AI Models Emit 50 Times More CO₂ for Some Prompts, Study Finds

Researchers from the Hochschule München University of Applied Sciences in Germany recently conducted a study to measure and compare the CO2 emissions of different large language models (LLMs) when answering a set of standardized questions. The findings highlight a critical trade-off between the accuracy of the models and their environmental impact, emphasizing the importance of making informed choices about AI usage. Key Insights from the Study The study evaluated 14 LLMs, ranging from seven to 72 billion parameters, on 1,000 benchmark questions across various subjects. Parameters are the components that LLMs use to learn and process information. The focus was on two types of models: reasoning-enabled models, which use additional tokens for internal computation and reasoning, and concise response models, which generate answers directly. Reasoning-enabled models, on average, required 543.5 thinking tokens per question, while concise models only needed 37.7 tokens. These “thinking tokens” are intermediate steps that reasoning models take to ensure more detailed and accurate responses. However, the higher token requirement translates to substantially higher energy consumption and CO2 emissions. For instance, the reasoning-enabled Cogito model, with 70 billion parameters, achieved 84.9% accuracy but produced three times more CO2 emissions than similar-sized models that generated more concise answers. Environmental Impact Variances The subject of the questions also influenced the CO2 emissions. Questions that demanded complex reasoning, such as those involving abstract algebra or philosophy, led to emissions up to six times higher than more straightforward subjects like high school history. This suggests that the type of task can significantly affect the environmental footprint of LLMs. The researchers noted that none of the models that kept emissions below 500 grams of CO2 equivalent reached an accuracy rate higher than 80%. This highlights a direct correlation between the level of detail and the amount of energy consumed, indicating an accuracy-sustainability trade-off inherent in these technologies. Real-World Implications To illustrate the real-world implications, the study provided examples of the emissions generated by different models. For instance, answering 600,000 questions with the DeepSeek R1 model (70 billion parameters) would produce CO2 emissions equivalent to a round-trip flight from London to New York. Conversely, the Qwen 2.5 model (72 billion parameters) could answer over 1.9 million questions with similar accuracy while generating the same emissions. This disparity underscores the efficiency differences among models and the potential benefits of choosing the right model for the task. Practical Recommendations The researchers hope their study will encourage users to be more selective and thoughtful about their AI use. They suggest that users can significantly reduce emissions by prompting AI to generate concise answers or by using high-capacity models only for tasks that genuinely require their power. Understanding the exact CO2 cost of AI-generated outputs can help users make better-informed decisions, such as avoiding casual or frivolous queries that do not necessitate complex reasoning. Caveats and Generalizability While the study provides valuable insights, the researchers acknowledged several limitations. The emissions are highly dependent on the structure of local energy grids, and the hardware used in the study may vary from that employed in other settings. Additionally, the generalizability of the results is limited by the specific models and questions examined. Despite these caveats, the study's findings offer a clear direction for reducing the environmental impact of LLMs. Industry Reaction and Company Profiles Industry insiders are taking note of the study's findings, recognizing the need for a more sustainable approach to AI development and usage. Tech companies are beginning to explore ways to optimize their models for both performance and energy efficiency. For example, Google has been investing in renewable energy sources to power its data centers, which could mitigate the environmental impact of its LLMs. Similarly, Microsoft is researching methods to reduce the computational overhead of reasoning models. The Hochschule München University of Applied Sciences, where the research was conducted, is known for its interdisciplinary approach to applied sciences, including a strong focus on sustainable technologies. The institution's commitment to exploring the environmental implications of emerging technologies aligns with broader efforts to address climate change through technological innovation. In conclusion, the study reveals a significant environmental cost associated with the operation of reasoning-enabled LLMs. By adopting more conscious practices, users and tech companies can mitigate this impact without compromising on the utility and accuracy of AI technologies. As AI continues to integrate into everyday life, the balance between accuracy and sustainability will be crucial in guiding future developments.