University of Michigan Engineering researchers found that generative AI can use 1.4 to 180 times more energy to process non-English text than English, while often producing lower-quality answers.

The team tested Alibaba’s Qwen 3, Google’s Gemma 3 and Meta’s Llama 3.1 on reading-comprehension questions, translated math problems and open-ended prompts, measuring energy with the ML.Energy framework.

The gap varied sharply by language: Hindi prompts used 7.9 times and Bengali prompts 15.3 times the energy of equivalent English prompts, while most Western European languages needed no more than twice as much.

English remained more energy-efficient across different tasks, model sizes and computer processors.

A major reason is that non-English text often requires more tokens—the units models process and generate—which use more memory, can require smaller processing batches and may raise costs for services that charge by token.

Researchers estimate that about 5.2 billion people use a first language for which text generation requires more than twice the energy needed for English, potentially increasing costs and discouraging companies from supporting those languages.

To close the gap, the researchers call for more language-specific training resources, including crowdsourcing native speakers to translate material and build data for underrepresented languages.

They also recommend developing models tailored to particular languages or reliably translating prompts into English before processing them with a general-purpose model.