Polish is the best-performing language when it comes to AI

In an absorbing new study, researchers were surprised to find that the English language wasn't as effective in artificial intelligence (AI) accuracy as originally thought.

Who’d have thought that when it came to prompt-effectiveness for AI, Polish would be the most effective language?

A study from the University of Maryland and Microsoft has found that prompts made in Polish yielded an 88% accuracy rate compared to that of 26 other languages. English, rather disappointingly, was found languishing in sixth on the list (out of 26).

Despite Polish being a difficult language (especially with pronunciation) to learn, doesn’t mean that machines will encounter this when scraping the web.

The structural elements of the Polish language could well mean that it is best suited for giving commands, and its rich and complicated grammatical structure can lend itself well to reducing ambiguity and heightening nuanced expression.

A retro schoolteacher with algebra on the blackboard. — Getty Images

The elements of language

Six major large language models(LLMs) were used, including OpenAI’s o3-mini , Google’s Gemini, Alibaba’s Qwen, Meta’s Llama, and DeepSeek. The bots were tested in 26 languages, with English, Chinese, Persian, and Italian among them.

With English being the most prominent language of AI, there may be numerous inaccuracies in the extensive set of training materials available.

As the language often includes a lot of casual speech, slang, poor grammar, and sarcasm, it may be the case that there is more distortion and ambiguity.

This looseness seems to explain why ChatGPT, for example, might commonly hallucinate or rattle off a string of words that don’t always add up.

A young man actually in the bottom of a barrel. — Getty Images

Scraping the bottom of the barrel

When I was using it the other day, it said “to be fare” instead of “to be fair,” which made me feel startled, and when I mentioned it, it said “ah yeah, I just made a typo” as if it didn’t matter at all.

Another language that severely underperformed was Chinese, which also has large swathes of content available to scrape.

After the Polish language, the Romance languages also performed well. French’s success rate was 87%, followed by Italian at 86%, and Spanish at 85%. These languages have a particular level of fluidity, in terms of looser grammar and high expressiveness.

The study suggests that future LLMs may be trained in more diverse languages, as companies look to maximize their success rates.