The artificially intelligent large language models (LLMs) behind chatbots may 鈥渢hink鈥� in English, even if asked questions in other languages. This is because their training data is biased, encoding concepts more common in English language cultures.
LLMs have become popular since the November 2022 release of ChatGPT, and many can answer questions put to them in a number of languages, responding in the same language.
To see which language the LLMs were actually using to process queries, , and their colleagues at the Swiss Federal Institute of Technology in Lausanne looked at three versions of the Llama 2 model, made by Facebook-owner Meta.
Advertisement
鈥淲e opened up these models and looked at each of the layers,鈥� says Veselovsky. LLM models are made up of layers of processing. These translate written prompts into tokens 鈥� which are words or sections of words 鈥� then try to contextualise each token to provide an answer.
鈥淓ach of these layers does something to the input, the original prompt that you give it,鈥� says Veselovsky. 鈥淲e wanted to see, can we see that the internal layers are actually processing in English?鈥�
The Llama 2 models were chosen because they are open-source, so available to the public, and therefore it is possible to look at each stage of the processing, unlike with other LLMs, such as those behind ChatGPT. The method used should extrapolate to any open-source model.
Three types of prompt were fed to the models in Chinese, French, German and Russian: one asking it to repeat the word it was given; another asking it to translate from one non-English language to another; and the third asking it to fill in a single-word gap in a sentence, such as 鈥淎 ___ is used to play sports like soccer and basketball鈥�.
The researchers tracked back the processes the LLM went through to answer each prompt. They found that the path of processing through the layers almost always passes via what they call the English subspace. If asked to translate Chinese to Russian, the correct Russian characters travel through the English subspace, before going back to Russian, says Veselovsky, which is a strong indication that English is being used by the models to help them understand concepts.
The results are unsurprising, but could be a cause for concern. 鈥淭here鈥檚 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,鈥� says at the Center for Democracy & Technology in Washington DC. 鈥淏ut 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.鈥�
It is a big issue for artificial intelligence, says at the University of Oxford. 鈥淭he dominance of English reduces diversity,鈥� she says. 鈥淚f English is the main language in which systems process queries, we will likely lose concepts and nuances that can only be appreciated in other languages.鈥�
There are also more fundamental risks with encoding generative AIs used around the world with Anglocentric values, such as incorrect answers that the AI 鈥渉allucinates鈥�, says Bhatia. 鈥淚f a model is used to generate text in a language it has not been trained on, it may result in culturally irrelevant hallucinations, and if a model is used to make asylum decisions for a community that doesn鈥檛 fit within the Anglocentric imagination of society, the model may stand between an individual and access to safety,鈥� she says.
arXiv
