Is Polish really the best language for AI?

Polish as AI’s best language is a myth, but a useful one. It asks whether the language you use changes what AI gives you, and the answer is yes. Each language brings its own concepts and culture, while the models learned those of American English.

The claim that Polish is the most efficient language for AI traces back to a single benchmark. The test measured how well models find information in long texts.¹ Polish scored 88%. English came sixth with 83.9%.¹ The Polish test drew on one volume of a single novel.² One of the authors later said the team never researched prompting.²

The difference between Polish and English was not statistically significant.²

So the benchmark says nothing about how well AI works for someone writing in Polish. A request written in Polish carries Polish concepts and culture, and the test never measured them.

A language is a whole system

Grammar and vocabulary are one part of a language. Beyond them, every language holds concepts others could not match.³ Its key words also carry the values of the culture behind it.⁴

Research rejects the idea that a language fixes what its speakers can think.⁵

The same research supports a weaker effect, in which a language directs what its speakers pay attention to.⁵

Each language also carries the history of what its speakers used it for. Before Christian Wolff, very little philosophy was written in German, and his work helped make German a language of philosophy.⁶ In French, Diderot’s Encyclopédie gave progressive thought a wide influence before the Revolution.⁷ Under the partitions and later under communist censorship, Polish writers learned to carry political meaning in allusion.⁸ English grew with trade across an empire claimed to cover a quarter of the Earth’s surface.⁹

The philosophy built in German later travelled into other languages. Polish philosophical terms were often modelled on German ones, as przeżycie translates Erlebnis.¹⁰

Every conversation also rests on what goes unsaid. People leave out what the listener already knows.¹¹

Say “go and park the car” at home, and the listener knows which car and where to park it.

Writing in Polish or in English

AI models treat American English as the default, and the English they learn from leans American in spelling and vocabulary.¹²

But most people who use English are not native speakers.¹³ A gap between the model’s language and the user’s is therefore the common case. A Polish user meets it whether writing in Polish or in English.

Writing in Polish leaves part of that gap open. Even in the user’s own language, models can swap some local references for globally dominant ones.¹⁴

Writing in English crosses the gap twice. The user first moves their own thinking into a foreign language, and people judge risks and moral dilemmas slightly differently in one.¹⁵ The model then swaps local references for global ones far more often than in the user’s own language.¹⁴

Logic and code lose little when the language changes.

• Reasoning in the user’s language instead of English costs only a few per cent of accuracy.¹⁶
• The language of a coding instruction moves the results by a few points.¹

Meaning rooted in a culture loses more.

• On tasks grounded in a culture, prompts written in the native language score about 8 points higher than English translations.¹⁸
• When code has to follow local rules, even the best AI in the test gets more than a third of the tasks wrong.¹⁷
• What goes unsaid is lost as well, because the model sees only the written request. Unclear requirements cut the share of coding tasks models solve by about 7 points on average.¹⁹ When a model flags a requirement as unclear, it is right only about half the time.¹⁹

Greek, Latin, French and now English

Could then AI make the world’s cultures more alike? Common languages have done something similar before, by spreading their concepts into the languages around them. Greek spread from Greece to the Middle East, wherever Greeks or Hellenised kings ruled.²⁰ Its philosophy then passed into Latin, as Cicero coined Latin terms for Greek ideas, and those terms entered languages across Europe.²¹ French later carried the prestige and universalism of the Republic.²² English is today’s common language, and scholars still disagree whether it works as a neutral tool for communication²³ or as a project of power.²⁴

In AI, the swaps of local references for global ones point the same way.¹⁴ Writing in your own language would slow that drift without stopping it.

Working in a language other than American English

Polish is not the best language for AI, but the language you use with AI still matters in practice.

• Prompts for logic, general reasoning and code could be written in whichever language is easiest, since the language changes little in their results.
• Anything tied to a culture, such as values, key words and local references, could come out better written in your own language.
• Local references in an answer would still be worth checking, even in your own language.
• Even the obvious would need to be written down, because AI knows only what you write.
• Anyone working with AI in a language other than American English is most likely to lose exactly what matters most in their own culture, and least likely to notice.