If the famous “Prompts” or instructions that we give to an AI such as Chatgpt or Depseek determine the final result, they also have another secondary incidence: the level of pollution it can generate.
To generate the answers, the model uses tokens, that is, words or fragments of words that become a chain of numbers that the LLM (large language models) can process.
This conversion, as well as other computer processes, produces CO₂ emissions. However, many users are unaware of the considerable carbon footprint associated with these technologies. Now, a team of German scientists, led by Maximilian Dauner, measured and compared the CO₂ emissions of different LLM already trained through a set of standardized questions. The results have been published in Frontiers in Communications.
“The environmental impact of questioning the trained LLMs is strongly determined by its reasoning approach, since explicit reasoning processes significantly increase energy consumption and carbon emissions -explains Dauner in a statement -we discovered that The models based In reasoning they produced up to 50 times mtos CO₂ emissions that the concise response models ”.
Dauner’s team Evaluated 14 LLM models with between 7 billion and 72 billion parameters in 1000 questions reference on various subjects. The parameters determine how the LLM learn and process information.
The reasoning models, on average, generated 543.5 tokens of “thought” per question, while the concise models required only 37.7 tokens per question. Thought tokens are additional tokens that the LLM of reasoning generate before generating an answer. A greater tokens footprint always implies higher CO₂ emissions. However, this does not necessarily mean that the resulting responses are more correct, since the complexity of the details is not always essential for correction.
The most precise model was the cogito model, based on reasoning, with 70 billion parameterswhich reached an accuracy of 84.9 %. The model produced three times more CO₂ emissions than similar size models that generated concise responses.
“Currently, we observe a clear compensation between precision and sustainability inherent to the LLM technologies -adds Dauner -none of the models that maintained the emissions below 500 grams of CO₂ equivalent managed an accuracy greater than 80 % when answering the 1000 questions correctly. The equivalent co₂ is the unit used to measure the climate impact of various greenhouse gases ”.
The theme that it was also resulted in significantly different levels of CO₂ emissions. The questions that required long reasoning processes, for example, abstract algebra or philosophy, generated emissions up to six times greater than those of simpler subjects, as high school history. The study indicates that the objective is to help users to make more informed decisions about their own use of AI.
“It is possible significantly reduce emissions by inciting AI to generate concise responses or limiting use of high capacity models to tasks that really require it, ”says Dauner.
The choice of the model, for example, can make a significant difference in CO₂ emissions. For example, if Deepseek R1 (70,000 million parameters) answered 600,000 questions, They would generate CO₂ emissions equivalent to those of a round trip from London to New York. On the other hand, Qwen 2.5 (72,000 million parameters) can answer more than triple questions (approximately 1.9 million) with similar precision rates, generating the same emissions.
The study details that its results They could be affected by the election of hardware used in the studyan emission factor that can vary regionally depending on the combinations of the local electricity network, and the models examined. These factors could limit the generalization of the results.
“If users know the exact cost of CO₂ of the results generated by AI, such as becoming coincidentally an action figure, They could be more selective and reflective on when and how they use these technologies”, Concludes Dauner.