Breaking News: The Environmental Impact of AI Models
AI Models and Their Carbon Footprint
In a groundbreaking study published in Frontiers in Communication, researchers have shed light on the environmental impact of artificial intelligence (AI) models. The study reveals that the size and complexity of AI models directly affect their carbon footprint, with larger models consuming significantly more energy and generating higher CO2 emissions.
The research team measured the number of tokens generated by AI models, which are digital information units that reflect the workload. They also tracked the power consumption of the local computers running these large language models (LLMs), using a conversion factor of 480 grams of CO2 equivalents per kilowatt-hour.
Token Generation and Computing Power
The study found that larger AI models, which deliver more correct answers, require more computing power. This is because these models generate additional “thinking tokens” for complex tasks, reflecting their effort in “rethinking.”
For instance, the smallest variant of Deepseek-R1 generated up to 14,187 tokens for a single math problem, compared to just a handful of tokens for classic models providing simple answers.
CO2 Emissions and Model Size
The CO2 footprint of AI systems depends on both the model size and the type of questions asked. “We see a clear compromise between correctness and sustainability: None of the large AI models with more than 80 percent correct answers remained with less than 500 grams of CO2 emissions for the thousand questions,” reports Dauner.
The smallest model, Qwen with only seven billion parameters, emitted around 27.7 grams of CO2 for all 1,000 questions, but was only correct for one third of all questions. In contrast, the largest version of Deepseek-R1 with 70 billion parameters emitted more than 2,000 grams of CO2 but provided around 80 percent correct answers.
Reasoning Models and CO2 Emissions
Reasoning models, which handle more complex tasks, cause significantly higher CO2 emissions. The study found that reasoning-capable models can emit up to 50 times more CO2 than simple LLMs.
For example, if Deepseek-R1 answers 600,000 user questions, it generates as many greenhouse gases as a return flight from London to New York. In contrast, the non-reasoning model QWEN 2.5 can answer three times as many questions for the same amount of CO2 with similar correctness.
Implications for Everyday AI Usage
The study suggests that users can significantly reduce emissions by choosing simpler AI models for straightforward tasks and limiting the use of powerful models to tasks that truly require high computational power.
“Users can significantly reduce emissions if they promptly ask the AI systems to provide simple, short answers without thinking for a long time,” says Dauner. “In addition, we should limit the use of the strongest models to the tasks that really need this computing power.”
