Not only are Polish programmers among the best in the world, it turns out the Polish language itself can outperform artificial intelligence. A new international study has revealed a surprising result: AI chatbots such as ChatGPT perform better when prompted in Polish than in English or Chinese, especially on tasks requiring a long, complex context.

What the study found

Researchers from Microsoft, the University of Maryland, and the University of Massachusetts Amherst devised a benchmark called ONERULER to test large language models across 26 languages for tasks with long-context chains.

Among the tested models and languages, Polish achieved the highest average accuracy (≈ 88 %) on the most complex tasks. In contrast, English ranked sixth, with about 83.9 % accuracy, while Chinese was among the weaker performers, with roughly 62.1 % accuracy.

The top 5 languages in the study were, in order:

1. Polish
2. Russian
3. French
4. Italian
5. Spanish

Why is Polish doing so well?

The researchers acknowledged that the result was “surprising,” especially because English and Chinese dominate AI training datasets.

They suggest several possible contributing factors:

• Script type: Languages using Latin or Cyrillic scripts tended to fare better.
• Data “resource” level: How much training data is available for a language—Polish may have more usable, curated corpora than assumed.
• Linguistic structure: The morphology and syntax of Slavic languages (like Polish) might give them an advantage in encoding meaning in longer contexts.

Implications and reactions

This finding has several interesting implications. It challenges the assumption that English is always the default “best” language for AI prompt engineering or large-language-model interaction, underscores the importance of investing in language-specific AI tools and resources, rather than relying solely on English-centric models, and also could encourage countries and institutions to ramp up development of national AI models tailored to their languages.

Overall, the study injects a fresh perspective into AI and multilingual modeling: even languages that are less dominant globally may have unexpected strengths when it comes to deep, context-heavy tasks.

Main Photo Source: www.freepik.pl