Generative AI uses more energy on non-English text
futurity.org

Generative AI uses more energy on non-English text

Archived — this story has rotated out of today’s deck. It is kept here in full.

The gist

A University of Michigan study found generative AI uses 1.4 to 180 times more energy for non-English text.
The gap can price non-English speakers out of AI and discourage companies from serving their languages.

Background

The study comes from researchers at the University of Michigan's AI Lab, led by first author Naihao Deng, a doctoral student in computer science and engineering, and co-author Rada Mihalcea, who heads the lab. They measured how much energy the Qwen 3, Gemma 3 and Llama 3.1 models used on tasks in 122 languages, including math problems and open-ended questions. The researchers trace the disparity to tokenisation: some writing systems produce far more tokens to convey the same meaning, loading memory and computing resources. The work was presented at the Empirical Methods in Natural Language Processing (EMNLP) conference of the Association for Computational Linguistics and partly funded by the National Science Foundation and DARPA (the US Defense Advanced Research Projects Agency).

How it unfolded

  1. Nov 5, 2026University of Michigan Engineering publishes the study on its news site.
  2. Nov 8, 2026TMCars News reports the findings, noting they were presented at the EMNLP conference of the Association for Computational Linguistics.
  3. Nov 9, 2026Futurity reports the study and its presentation at the conference.
  4. Nov 10, 2026Michigan Engineering posts that generative AI uses 1.4 to 180 times more energy to produce text outside English.
  5. Nov 11, 2026The University of Michigan College of Engineering repeats the 1.4 to 180 times figure on LinkedIn.
  6. Nov 14, 2026A Hacker News discussion thread links to the story.

Key terms

tokenisation
A basic unit of text that a language model processes, such as a word, part of a word, or symbol.
Empirical Methods in Natural Language Processing (EMNLP)
A major academic conference for research in natural language processing and artificial intelligence, organized by an ACL special interest group.
DARPA
A U.S. Department of Defense agency that develops emerging technologies for military use.

Source: Wikipedia

Who’s saying what

Expert
First author Naihao Deng says languages that consume more energy also tend to get lower accuracy from the models, and that this double penalty may price some people out of AI and make companies reluctant to serve certain languages.
Official
Co-author Rada Mihalcea says developers should build more resources for languages with an energy and cost gap, including involving native speakers in translating training materials.
Analysts
The researchers recommend optimizing services for more languages, routing queries to language-specific models, or translating prompts to English first when the translation is reliable.

Still unverified

TMCars News states that its article was translated by artificial intelligence and may contain inaccuracies.

Sources

See today’s stories in the app Gistline — free on the App Store