Michigan study: LLMs burn 1.4 to 180 times more energy on non-English prompts, with Hindi at 7.9x and Bengali at 15.3x
- University of Michigan researchers measured Qwen 3, Gemma 3 and Llama 3.1 and found generative AI uses 1.4 to 180 times more energy on non-English text, with Hindi at 7.9x and Bengali at 15.3x the energy of an equivalent English prompt.
- About 5.2 billion people would spend more than twice the energy to prompt in their first language, based on the team's measurements combined with Ethnologue demographic data.
- Most of the gap comes from output length, not energy per token: tokens cluster near 0.1 J across languages, but models loop and repeat text (one sentence 45 times in the worst case), so per-token pricing charges non-English users more for worse answers.
- First author Naihao Deng says the languages that burn more energy also tend to get lower accuracy from the model, a double penalty that could price people out of AI and make companies reluctant to serve those languages.
- The authors recommend training models optimized for specific languages or adding translation pipelines to services; the work was presented at EMNLP and partly funded by NSF and DARPA.
Hacker News 의견들
Half the article goes by before it tells you which models they tested: Qwen 3, Gemma 3, Llama 3.1. That belongs in the first paragraph.
The abstract does say the disparity holds across models, so the placement is fine in the paper. The news writeup and the headline are the weak part.
Those models were relevant 18 months ago. Is there any reason not to rerun the benchmarks on current models before submitting?
How would you even measure energy for models you only reach through Anthropic or OpenAI APIs? You can't run them on your own box and watch the wattmeter.
Does the model think in English internally? In the DeepSeek app I can see the reasoning tokens, and when I ask in Korean it often switches to English mid-way and translates back.
Nobody really knows what happens below the token level. I read that the early layers act like translation, the middle does the thinking, and the last layers translate back.
Google AI overviews refusing to answer in English unless I add 'answer in English only' is maddening.
Counting native speakers only penalizes English and French and makes them look efficient. There are about 3x more non-native speakers, and a lot of training text is written by them.
The presentation is rough but the table on page 6 is gold: the worst cases are token explosion, one sentence repeated 45 times. Almost everything clusters at 0.1 J/token with outliers up to 4.8 J/token, 900 runs per language.
I reproduced chunks of it on our energy-priced models. On Qwen3.8-27B, English was 62 J per request, Shan 1407 J, Tibetan 1455 J, and Tigrinya scored 0/76 on the strict parser. Chinese and Russian actually came out cheaper than English on the DeepSeek model.
That Chinese result is the interesting bit. I would have expected double-byte scripts to cost more, but these are Chinese-trained models.
Different tokenizers per script would fix a lot of this.
What surprised me is that the gap is mostly output length, not energy per token. With per-token pricing that is a fairness problem, not just an efficiency one.
Doesn't this mean there is very little transfer between languages? Only reading and writing should differ. I would like to see a run where they post-train on translations of the same texts.
I already prompt in English and translate the output when I need another language. Most multilingual LLM users figured that out a while ago.