The large-language-models (LLMs) behind AI chatbots ‘think’ in English, even when being asked questions in other languages, new research shows.
To investigate this phenomenon, researchers at the Swiss Federal Institute of Technology in Lausanne looked at three versions of these AI chatbot models: opening them up to see the various “layers” that make up these LLMs’ inner processing.
“We opened up these models and looked at each of the layers,” researcher Veniamin Veselovsky told the New Scientist. “Each of these layers does something to the input, the original prompt that you give it. We wanted to see, can we see that the internal layers are actually processing in English?”
The ‘English subspace’
The models, which were chosen on account of their open-source nature, were fed three types of prompts in four languages: French, German, Russian, and Chinese. The first prompt-type asked the LLM to repeat the word it was given. The second requested that the LLM translate from one non-English word to another. And the third and final prompt asked the LLM to fill a one-word gap in a sentence.
The researchers then managed to backtrace all the different changes and processes the LLM had to go through in order to arrive at the answers to these prompts. What they found was that all of these LLMs and all of these layered processes have one thing in common: they all pass through what they coin the “English subspace.”
This basically means that instead of translating straight from French to German, it takes a detour and translates from French, to English, and then to German, or vice versa. According to Veselvosky, this is significant because it suggests that these LLMs are using English in order to understand certain concepts.
Speaking to the New Scientist, Aliya Bhatia of the Center for Democracy & Technology in Washington DC explained why these results may be concerning.
“There’s more high-quality data available in English and some UN languages to train models than in most other languages and as a result, AI developers train their models mostly on English-language data,” she explained.
“But using English as the intermediary through which to teach a model how to analyse language risks superimposing a limited world view onto other linguistically and culturally distinct regions.”
Featured Image: Ideogram
