Chatbots “think” in English even when working with other languages

Scientists from the Federal Polytechnic School of Lausanne studied several versions of the large language model Llama 2 from Meta (recognized as extremist and banned in the Russian Federation) and came to the conclusion that the main AI processes take place in English, even when queries to them were formulated in other languages. Llama 2 is open source, which allowed researchers to study AI at every stage of request processing.

According to scientists, the work of AI models is as follows: first, written clues are translated into tokens, then each of them is contextualized and then an answer is provided. The researchers tried asking the AI ​​in Russian, French and German: each time the request was processed through the English subspace.

According to scientists, AI operating only through English-language subsystems could lead to the spread of English-related worldview restrictions to other culturally and linguistically distinct regions. According to the researchers, there are fundamental risks associated with the spread of generative AI with Anglocentric values.

Why is interest in ChatGPT decreasing in the socio-political sphere? Find out in the material “Kommersant” “Artificial intelligence did not come out”.

Evgeniy Fedunenko