Regardless of the questions we ask AI, and whether the answers are correct or not, all of these computational processes result in CO2 emissions, which contribute to climate change and global warming. Researchers in Germany measured and compared the CO2 emissions of different AI models, and here's what they found.
Regardless of the questions we ask AI, and whether the answers are correct or not, all of these computational processes result in CO2 emissions, which contribute to climate change and global warming. Researchers in Germany measured and compared the CO2 emissions of different AI models, and here's what they found.
Researchers at the Munich School of Applied Sciences evaluated 14 LLMs with a range of 7 billion to 72 billion parameters on 1,000 questions across a variety of subjects. They found that reasoning models resulted in 50 times more CO2 emissions than others, Tech Xplore reports .
Reasoning models generated an average of 543.5 tokens per question, while models providing concise answers required only 37.7 tokens per question.
A larger token footprint always means higher CO2 emissions. However, this does not necessarily mean that the answers obtained are more correct, as detailed information is not always necessary for correctness.
The most accurate model was the American Cogito reasoning model with 70 billion parameters, which achieved an accuracy of 84.9%. The model produced three times more CO2 emissions than similarly sized models that generated concise answers.
“None of the models that kept emissions below 500 grams of CO2 equivalent achieved more than 80% accuracy when answering 1,000 questions correctly,” the study says. CO2 equivalent is a unit used to measure the climate impact of various greenhouse gases.
The amount of CO2 emissions also depended on the topic of the questions. Questions that required long reasoning processes, such as abstract algebra or philosophy, resulted in 6 times higher emissions than simpler subjects, such as high school history.
The choice of model can also have a significant impact on CO2 emissions. For example, if DeepSeek R1 (70 billion parameters) answers 600,000 questions, this will result in CO2 emissions equal to a round-trip flight from London to New York.
Meanwhile, Alibaba's Qwen 2.5 (72 billion parameters) can answer more than three times as many questions (about 1.9 million) with similar accuracy, while generating the same emissions.
The researchers hope their work will encourage people to make more informed decisions about the use of AI.
“Users can significantly reduce emissions by encouraging AI to generate concise responses or by limiting the use of high-power models to tasks that truly require such power,” the researchers note.
