AI Models Burn More Energy Processing Non-English Text

Generative artificial intelligence systems consume significantly more energy when processing and generating text in languages other than English, a disparity that can reach up to 180 times the energy expenditure required for English prompts, according to a new study from the University of Michigan. The research, presented at the Empirical Methods in Natural Language Processing conference of the Association for Computational Linguistics, reveals that this energy inefficiency is often accompanied by lower-quality output, creating a double penalty for the billions of people who rely on AI tools in their native tongues. The findings suggest that current AI infrastructure disproportionately favors English speakers, potentially reinforcing existing global inequalities in access to and utility of advanced computational technologies.

The study, titled “The Language-Energy Divide: Measuring Energy Costs of Multilingual LLM Inference,” was conducted by researchers at the University of Michigan Engineering. Naihao Deng, a doctoral student in computer science and engineering from the U-M AI Lab and the first author of the study, noted that languages typically associated with higher energy consumption also tend to yield lower accuracy in model outputs. This correlation implies that users of non-English languages not only face higher computational costs but also receive less reliable information, a combination that could price certain populations out of the AI ecosystem and discourage companies from optimizing their models for diverse linguistic markets.

The researchers quantified this disparity by measuring the energy usage of three prominent open-source AI models: Alibaba’s Qwen 3, Google’s Gemma 3, and Meta’s Llama 3.1. To assess the models’ performance across different languages, the team utilized the Belebele dataset, a resource designed to evaluate the reading comprehension of AI models in 122 different languages.

The measurements revealed that AI models used between 1.4 and 180 times more energy to generate non-English text compared to English. While most Western European languages required only about twice the energy expenditure or less, languages with non-Latin scripts and limited digital training data showed much higher disparities. For instance, Hindi and Bengali, each spoken by hundreds of millions of people, required 7.9 and 15.3 times more energy, respectively, than an equivalent English prompt. This significant gap highlights how the structural characteristics of a language and the availability of digital content for training directly influence the computational efficiency of AI systems.

The primary driver of this increased energy consumption is the higher number of tokens required to represent non-English text. Tokens are the basic units of data used by AI models, which can represent entire phrases, words, or even single syllables and characters. When models process languages that use scripts with complex character combinations, such as the Devanagari script used in Hindi, Nepali, and Marathi, or the Bengali script, the number of tokens generated increases substantially. These scripts feature primary characters that change form and combine with accents to produce millions of unique characters, necessitating a larger token count for the same semantic content compared to English.

This increase in token count has a cascading effect on energy usage. A higher number of tokens consumes more computer memory, leaving less available for processing information. Consequently, the model processes the tokens in smaller batches, which increases the energy cost per token. Additionally, because many companies price model access based on the number of tokens generated, this technical inefficiency translates directly into higher financial costs for users and service providers alike.

The researchers also investigated whether this disparity persisted across different model sizes and hardware configurations. They found that the energy disparity remained consistent even at larger model sizes and across different types of computer processors. To further validate their findings, the team tested the models on other types of tasks, including grade-school math word problems and open-ended user prompts. Since these datasets were not originally multilingual, the researchers translated the English questions into seven other languages using Meta’s translation tool, No Language Left Behind. The results confirmed that the energy inefficiency was not limited to reading comprehension tasks but was a broader characteristic of how these models process non-English inputs.

According to the researchers’ measurements and demographic data available in Ethnologue, approximately 5.2 billion people currently have to use more than twice as much energy to generate text in their first language compared to English speakers. This widespread impact underscores the global scale of the issue, affecting a majority of the world’s population. The study highlights that the current state of AI development is not neutral with respect to language, but rather biased toward English in terms of both efficiency and cost. This bias could hinder the adoption of AI technologies in regions where English is not the primary language, limiting the potential benefits of these tools for a significant portion of the global community.

To address this challenge, the researchers propose several strategies for improving energy efficiency and reducing the language-energy divide. One recommendation is for AI developers to train models that are optimized for specific languages, rather than relying solely on generalized multilingual models. Another approach involves incorporating translation pipelines into AI services, where user prompts are translated to English before being fed to a more generalized model, provided that 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 of the study, emphasized the need for more resources dedicated to languages with significant energy and cost disparities. She suggested that crowdsourcing native speakers to help translate material or create resources could help improve the efficiency of these models by providing better training data.

The research was partially funded by the National Science Foundation and the Defense Advanced Research Projects Agency (DARPA), with additional support from Ford, Laude Institute, OpenAI, and the Survival and Flourishing Fund. The study utilized the University of Michigan’s Great Lakes high-performance computing cluster to conduct the energy measurements. By developing and applying the ML.Energy framework, created at Michigan Engineering, the team was able to accurately measure the energy used to answer each question in the benchmark datasets.

Subject of Research: Energy consumption disparities in multilingual generative AI models

Article Title: Generative AI uses more energy on non-English text

Article References: Generative AI uses more energy on non-English text. Original publication (Bibliographic details unavailable.)

Image Credits: AI Generated

DOI: Not provided

Keywords: generative AI, energy efficiency, multilingual models, computational cost, linguistic bias, tokenization, University of Michigan, AI ethics, language processing, non-English text, Generative, uses