Some of your AI queries can emit 50 times more CO₂ than others, study finds

AI Answers Come at a Carbon Cost, German Study Finds (Image: Canva)

In a world where AI seems to know everything, the true price of its knowledge might be heavier than expected. A new study from Germany reveals that asking complex questions to AI could be increasing our carbon footprint more than we realise.

‘Thinking’ burns more energy

Researchers at Hochschule München University of Applied Sciences evaluated 14 large language models using 1,000 standardised questions. They discovered that models designed for detailed reasoning produce far more carbon emissions than those giving short, direct answers.

Models like Cogito, with 70 billion parameters, produced up to 50 times more CO₂ than simpler response models. Though Cogito reached a high accuracy of 84.9%, it came with a carbon cost nearly three times that of other models of similar size. The study found that reasoning-heavy models generated an average of 543.5 tokens per answer. In contrast, concise models used just 37.7 tokens.

These extra tokens, referred to as ‘thinking tokens’, significantly increase the energy needed for each query. More tokens lead to more processing, which means more emissions. But a longer answer doesn’t always mean a more correct one.

Subject matters and sustainability

The type of question also plays a role in emissions. Abstract subjects like philosophy and algebra pushed emissions up to six times higher than simpler queries, such as high school-level history.

The research highlights a clear trade-off between accuracy and sustainability. No model emitting less than 500 grams of CO₂ equivalent achieved over 80% accuracy on the test set.

Making smarter AI choices

Lead researcher Maximilian Dauner stressed the need for mindful AI use. He said that users can lower emissions by asking for shorter answers or using powerful models only when needed.

The study pointed out that using DeepSeek R1 to answer 600,000 questions would release the same emissions as a return flight from London to New York. On the other hand, Qwen 2.5 could answer over 1.9 million questions with similar accuracy and the same carbon output.

Researchers noted that the findings could vary depending on the hardware and regional energy sources. Still, they believe transparency about the CO₂ cost of AI tasks can help users make better choices.

“If people knew how much carbon goes into their AI selfies or trivia answers, they might use the tech more wisely,” Dauner said.

The study was published in Frontiers in Communication.