AI software works better in some languages than in others, according to a new study.
Researchers have said users get better results when they’re using certain languages to communicate with large language models (LLMs), which are the systems behind AI chatbots.
The researchers tested software made by six different AI companies. Native speakers of 26 different languages then gave the AI the same tasks to do.
They were quite complex tasks. For example, the LLMs were given a large amount of data and asked to find one piece of information that didn’t belong.
The researchers looked at how effective the different languages were at doing the tasks.
For the tasks that gave LLMs the largest amounts of information to work with — tens of thousands of words or more — they found that Polish was best, giving results that were accurate 88% of the time.
Next was French at 87%, followed by Italian at 86%, Spanish at 85% and Russian at 84%.
English was sixth, just below 84%.
The three least accurate languages were Swahili, Tamil and Sesotho, a language spoken in Lesotho and South Africa. Sesotho was accurate less than 50% of the time.
The researchers said this was probably because there’s not as much online content written in these languages for LLMs to be trained on.
However, Polish was more accurate than English or Chinese, even though there’s far less online content in Polish — and Chinese was actually fourth from the bottom.
The hope is that in the future, people will get the same good results in every language.
