striking: LLMs overwhelmingly favored the language of the query itself. If asked in English, they sourced answers from English articles even when more accurate or nuanced perspectives existed in other languages.

What Happens When There’s No Article in the User’s Language?

This scenario is common for speakers of low-resource languages. In such cases, the LLMs defaulted to content in high-resource languages, with English dominating. For example, if a user asked a question in Sanskrit about an Indian figure, the model’s response would be shaped almost entirely by English content.

This bias creates what Sharma calls “information cocoons,” trapping users in linguistic filter bubbles that reflect the dominant language rather than a balanced global perspective.

A Real-World Example

Imagine three users asking about the India–China border dispute:

• A Hindi-speaking user receives an answer influenced heavily by Indian narratives.
• A Chinese-speaking user sees answers aligned with Chinese sources.
• An Arabic-speaking user without access to Arabic-language coverage—gets an English-centric perspective.

This scenario underscores how the same event can be portrayed in vastly different ways depending on the language of the query, leading to fragmented and potentially distorted understandings of global issues.

The Larger Implications

The study highlights a growing issue of linguistic imperialism in AI: a world where high-resource languages dominate the global narrative, overshadowing minority perspectives. This dynamic is especially concerning in coverage of sensitive topics such as international conflicts, trade disputes, and politics, where balanced and diverse viewpoints are crucial.

Sharma warns, “The information you’re exposed to determines how you vote and the policy decisions you make. If we want to shift power to the people, AI systems must provide holistic, multilingual perspectives.”

Solutions and Future Directions

To address this issue, the Johns Hopkins team proposes several solutions:

1. Dynamic Benchmarking and Datasets: Build evaluation tools that focus on diversity and fairness across languages.
2. Data Mixture and Training Adjustments: Explore training strategies to ensure balanced representation of languages.
3. Retrieval-Augmented Generation (RAG): Enhance models’ ability to source and cross-reference multilingual data.
4. User Education: Promote information literacy and caution users against over-trusting AI outputs.
5. Language Diversity Warnings: Notify users when responses are disproportionately influenced by a single language.

The Call for Responsible AI Development

This research underscores a vital truth: concentrated power over AI technologies can perpetuate biases, manipulate narratives, and contribute to misinformation. To build trust, AI developers and policymakers must prioritize language equity and transparency.

“Society deserves AI systems that deliver consistent and fair information, regardless of a user’s language or location,” says Sharma. The team hopes their research will inspire more inclusive and globally representative AI development.

While AI holds the promise of democratizing knowledge, the dominance of English and other high-resource languages is creating a digital language divide. The Johns Hopkins study is a wake-up call to make AI truly multilingual, empowering users worldwide with equitable access to information.