Generative AI uses more energy on non-English text

Key takeaways

• Gen AI uses 1.4 to 180 times more energy to produce text outside the English language.
• Higher energy costs increase the barrier to using AI for billions of people worldwide—they must either pay more or switch to a non-native language.
• AI developers could improve energy efficiency by training models optimized for specific languages or incorporating translation pipelines into their services.

---

When processing and producing non-English text, generative AI burns more energy and sometimes provides less accurate answers, potentially reinforcing existing global inequalities, according to University of Michigan Engineering researchers.

The energy disparity is greatest in languages that don’t use Latin script and have less digital content available for training. AI models could use 1.4 to 180 times more energy to generate non-English text, with most Western European languages coming in at twice the energy expenditure or less. However, Hindi and Bengali, each spoken by hundreds of millions of people, require 7.9 and 15.3 times more energy than an equivalent English prompt.

“The languages that typically consume more energy also tend to have lower accuracy in the model output,” said Naihao Deng, a doctoral student in computer science and engineering from the U-M AI Lab and the first author of the study, which was presented at the Empirical Methods in Natural Language Processing conference of the Association for Computational Linguistics.

“Most people will want to use AI models by asking questions in their native languages, and they may feel pressured to use it as it’s incorporated into more parts of society, but the 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.”

At present, around 5.2 billion people have to use more than twice as much energy to generate text in their first language compared to English speakers, according to the researchers’ measurements and demographic data available in Ethnologue. But the researchers also provide recommendations to bridge the gap and ensure that users aren’t penalized for trying to use AI in their language of choice.

To help quantify AI energy use and improve efficiency, the National Science Foundation and the Defense Advanced Research Projects Agency (DARPA) partially funded the research.

Measuring AI’s language energy disparity

The researchers found the energy and accuracy disparity by measuring the amount of energy used by three open-source AI models, Alibaba’s Qwen 3, Google’s Gemma 3, and Meta’s Llama 3.1. They asked each model to answer multiple-choice questions included in Belebele, a dataset for assessing the reading comprehension of AI models in 122 different languages. They then measured the amount of energy used to answer each question using the ML.Energy framework, developed at Michigan Engineering.

To check for the disparities in other types of tasks, the researchers also asked the models to answer a collection of grade-school math word problems and open-ended user prompts included in other benchmarking datasets. Because these datasets aren’t multilingual, the researchers translated the English questions into seven other languages using Meta’s translation tool, No Language Left Behind.

All the models used less energy for English text, on average. The energy disparity persists at larger model sizes and across different types of computer processors.

Extra tokens, extra cost

When working in other languages, the models used more energy because they had to generate more tokens, the basic unit of data used by AI. Tokens can represent entire phrases and words or single syllables and characters. AI models convert prompts into tokens before processing them, and then convert the output tokens into human language for the final response.

This problem escalates for scripts whose characters require more code to display. For example, the Devanagari script used in 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.

A higher number of tokens takes up more computer memory and burns more energy. With more memory consumed in tokens, the model has less available for processing information. So, it processes the tokens in smaller batches, which uses more energy per token. This effect compounds the energy disparity compared to English text.

More tokens can also result in higher direct costs, since companies often price model access by the number of tokens generated.

Reducing AI’s language energy disparity

To help fix the cost disparity, the researchers recommend that companies optimize their services for more languages. This could entail routing user queries to models tailored to specific languages. Companies could also translate user prompts to English before feeding them to a more generalized model, provided the translation is reliable.

“AI developers should be creating more resources for those languages with an energy and cost disparity,” said Rada Mihalcea, the Janice M. Jenkins Collegiate Professor of Computer Science and Engineering, director of the U-M AI Lab and corresponding author of the study. “Crowdsourcing native speakers to help translate material or create resources to get enough training data could also help improve efficiency.”

Additional funding sources include Ford, Laude Institute, OpenAI and the Survival and Flourishing Fund.

The study used the University of Michigan’s Great Lakes high-performance computing cluster.

Related Stories