Did you know that digital pollution exists? Did you know that we are polluting the planet every time we use conversational AI? A scientific study conducted by researchers at the Munich Center for Digital Sciences and AI (MUC.DAI), explains the environmental impact of the use of artificial intelligence, and the ecological footprint it generates. In a world where more and more measures are being taken to combat climate change, we are not aware of the carbon emissions generated by the use of artificial intelligence.
According to the study, emissions vary depending on how much we make the AI think, so some questions are more polluting than others. Perhaps the solution is not to stop using this resource, but to take into account how it affects the relationship between technology and the environment, and the energy consumption generated by AIs.
Munich Center for Digital Sciences and AI study
Frontiers in Communication published this study led by Maximilian Dauner and Gudrun Socher, which measured precisely how much conversing with an AI pollutes the environment. The results showed that some questions can generate 50 times more CO2 emissions than others. The research tested 14 different language models. Through this, they were able to measure the difference in emissions according to the tokens generated.
Do questions or answers contaminate?
The answer is neither. It has been found that what really pollutes is the process of “thinking” the answer. The more you have to think to solve the question, the more you pollute. According to the study, a simple model generates an average of 38 tokens to answer a multiple-choice question. However, a question that requires “thinking” can generate up to 540 tokens. The test conducted in the study included the Cogito 70B model and the Qwen 2.5. The former was the most accurate with 84.9% accuracy, but it was also the most polluting; over 1,300 grams of CO2. The Owen 2.5, on the other hand, generated 426 grams of CO”, giving shorter responses.
Some questions contaminate more than others
Although it has become clear that it is the time spent “thinking” that consumes energy and therefore pollutes, the subject on which the question is asked also has an influence. The study has shown that there are academic subjects that reflect this trade-off. The AI considers questions related to World History to be much easier to answer, so they pollute less. However, subjects such as Abstract Algebra or Philosophy force the models to undergo longer reasoning processes, so the number of emissions is much higher.
DeepSeek R1 70B, the most polluting model
The study has shown that the most polluting model, without a doubt, is the DeepSeek R1 70B. It has been calculated that if it were asked 600,000 questions, it would pollute exactly the same as a flight from London to the United States. Although individual level emissions are very low, the problem lies in the massive use of artificial intelligence models. The researchers estimated that the current use of AI is equivalent to the total electricity consumption of Ireland. They do not intend to demonize the use of AI, but perhaps to think before using it. At the moment, no model has been developed that is very accurate and sustainable; the higher the accuracy, the higher the CO2 emissions.
Reduce the impact
The researchers recommend a number of ideas to reduce the impact we generate as individuals when using AI models.
- Use small models for everyday tasks.
- Avoid open-ended prompts.
- Encourage concise default response options.
At the enterprise level, they recommend the use of data centers with renewable energy, although this seems almost utopian when the reality is that companies such as Microsoft, OpenAI or Meta already have nuclear plants to support electricity consumption.