In which language do you think chatgpt? When we communicate with artificial intelligence such as chatgpt, Gemini or Claude, we do it in our language: that it is Italian, French, English, Chinese … and the Ai responds to us in the same language, in a coherent way. But how is it possible that one of the can understand and elaborate Answers in all languages of the world? In what language “do you think” before responding? These questions, which may seem only curiosity, are at the center of numerous research in recent years. Understanding how AI elaborates language helps us to understand its potential, limits and, above all, to control its risks.
To try to answer these questions, in March 2025, Anthropic – the company that develops Claude – He has published a research on how his model interprets language. The study showed that Claude 3.5 Haiku it works similarly to the brain of people bilingual: activate the same internal “areas” when it elaborates words in different languages. In practice, if a question in multiple languages arises, the same “circuits” are activated regardless of the language used. This process allows Claude to transfer concepts from one language to another, allows those who use it to write in your own language Mother without losing in the quality of the response and directs us towards the creation of to the increasingly transparent and safe.
In this article we see how language models work, what it means to say that aTo the “think” in multiple languagesand why understanding these mechanisms is essential to build more reliable, transparent and safe systems.
What are the Large Language Models “think”
Large linguistic models, or LLM (Large Language Models), are systems designed for interpret, trial And generate text. They are the basis of tools like Chatgpt, Gemini And Claudeand are capable of conversing, summarizing documents, answering complex questions and also translating between languages. To do all these operations, there have been no row line by human beings, but they were trained on huge quantities of textsfrom which they learn for themselves to recognize, relations And rules of language.
Even if we know what the main mechanisms underlying the choices of words and interpretation of requests by LLM, the logical strategies that develop during the training phase to face linguistic tasks are still unprecedenablealso by the developers themselves.
As the CEO of Anthropic himself said:
When a generative artificial intelligence does something, how to summarize a financial document, we have no idea (…) of why it makes the choices it makes, because choosing certain words rather than others, or because it occasionally makes a mistake despite being usually accurate.
Precisely for this reason, being able to understand the profound logic at the basis of linguistic choices or interpretation of the texts is a central point for having more and safer to the most sure.
Claude “Do you think in multiple languages”? Anthropic’s search
To try to interpret the operation of Claude and to understand if “thinks” differently depending on the language in which the question is asked, Anthropic has decided to observe which areas Yes “activate“When a question is asked in different languages and how these areas are connected among them, a bit like when you do a magnetic resonance imaging on a human brain.
They started from simple requests: they asked, for example, to Claude Di complete the sentence “The opposite of ‘little’ is …” in English, French And Chinese. Analyzing the internal activations of the model, they understood that:
- “Small”, “small” and “petit” always activate the same area, that is, regardless of the language used, it same concept (the “small” being) correspond to the same area model;
- When activating the area that contains the concept of “small” and that of “opposite”, the area that contains the concept of “large” is also activated, and this applies to all languages. For Claude, the mechanism with which the opposite of a word is generated does not depend on the language, but refers to a common, shared, abstract representation. Claude, therefore, is capable of generalize the relations among the concepts and to make them independent of the starting language.
This study, of course, also has limitations. Although only very short and easy to interpret requests have been studied, the research group was unable to explain all the operations that Claude was making to generate the answer. This is a sign of how much we are still far away to understand really in depth the functioning of these models.
We have not yet fully understood how to “think” the Ai
Understand in which language “think” A language model has very practical consequences, especially to design systems that work welland in several languages, that they are efficient And above all, safe. If a AI really manages to abstract the concepts and to connect them to each other, then it can transfer what has learned in a language to the others. This allows you to guarantee consistency And quality in the answersregardless of the user’s language.
The research in this area, however, is still at the beginning, and the opinions of the scientific community are not unanimous. Some studies argue that the most advanced language models use structures Truly multilingual, Like Claude. Others, however, observe that, even in the most sophisticated systems, English continues to play a dominant role, especially in the final passages of the generation of responses. The main reason is linked to training: most of the data on which these models are educated are, still today, in English language.
Despite all the limits, studies such as Anthropic represent a first step towards greater transparency of the internal mechanisms of linguistic models. The way to truly understand the internal mechanisms of AI is still long, but studies like this represent fundamental stages for a greater mastery of this technology.
