
A new study published in Frontiers in Communication reveals that AI emissions can rise up to 50 times. It occurs when chatbots employ complex reasoning instead of giving short responses. At Hochschule München University of Applied Sciences, researchers under the direction of Maximilian Dauner evaluated 14 large-language models (LLMs). It includes Cogito and DeepSeek.
They found that response complexity, particularly in subjects like abstract algebra and philosophy, sharply increases energy use and carbon release. Additionally, the results demonstrate the increasing carbon footprint of reasoning models. It aims for greater accuracy, but frequently at the expense of sustainability.
Why Complex Prompts Spike AI Emissions Rapidly
The study shows that AI emissions are not uniform across all chatbot responses. LLMs maintained low emissions when asked basic high school history questions. However, emissions increased sixfold when asked to use abstract reasoning. This is because reasoning models require more computation and energy due to their increased use of processing tokens.
Every chatbot responded to 1,000 benchmark questions in five different subjects. The subjects include high school history, international law, philosophy, high school math, and abstract algebra. For abstract algebra, DeepSeek R1 7B generated over 6,700 tokens per question. In contrast, simple prompts required fewer than 100. Thus, the token count is the main source of increasing emissions during advanced tasks, which directly affects the carbon impact.
Reasoning Models Raise Emissions for More Accuracy
According to the researchers, each question generated by the reasoning models generated an average of 543 thinking tokens. Concise-response models used just 37. Even with their thorough responses, the complexity did not always translate into greater accuracy. Indeed, across the dataset, models with emissions below 500 grams of CO₂ per response never achieved more than 80% accuracy.
Cogito, one of the leading reasoning models, achieved nearly 85% accuracy but produced three times more emissions than its smaller peers. The authors emphasized that the impact of carbon accurately scales, exposing a trade-off between sustainability and intelligence. Additionally, their data clearly shows that longer, more rational responses are often time-consuming and ineffective.
According to a study of visual data comparing different models, emissions rise most quickly in mathematical and philosophical prompts. Posts around the study quickly spread on tech forums, sparking debate about the future of eco-conscious AI. As a result, many users were surprised to learn that providing detailed answers could be environmentally costly.
AI Emissions Demand Smarter, Sustainable Innovations
The research team emphasizes the urgent need to develop smarter and leaner reasoning models. Optimizing token generation without losing accuracy could ease the carbon impact of next-gen AI systems. For sustainable deployment, this balance will be crucial, particularly in high-demand industries like legal and educational services.
A dual strategy is suggested by Dauner and his group: advance model design while, when practical, streamlining prompts. They point out that even with today’s capabilities, response brevity in challenging tasks. Abstract algebra is one of the tasks that can cut AI emissions without sacrificing accuracy. This makes generative AI more environmentally friendly.
Final Thoughts
The findings leave us with a challenging question: Can we build AI that thinks deeply without burning the planet? Currently, there is a real trade-off between environmental cost and performance. According to the study, the industry should reconsider how much carbon should be expended on each chatbot interaction.
Additionally, developers now have to give energy-efficient reasoning methods top priority when creating new models. It might also be necessary for regulatory agencies to establish emission guidelines for AI platforms that are used extensively. Ultimately, sustainability must become a core metric alongside accuracy in AI development.