A new study has revealed an unexpected linguistic winner in the world of artificial intelligence—Polish. Researchers from Microsoft and the University of Maryland have discovered that Polish outperforms all other tested languages when prompting large language models (LLMs), including those developed by OpenAI, Google, and Meta. The findings have surprised experts who assumed English, the dominant language in AI training data, would lead the rankings.
How the Study Was Conducted
The team evaluated six popular AI models across 26 different languages. They designed a set of “needle-in-a-haystack” tasks that required models to locate specific words hidden in long passages of text. These tests measured the models’ ability to handle long contexts and retrieve precise information accurately. Surprisingly, the models performed best when the prompts were written in Polish, achieving an average accuracy of 88 percent. English followed in sixth place with about 84 percent accuracy, while Chinese lagged behind at around 62 percent.
Why Polish Stands Out
Experts suggest that Polish’s grammatical structure and use of the Latin alphabet could be key factors behind its superior performance. The language’s complex inflection system might help AI models interpret instructions more precisely, reducing ambiguity during processing. Researchers also noted that even though Polish appears less frequently in AI training data, its structure allows for clearer tokenization and more efficient interpretation by the model.
What the Discovery Means
This study challenges the assumption that English is always the most effective language for interacting with AI systems. It shows that language structure, not just data quantity, plays a major role in model performance. For developers and prompt engineers, this finding opens up exciting possibilities for exploring multilingual strategies to improve AI accuracy and reasoning.
