The use of artificial intelligence is exploding around the world, but the technology’s language models are primarily trained in English, leaving many speakers of other languages behind.

On July 26, 2024, researchers at Stanford conducted a language test with a popular A.I. chatbot. They prompted the bot in Vietnamese to create a traditional poem known as “song thất lục bát,” which requires a specific structure of seven, seven, six, and eight words. However, the bot failed to adhere to the required format.

When asked for the proper Vietnamese term for a mother’s younger brother, the bot mistakenly provided terms for a father’s younger and older siblings. These errors highlight a significant issue: A.I. tools often struggle with languages outside of standard American English.

Why does this matter?

While A.I. has gained immense popularity in Western countries, many regions of the world remain excluded from this technological advancement. Experts express concern that the language gap could worsen existing technological inequalities, leaving diverse cultures behind.

Sang Truong, a Ph.D. candidate at Stanford, emphasized that even a few years of delayed access to effective technology could result in decades of economic setbacks. His team’s findings revealed that A.I. tools frequently misinterpret facts and vocabulary when applied to Vietnamese, primarily due to the language being classified as “low-resource.” This classification indicates a lack of sufficient data sets and online content for A.I. models to learn from.

What can be done?

Addressing the language gap in A.I. requires a concerted effort to develop models that are inclusive of various languages. Initiatives like Lelapa AI in Johannesburg aim to enhance A.I. technology for African languages, showcasing the potential for socially-grounded research to bridge this divide.

In conclusion, as A.I. continues to evolve, it is crucial to ensure that advancements benefit all languages and cultures, fostering inclusivity in the digital age.