Linguistic imbalance is the unequal representation and computational support of languages in training data and artificial intelligence systems. Large language models reflect these disparities because their abilities depend heavily on the languages present in their training data. section 3.3
English and a small group of high-resource languages dominate online data, while thousands of languages remain underrepresented or absent. Differences in grammar, morphology, and cultural nuance make simple word counts an incomplete measure of language support. section 3.3
When article counts mislead
Wikipedia illustrates the problem: articles about countries in Africa and Asia may favor English or colonial languages over local knowledge. Cebuano appears to have millions of articles, yet up to 99% may have been generated by an automated bot, limiting human-authored nuance and cultural context. section 3.3
