Michigan study: AI models use 1.4 to 180 times more energy on non-English text

Michigan study: AI models use 1.4 to 180 times more energy on non-English text

University of Michigan Engineering researchers report that generative AI burns more energy when it processes and produces non-English text, and sometimes gives less accurate answers. They say this could reinforce existing global inequalities. The study was presented at the Empirical Methods in Natural Language Processing (EMNLP) conference of the Association for Computational Linguistics.

The headline finding is that AI models could use 1.4 to 180 times more energy to generate non-English text. The gap is greatest in languages that do not use Latin script and have less digital content available for training. Most Western European languages come in at twice the energy expenditure of English or less. Hindi and Bengali, each spoken by hundreds of millions of people, need 7.9 and 15.3 times more energy respectively than an equivalent English prompt. Using the researchers' measurements and demographic data from Ethnologue, the source puts the number of people who must use more than twice as much energy to generate text in their first language, compared with English speakers, at around 5.2 billion.

Naihao Deng, a doctoral student in computer science and engineering at the U-M AI Lab and first author of the study, says the languages that consume more energy also tend to have lower accuracy in the output. He argues that this double penalty makes it more likely that certain people will be priced out of AI, and that companies will be more reluctant to serve models in certain languages. People may also feel pressured to use AI as it becomes part of more areas of society.

How it was measured: the team tested three open-source models, Alibaba's Qwen 3, Google's Gemma 3 and Meta's Llama 3.1. Each model answered multiple-choice questions from Belebele, a reading-comprehension dataset covering 122 languages. Energy per question was measured with the ML.Energy framework, developed at Michigan Engineering. To check other kinds of tasks, the team also used grade-school math word problems and open-ended user prompts from other benchmarks. Those datasets are not multilingual, so the researchers translated the English questions into seven other languages with Meta's No Language Left Behind translation tool. On average, all the models used less energy for English. The disparity persists at larger model sizes and across different types of processors.

The cause is tokens. In other languages the models had to generate more tokens, the basic unit of data AI uses, which can stand for whole words and phrases or single syllables and characters. The problem grows for scripts whose characters need more code to display. The Devanagari script used for Hindi, Nepali and Marathi, and the Bengali script, have primary characters that change form and combine with accents to produce up to millions of unique characters. More tokens take up more memory, which leaves the model less room for processing, so it works through tokens in smaller batches, and that uses more energy per token. More tokens can also raise direct costs, since companies often price model access by the number of tokens generated.

The researchers recommend that companies optimize their services for more languages, for example by routing user queries to models tailored to specific languages. Companies could also translate user prompts to English before passing them to a more generalized model, provided the translation is reliable. Rada Mihalcea, the Janice M. Jenkins Collegiate Professor of Computer Science and Engineering, director of the U-M AI Lab and corresponding author, says AI developers should create more resources for languages with an energy and cost disparity. She adds that crowdsourcing native speakers to help translate material or create resources to gather enough training data could also improve efficiency.

The National Science Foundation and DARPA partially funded the work. Additional funding came from Ford, Laude Institute, OpenAI and the Survival and Flourishing Fund. The study used the University of Michigan's Great Lakes high-performance computing cluster.

Key facts

  • Generating non-English text took 1.4 to 180 times more energy than English across the three open-source models tested (Qwen 3, Gemma 3, Llama 3.1); most Western European languages came in at twice English or less.
  • Hindi needed 7.9 times and Bengali 15.3 times the energy of an equivalent English prompt; the gap is largest for non-Latin scripts with less training content.
  • Higher-energy languages also tended to have lower accuracy, which the authors call a double penalty; around 5.2 billion people must use more than twice the energy to generate text in their first language.
  • The cause is more tokens per output, which fills memory and forces smaller batches, raising energy per token; token-based pricing can raise direct costs too.
  • The researchers recommend language-specific models, or translating prompts to English first if the translation is reliable, plus crowdsourced resources from native speakers.

Why it matters

Most people want to ask AI questions in their native language, and the study finds that doing so can cost far more energy than doing it in English. Because companies often price access by tokens generated, extra tokens can mean higher direct costs for the user. Deng's worry is that the energy penalty and the accuracy penalty arrive together, so some people could be priced out of AI and companies could be less willing to serve certain languages. The source frames this as a way existing global inequalities could be reinforced.

Who it affects

Speakers of languages outside English are affected most, especially languages that do not use Latin script and have less digital content for training. Hindi and Bengali speakers face 7.9 and 15.3 times the energy of English. The source estimates around 5.2 billion people must use more than twice as much energy to generate text in their first language. AI developers and companies that serve models, and that price access by tokens, are the ones the recommendations are aimed at.

How to use it

The source gives its advice to companies and developers rather than end users. Optimize services for more languages, possibly by routing user queries to models tailored to specific languages. Or translate user prompts to English before feeding them to a more generalized model, provided the translation is reliable. Mihalcea adds that crowdsourcing native speakers to help translate material or create resources could supply enough training data to improve efficiency. The measurements were made with the ML.Energy framework, developed at Michigan Engineering.

How solid is it

The work comes from University of Michigan Engineering and was presented at the EMNLP conference of the Association for Computational Linguistics. The source for this account is the university's own write-up. Energy was measured directly with the ML.Energy framework on three open-source models, using the 122-language Belebele dataset, and the gap held at larger model sizes and across different processor types. The source gives only ratios relative to English, not absolute energy figures. It does not say which models or languages produced the 180 times figure. The subtitle's 'up to 15 times' refers to widely spoken languages, while 180 times is the top of the overall range.

Risks and caveats

Only three open-source models were measured, so the results should not be read as covering every model. The math and open-ended prompt datasets were made by machine-translating English questions into seven other languages with Meta's No Language Left Behind tool. The source does not list those languages. The translate-to-English fix works only if the translation is reliable, per the researchers. The 5.2 billion figure combines the researchers' measurements with Ethnologue demographic data. Funding included OpenAI, Ford, Laude Institute and the Survival and Flourishing Fund, alongside the NSF and DARPA.

“The languages that typically consume more energy also tend to have lower accuracy in the model output”

— Naihao Deng, doctoral student in computer science and engineering at the U-M AI Lab and first author of the study