Whenever we ask for something to an artificial intelligence, it replies. But behind each answer, even the incorrect or unsolicited ones, there is an invisible process: the generation and processing of token. The tokens, or words or fragments of words converted into numerical strings to be interpreted by Large linguistic models (LLM)consume energy. And this consumption inevitably involves carbon dioxide emissions (CO₂).
A team of German researchers recently measured and compared the emissions generated by several LLM already trained, answering a set of standardized questions. The results, published in the magazine Frontiers in Communicationare surprising: some types of prompts, or the applications or instructions given to the AI, can cause up to 50 times more emissions compared to others.
According to the study, conducted by Maximilian Dauner of the Hochschule München University of Applied Sciences, the models that use explicit reasoning processes to get to an answer – so -called Reasoning Models– They are much more energetic than those that provide concise and direct answers.
Analyzing 14 LLM models different, with dimensions varying from 7 to 72 billion parametersthe researchers have placed 1,000 questions about heterogeneous issues. The parameters, in this context, represent the fundamental units through which the model learns and reworks information.
The results show that the models based on reasoning generate on average 543.5 token “of thought” For each question, while the synthetic models only employ them 37.7. THE “Thinking token” These are those intermediate passages generated before reaching the final response and, as expected, involve a much higher energy load.
More token equals multiple emissionsbut this does not necessarily guarantee more correct answers. The accuracy, in fact, does not always depend on the complexity of the reasoning.
Energy efficiency also changes according to the type of demand and the model used
The most accurate model of the study was Cogitoequipped with 70 billion parameterswhich has reached an accuracy of the84.9%. However, he produced tripled emissions Compared to models of similar dimensions that offered shorter answers, as Dauner explained:
We clearly witness a compromise between accuracy and sustainability.
None of the models it has maintained emissions lower than 500 grams of equivalent co₂ he managed to overcome the80% of accuracy. There Equivalent It is the standard unit used to measure the climatic impact of the various greenhouse gases.
Also The nature of the demand significantly affects emissions: complex questions, such as those of Abstract algebra or philosophygenerate up to six times more emissions Compared to simpler questions, such as those of History of high school.
How to use artificial intelligence in a conscious and sustainable way
The researchers hope that these results induce users to adopt a more reasoned use of AI “, as Dauner pointed out:
It is possible to significantly reduce emissions simply by asking for concise answers or by reserving the most powerful models only to the tasks that justify their use.
To give a concrete example: if you use the model Deepseek r1 (70 billion parameters) to respond to 600,000 questionsemissions comparable to a Round trip flight London – New York. In comparison, Qwen 2.5 (72 billion parameters) can respond to About 1.9 million questions maintaining similar accuracy levels, but with it same environmental impact.
It should be borne in mind that the results of the study are also influenced by other factors, such as The hardware usedthe Features of the local electricity grid he is models analyzedelements that can limit the generalization of data.
Finally, Dauner said:
If users were fully aware of the environmental cost of the responses generated by the AI - even for trivial uses how to transform themselves into a video game character – they would be able to adopt a more selective approach and responsible for the use of these technologies.
