A 2025 paper from the Stanford Institute for Human-Centered Artificial Intelligence found that many popular large language models do not work well in languages other than English. The online world is dominated by English, and many developers rely on English-language data. This leaves many non-English speakers with weaker AI support.
Practical problems appear when AI helps with daily tasks. For example, asking ChatGPT to write an email in Tamil can give a muddled draft in English. Efforts to add more multilingual data sometimes make matters worse because scraped web texts can include machine-translation mistakes. Experts say companies should work with local communities, check multilingual data, and partner with grassroots developers to make AI more useful for everyone.
Difficult words
- dominate — Be stronger or control a place or areadominated
- rely — Need something or depend on it
- multilingual — Using or including many different languages
- scrape — Collect text or data from websites automaticallyscraped
- muddle — Be unclear or mixed up in meaningmuddled
- grassroots — Local people or small community groups
Tip: hover, focus or tap highlighted words in the article to see quick definitions while you read or listen.
Discussion questions
- Have you used AI tools in your own language? What problem did you see?
- How could local communities help improve AI for their language?
Related articles
Inequality and Pandemics: Why Science Alone Is Not Enough
Matthew M. Kavanagh says science can detect viruses and make vaccines fast, but rising inequality makes pandemics worse. He proposes debt relief, shared technology, regional manufacturing and stronger social support to stop future crises.
