Artificial intelligence, particularly generative AI, is becoming integral to various aspects of daily life, from search engine optimization to automated email composition. However, the computational intensity of these AI models has notable environmental implications. According to a report by the Energy Department, AI-driven data centers could potentially increase their share of the national electricity consumption from 4.4% to 12% by 2028, necessitating increased reliance on energy sources such as coal and natural gas.
A recent study published in the journal Frontiers in Communication provides a quantitative analysis of the environmental impact of AI chatbots. The research evaluates 14 large language models (LLMs) by measuring their energy consumption and subsequent carbon dioxide emissions. The study highlights that models with larger parameters consume exponentially more energy, correlating with higher accuracy until a saturation point is reached.
Maximilian Dauner, the study’s lead author, emphasizes the importance of selecting appropriately sized models for specific tasks. Smaller, less resource-intensive models can efficiently handle simpler queries without significant accuracy trade-offs, optimizing both performance and energy use.
The study involved subjecting each AI model to 500 multiple-choice and 500 free-response questions across five subjects. Energy usage was calculated per model and converted to CO2 equivalents. Initial findings indicate that models responding to logic-based questions, such as abstract algebra, required more computational power, thereby producing more emissions compared to fact-based queries, like those in history.
Furthermore, the research shows that chatbots employing detailed, step-by-step reasoning tend to have higher energy demands than their peers, without significantly improving accuracy. For instance, the highest-emitting model, DeepSeek-R1, demonstrated similar accuracy levels to models with a quarter of its emissions.
The data suggests that AI models optimized for reasoning tasks are less energy-efficient. For example, DeepSeek-R1 emitted 4.8 grams of CO2 per abstract algebra answer, while other models like Cogito 70B and Qwen2.5 72B emitted significantly less per response across various subjects.
The carbon footprint of these models is notably higher for subjects like high school mathematics and abstract algebra, with emissions per answer reaching up to 4 grams of CO2 for certain models. In contrast, fact-based subjects such as international law and philosophy showed lower emissions, underscoring the need for task-specific model deployment.
The study underscores the necessity for strategic model selection in AI applications to balance accuracy with environmental sustainability. By leveraging data-driven insights, developers and investors can optimize the deployment of AI models to enhance efficiency and minimize ecological impact, aligning technological advancement with environmental responsibility.
