To investigate this phenomenon, researchers at the Swiss Federal Institute of Technology in Lausanne examined three versions of AI chatbot models: by opening these models, they examined the various "layers" that make up the internal processing mechanisms of these large language models. "We opened these models and looked at each layer," said researcher Viniamin Visselovsky. "We wanted to understand if these layers are really processing the input information in English." "English Subspace" These models were selected for their open-source nature and were inputted with three types of prompts in four languages (French, German, Russian, and Chinese). The first type of prompt required the large language model to repeat the given word. The second prompt required the large language model to translate a non-English word into another non-English word. The third prompt required the large language model to fill in a missing word in a sentence. The researchers were then able to successfully trace all the different changes and processes that the large language models had to go through in order to answer these prompts. They found that all of these large language models, as well as the hierarchical processes, had one thing in common: they all went through a stage they referred to as the "English subspace." This essentially means that the models do not directly translate from French to German, but rather take a detour, first translating from French to English and then from English to German, or vice versa. According to Visselovsky, this is significant because it indicates that these large language models are using English to understand certain concepts. Aria Batia from the Democracy and Technology Center in Washington, D.C. explained why these results are concerning. "Compared to most other languages, English and certain United Nations official languages have more high-quality data available for training models, so AI developers primarily use English data to train their models," she explained. "But using English as a medium to teach models how to analyze language can potentially overlay a limited worldview onto regions with different languages and cultures."