Gen AI burns more energy on non-English text
Original: Generative AI uses more energy on non-English text
Why This Matters
Energy inequality in AI hits non-English speakers hardest, compounding existing access gaps.
University of Michigan research finds generative AI models consume more energy processing non-English text than English, raising equity and environmental concerns for global AI deployment.
Researchers at the University of Michigan have found that generative AI systems use more energy when processing text in non-English languages compared to English. The finding points to a structural inefficiency baked into most large language models, which are predominantly trained on English-heavy datasets. Because tokenizers used by these models are optimized for English, non-English text typically requires more tokens to represent the same semantic content — meaning the model does more computational work per equivalent passage. That extra computation translates directly into higher energy consumption. The implications are twofold: users in non-English-speaking regions face a hidden environmental cost, and the energy disparity adds another dimension to ongoing debates about AI equity and accessibility. The research adds quantitative weight to what linguists and AI fairness researchers have long flagged as a tokenization bias against non-Latin and lower-resource languages.