New study warns: AI’s carbon footprint varies wildly by model and query
A fresh analysis published in Frontiers reveals some AI prompts emit up to 50 times more CO2 than others—fueling a growing environmental concern around AI’s energy use.
Researchers assessed 14 large language models (LLMs), ranging from 7 billion to 72 billion parameters. The bigger and more complex the model, the more tokens they generated to produce answers, driving up emissions. Reasoning-powered models averaged 543.5 “thinking” tokens per question—a stark contrast to simpler models needing just 37.7 tokens. 
The type of question matters too. Straightforward topics like high school history generated six times fewer emissions compared to dense subjects like abstract algebra or philosophy. Accuracy also comes at an environmental cost. The most accurate model, Cogito, emitted three times the CO2 of similar-sized, concise-answer models.
The study highlights a “accuracy-sustainability trade-off.” None of the models keeping emissions below 500 grams of CO2 hit more than 80% accuracy.
Model efficiency varies sharply as well. DeepSeek R1 (70B parameters) requires as much CO2 to answer 600,000 questions as a round-trip London-New York flight. Meanwhile, Qwen 2.5 (72B parameters) can answer nearly 2 million questions with similar emissions and accuracy.
Hochschule München’s Maximilian Dauner urges users to consider AI’s environmental cost:
“If users know the exact CO₂ cost of their AI-generated outputs, such as casually turning themselves into an action figure,"
"they might be more selective and thoughtful about when and how they use these technologies.”
Data centers already sucked up 4.4% of U.S. electricity in 2023 and are projected to hit 6.7–12% by 2028, driven in part by AI demand. Worldwide data centers have exploded from 500,000 in 2012 to over 8 million in September 2024.
Simple Google searches consume far less power than an average ChatGPT request, which uses 10x more electricity, according to the International Energy Agency.
This study spotlights the urgent need to balance AI progress with environmental sustainability as AI adoption skyrockets globally.
