Neural networks can use 1.4–180 times more energy to process texts in other languages than for similar queries in English. This conclusion was reached by engineers at the University of Michigan who studied the performance of several generative models.
The results were presented at the EMNLP conference of the Association for Computational Linguistics.
The most noticeable difference was found in languages with non-Latin scripts and a limited amount of digital material for training AI. For example, processing queries in Hindi requires on average 7.9 times more energy, and in Bengali 15.3 times more, than in English. For Western European languages, the difference usually does not exceed twofold.
The researchers also established that increased energy consumption may be accompanied by a decrease in the accuracy of answers. By their estimate, around 5.2 billion people use languages whose processing requires more than twice as much energy.
The experiments were carried out with the Qwen 3, Gemma 3 and Llama 3.1 models. They were given tasks in 122 languages, as well as mathematical problems and open-ended questions. Energy consumption was measured using the ML.Energy system.
The scientists cite the specifics of tokenisation – the process of dividing text into small fragments that the neural network processes – as one of the reasons for the difference. Some scripts generate significantly more tokens to convey the same meaning. This increases the load on memory and computing resources.
Among possible solutions to this problem, the authors of the study name the development of models adapted to specific languages and the translation of queries into English before processing.
“Developers should create more resources for languages with an energy and cost gap, including involving native speakers in translating training materials,” noted the study’s co-author, head of the AI laboratory at the University of Michigan, Rada Mihalcea.
