Why do some AI answers burn 50× more CO₂ than others?

AI models consume energy and emit CO₂ each time they process a question, but the level varies widely depending on the model’s reasoning approach.

German researchers discovered that AI models using "reasoning" to generate extensive step-by-step content before giving a final answer can emit up to 50 times more carbon than models that provide short, direct responses.

This happens because reasoning models produce many more tokens internally during processing. Tokens are the words or fragments transformed into numerical data that models analyze. More tokens mean more computation, which increases energy use and CO₂ output.

What causes the huge difference in AI CO₂ emissions?

The main factor behind the large disparity is the number of "thinking" tokens generated. Reasoning models produced an average of 543.5 additional tokens per question, compared to just 37.7 tokens from concise-answer models.

Since generating each token requires computing power, reasoning-enabled AI inherently demands much more energy, pushing CO₂ emissions dramatically higher.

Did you know?
The term "Artificial Intelligence" was coined in 1956 at a conference at Dartmouth College, which is considered the birth of AI as a research field.

How does reasoning in AI models affect carbon output?

Reasoning models are designed to deliver detailed, transparent explanations, often necessary for complex tasks such as algebra or philosophy. However, this added computation results in up to 50 times more CO₂ emissions per query.

The study found that the highest accuracy model, Cogito, emitted three times more CO₂ than other similarly sized models that provided briefer responses. This underscores a trade-off between accuracy and environmental impact.

ALSO READ | University of Illinois builds scalable modular superconducting quantum processors

Comparing accuracy and sustainability in AI answers

Not all detailed reasoning improves answer accuracy. Some models that produce high emissions do not necessarily provide better answers, creating a dilemma between environmental sustainability and AI performance.

Researchers observed that models emitting under 500 grams of CO₂ equivalent rarely surpassed 80% accuracy across 1,000 test questions.

What can users do to reduce AI’s carbon footprint?

The German researchers recommend prompting AI for concise answers when possible and reserving larger, reasoning-enabled models for cases requiring profound analysis. Choice of model also affects emissions; some large models answer significantly more queries per equivalent CO₂ emitted.

Being aware of AI’s carbon impact empowers users to make informed decisions, balancing their need for detailed answers with environmental responsibility.