What language does my AI really speak?

When you type a question into ChatGPT, Gemini or Claude, the response feels instantaneous. If you ask in English, it replies in English. If you switch to Spanish, it effortlessly pivots along with you. When my Iranian parents-in-law visited, I created an English to Farsi GPT and the language barrier almost disappeared.

Photo by Google DeepMind on Unsplash‍

It’s easy to assume the model is “thinking” in English and translating on the fly.

But the reality is stranger.

Large language models don’t truly speak any human language at all. They operate in mathematics — probabilities, vectors and patterns. Language is just the interface layer. And understanding that is the key to getting better results.

The universe in tokens.

At its core, a language model doesn’t see words; it sees numbers.

Text is broken into tokens — fragments that might be whole words, parts of words or even single characters. Each token becomes an ID, which the model represents as a vector: points in a vast, high-dimensional space.

During training, the model learns that “cat”, “gato” and “katze” tend to occupy nearby regions of that space. Not because it knows they’re translations, but because they appear in similar contexts. Meaning emerges from proximity.

In other words, the model builds a language-agnostic representation of concepts.

Attention is all you need.

How does it connect these concepts so efficiently? The breakthrough moment happened in 2017 with a Google research paper titled Attention Is All You Need. This paper introduced the “Transformer” architecture, the “T” in ChatGPT.

Before Transformers, AI struggled with long sentences because it processed words sequentially. By the time it got to the end of a paragraph, it often forgot the beginning.

The Transformer changed everything with a mechanism called “self-attention”. This allows the model to consider every word in relation to every other word at once, regardless of their distance in the sentence.

Crucially, this mechanism works across languages. The attention mechanism can identify that the subject of a sentence in a Hindi paragraph relates to a verb that appears much later, just as easily as it can in English. It maps the relationships between concepts, not just the definitions of words.

The English elephant in the room.

While the architecture is language-agnostic, the training data is not balanced. If you ask a language model a question, it is statistically more likely to draw on English data to form its outputs.

Why is English so dominant? Firstly, volume. The sheer amount of digitised text on the open internet is heavily skewed toward English. Secondly, history. The internet’s roots are in the US military and American academia. Finally, code itself. The foundational syntax of virtually all major programming languages — Python, Java, C++ — is based on English vocabulary (if, else, while, print), which reinforces the dominance of English technical content.

Because the model has “read” more English than anything else, its representations are simply better calibrated in English.

Why context and culture matter.

If the AI knows data in all languages, why does it matter which language you use to prompt it?

When you ask a question in English, you are doing two things: you are asking for English words back, and you are signalling a cultural context.

Language is deeply tied to idiom, culture and nuance. If you ask about “football” in American English, the model knows to weigh data about quarterbacks and touchdowns. If you ask about “fútbol” in Spanish, it knows to prioritise data about strikers and penalty kicks. (And if you ask in Australia where “football” could be half a dozen different things, it will probably ask you what exactly you are referring to so it gives you an appropriate answer…)

The model weights the language of your prompt heavily because it’s the best clue it has regarding the intent of your question. It is trying to mirror your cultural framework to provide the most relevant answer. It assumes that if you are speaking English, you want English sources, English cultural norms and English idioms.

Hacking the language barrier.

So, how do you use this information to get better results, especially for niche topics?

If you are researching a topic that is deeply tied to a specific non-English culture (say, Japanese urban planning or Italian culinary history) don’t just ask in English and hope for the best.

You need to explicitly steer the model toward non-English sources. Instead of asking, “What are the benefits of the Mediterranean diet?”, try this:

Acting as a researcher, synthesise findings on the Mediterranean diet. Please prioritise primary sources and cultural analyses written originally in Italian and Greek, summarising their core findings in English for me.

By explicitly directing the model’s attention toward specific language datasets, you can force it to bypass its default English bias and tap into its deeper, multilingual knowledge base. Your AI is a polymath, sometimes you just need to tell it which shelf to take the books from.

Ready to turn your business challenges into your biggest opportunities? Let’s talk.