18.8 C
New York

AI chatbots ‘think’ in English, research finds


A recent study has shown that artificial intelligence chatbots primarily process information in English, even when presented with questions in other languages.

To explore this research, a team from the Swiss Federal Institute of Technology in Lausanne examined three versions of these AI chatbot models by delving into the different layers that constitute their internal processing.

“We dissected these models layer by layer,” explained researcher Veniamin Veselovsky in an interview with the New Scientist. “Each layer plays a unique role in processing the input, or the initial prompt. We aimed to determine if these internal layers essentially interpret information in English.”

The ‘English subspace’

The researchers selected open-source models and presented them with three types of prompts in four different languages: French, German, Russian, and Chinese. The prompts involved tasks like repeating a given word, translating between non-English words, and filling in a one-word gap in a sentence.

Upon examining the processes undergone by the chatbots to respond to these prompts, the researchers discovered a common thread across all models and layers, which they referred to as the “English subspace.”

Essentially, this means that instead of directly translating between two non-English languages, the chatbots first translate the information into English before proceeding to the target language. Veselvosky highlighted the significance of this finding, indicating that English is instrumental in the bots’ comprehension of certain concepts.

Commenting on these findings, Aliya Bhatia from the Center for Democracy & Technology in Washington DC expressed concerns over the potential implications.

“English and some UN languages offer richer data sources for training AI models compared to other languages. Consequently, developers often train their models predominantly on English data,” Bhatia noted.

“However, utilizing English as an intermediary for teaching language analysis to models runs the risk of imposing a narrow perspective onto linguistically and culturally diverse regions.”

Featured Image: Ideogram

Related articles

Recent articles