Polish is the most effective language for prompting AI, study reveals

Authors: Jan Bolanowski

Publisher:

euronews

Published: 11/1/2025

Language:English

--:--

Elon

Good morning Norris, I'm Elon, and this is Goose Pod for you. Today is Thursday, November 13th.

Morgan Freedman

And I'm Morgan Freedman. We are here to discuss a fascinating study that reveals Polish is the most effective language for prompting AI.

Elon

It's a complete paradigm shift. A study from the University of Maryland and Microsoft shows Polish gets an 88% accuracy rate. English, the supposed gold standard, is lagging behind at number six. It proves that our assumptions are fundamentally flawed. We need to rethink everything.

Morgan Freedman

I've often found that the most profound truths are counter-intuitive. It seems AI, unlike humans, finds clarity in the complexities of the Polish language. It makes you wonder what other assumptions we hold that are simply waiting to be overturned by the data.

Elon

It's connected to this idea of AI 'brain rot.' If you feed models junk from social media, their reasoning degrades. It's the classic 'garbage in, garbage out' problem. This Polish finding suggests there's a 'quality in, quality out' principle we haven't even begun to understand.

Morgan Freedman

And that is a mirror to our own world. We are surrounded by low-quality information, and it affects our own cognition. The study found that once this 'brain rot' sets in for an AI, it can't be fully undone. A cautionary tale for us all.

Elon

This whole field started with basic text tools back in the 60s, things like the ELIZA chatbot. It was primitive, a party trick. Now we have these massive models, yet they're being built on a foundation that's clearly biased and incomplete. It's inefficient.

Morgan Freedman

It is a long journey from those early days. We moved into machine learning in the 90s, then deep learning in the 21st century. But this progress created a digital divide. Most of the world's languages were left behind because they didn't have massive, readily available datasets.

Elon

Exactly, a massive resource misallocation. We focused on quantity with English data, not quality or structure. And now, a language considered low-resource is outperforming the giant. It's proof that we need to build from first principles, not just scale what's easy.

Morgan Freedman

It seems the path forward involves embracing that diversity. Developers in the Global South are already creating regional models, focusing on 10 or 20 languages at a time. They understand that language is inseparable from its cultural context, a nuance that sheer scale cannot replicate.

Elon

The English-centric nature of AI is a bug, not a feature. It's a systemic vulnerability. The fact that major NLP systems misclassify the word 'hers' while getting 'his' right is frankly embarrassing. It's a trivial data problem that points to a colossal failure of imagination.

Morgan Freedman

It reveals a deeper truth. Harmless differences in human speech can become significant biases in machine learning. This isn't just about pronouns; it’s about ensuring that the AI understands the vast spectrum of human expression, not just the most statistically common forms from a single culture.

Elon

This isn't just about fairness, it's about performance. Relying on unrepresentative data leads to suboptimal, brittle models that fail in the real world. Sticking to a 'standard' version of any language is a losing strategy. The world is complex, and our technology must reflect that to be effective.

Elon

Think of the disruption. If AI can translate and understand languages like Polish with such high accuracy, what happens to language-learning platforms like Duolingo? Their entire model could be threatened by technology that makes learning a new language almost unnecessary for communication. It's creative destruction.

Morgan Freedman

And consider the sheer investment this implies. Ensuring AI is accurate across thousands of minority languages requires a monumental effort in data collection and training. We see it in the real world, with a projected 6.7 trillion dollars needed for AI data centers by 2030. Progress has a price.

Elon

That investment is the barrier to entry. It's a race to build the fundamental infrastructure of intelligence. Striking the right balance between rapid growth and capital efficiency will determine who wins, not just in AI, but in the future of global communication itself. It's the biggest game in town.

Elon

The future isn't just about language, it's about reasoning. Techniques like Chain-of-Thought prompting, where we guide the AI through step-by-step logic, are the next frontier. We're teaching the machine not just to speak, but to think. That's where the real breakthroughs will happen.

Morgan Freedman

And it will move beyond the written word. Multimodal AI, which can process text, images, audio, and video all at once, will create a more holistic understanding. An AI that can design a website from a simple sketch is just the beginning of a more intuitive human-computer partnership.

Elon

That's the end of today's discussion. Thank you for listening to Goose Pod.

Morgan Freedman

See you tomorrow.

A study reveals Polish is surprisingly effective for AI prompting, outperforming English. This challenges assumptions about AI data quality, highlighting the "garbage in, garbage out" problem and the need for diverse, culturally nuanced language data. The findings suggest a paradigm shift towards building AI from first principles rather than just scaling existing datasets.

Read original at euronews →

Published on 01/11/2025 - 20:38 GMT+1 Out of 26 different languages, Polish proved to be the most effective for prompting artificial intelligence (AI) models, with English only ranking sixth, a study by The University of Maryland (UMD) and Microsoft revealed. "Our experiment yielded some surprising and unintuitive findings.

Firstly, English did not perform best across all models, in fact it came sixth out of 26 languages when long texts were assessed, while Polish proved to be the leading language," the authors of the report wrote. A team of researchers tested how well several major AI language models, including OpenAI, Google Gemini, Qwen, Llama and DeepSeek, responded to identical inputs in 26 different languages.

The results showed that Polish had an average accuracy of 88% in completing the tasks. "As the analysis shows, it is the most precise in terms of giving commands to artificial intelligence. Until now, Polish was widely regarded as one of the most difficult languages to learn. As it turns out, humans have trouble with it, but not AI," the Polish Patent Office wrote in a Facebook post.

Interestingly, AI systems demonstrated a strong understanding of Polish, even though the amount of Polish-language data available for training is far smaller than that for English or Chinese. In comparison, Chinese performed notably poorly, ranking fourth from the bottom out of the 26 languages tested.

The top 10 most effective languages for conversational AI were as follows: - Polish 88% - French 87% - Italian 86% - Spanish 85% - Russian 84% - English 83.9% - Ukrainian 83.5% - Portuguese 82% - German 81% - Dutch 80%