Generative artificial intelligence uses more energy on texts in languages other than English, researchers announced at the Empirical Methods in Natural Language Processing conference in Budapest.
According to researchers at University of Michigan Engineering, generative AI uses more energy (from 1.4 to as much as 180 times more) when processing and producing texts in languages other than English, and sometimes gives less accurate answers, potentially deepening existing global inequalities.
Higher energy costs raise the barrier to using artificial intelligence for billions of people worldwide, who must either pay more or switch to a language other than their native one.
The disparity in energy use is greatest for languages that do not use the Latin alphabet and have less digital content available for learning. Most Western European languages use about twice as much energy. But Hindi and Bengali, which are spoken by hundreds of millions of people, require 7.9 and 15.3 times more energy, respectively, than an equivalent message in English.
“Languages that typically use more energy also typically have lower accuracy of model results,” said Naihao Deng, a doctoral student in computer science and engineering at the University of Michigan AI Lab and first author of the study.
The researcher noted that most people will want to use AI models by asking questions in their native languages and may feel compelled to do so. He added that the “double penalty” (higher costs and worse results) makes it more likely that some groups of people may be excluded from access to AI because of prohibitive costs, and that companies may be reluctant to make their models available in certain languages.
Currently, about 5.2 billion people must use more than twice as much energy to generate text in their native language as English speakers do, according to measurements by the researchers and demographic data from the Ethnologue linguistic database.
The researchers uncovered the disparity in energy use and accuracy by measuring the amount of energy consumed 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 from Belebele, a dataset used to evaluate how well AI models comprehend written text 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 test for disparities in other types of tasks, the researchers also asked the models to answer a set of elementary school math word problems and open-ended user questions from other test datasets. Because those datasets are not multilingual, the researchers translated the questions from English into seven other languages using Meta’s No Language Left Behind translation tool.
On average, all of the models used less energy for English-language text. The difference in energy use persisted at larger model sizes and across different types of computer processors.
When working in other languages, the models used more energy because they had to generate more tokens, the basic units of data used by artificial intelligence. Tokens can represent whole phrases and words or individual syllables and characters. AI models convert prompts into tokens before processing them, and then convert the output tokens into human language to produce the final answer.
The problem is more pronounced 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 base characters that change form and combine with accents, creating up to millions of unique characters.
More tokens take up more computer memory and use more energy. With more memory taken up by tokens, the model has less space available to process information. It therefore processes tokens in smaller batches, which uses more energy per token.
More tokens can also mean higher direct costs, because companies often price access to a model based on the number of tokens generated.
To close the cost gap, the researchers recommend that companies optimize their services for more languages. That could mean routing users’ queries to models tailored to specific languages. Companies could also translate users’ messages into English before passing them to a more general model, provided the translation is reliable.
“AI developers should build more resources for the languages where there are disparities in cost and energy use,” said Prof. Rada Mihalcea, director of the University of Michigan AI Lab (U-M AI Lab) and the study’s corresponding author. She said that “crowdsourcing native speakers – to help translate materials or create resources to get enough training data” could help improve efficiency.
Crowdsourcing is a method in which a company, institution or organization assigns specific tasks, ideas or services to a broad, unidentified group of people (usually via the internet) instead of only its own employees.
Paweł Wernicki (PAP)
