Generative artificial intelligence (AI) that produces text on request uses more energy when working with non-English languages and sometimes produces less accurate answers. These are the conclusions of researchers from the University of Michigan College of Engineering. In their view, this could widen inequality between language groups.
The gap is especially large for languages with non-Latin scripts and small amounts of training material. Depending on the language, models may use 1.4 to 180 times more energy than when working with English. For most Western European languages, the increase does not exceed twofold. Queries in Hindi and Bengali require 7.9 and 15.3 times more energy, respectively.
"Languages that consume more energy also tend to have lower answer accuracy," says Naihao Deng, the study's first author from the University of Michigan AI Laboratory. The work was presented at the Empirical Methods in Natural Language Processing conference.
According to estimates based on measurements and Ethnologue demographic data, about 5.2 billion people use more than twice as much energy to generate text in their native language as native English speakers. The authors propose ways to narrow the gap.
The researchers tested three open models: Qwen 3 from Alibaba, Gemma 3 from Google and Llama 3.1 from Meta. The models were asked multiple-choice questions from the Belebele set to assess reading comprehension in 122 languages. Energy was measured using the ML.Energy framework, a set of tools for such measurements. In addition, the models answered school math problems and open-ended queries; the English questions were translated into seven languages using Meta's No Language Left Behind tool.
On average, all models used less energy for English text. The gap persisted as model size increased and across different processor types.
The reason is that for other languages, models generate more tokens, the basic units of data. A token can represent a phrase, word, syllable or character. The more tokens there are, the more memory they occupy and the more energy is consumed. The model processes them in smaller batches, which increases the cost per token. A larger number of tokens also raises monetary costs, since companies often charge by the number of tokens generated.
To narrow the gap, the researchers advise companies to optimize services for more languages: route requests to models tuned for a specific language, or translate them into English when translation is reliable. "AI developers should create more resources for languages with a gap in energy and cost," says Rada Mihalcea, the study's corresponding author. She also proposes involving native speakers to translate materials and create training data.
The study found a link between language, energy consumption and answer accuracy, but does not prove that a non-English language itself is the cause of the differences.