AI chatbot reasoning increases carbon emissions, study finds

Queries demanding logical reasoning from AI chatbots like OpenAI’s ChatGPT produce significantly more carbon emissions than simpler requests, according to a new study published in Frontiers.

Researchers from Germany’s Hochschule München University of Applied Sciences analysed 14 large language models (LLMs) and found that the environmental impact of chatbot queries varies greatly depending on complexity.

The study shows that questions involving abstract thinking — such as those in algebra or philosophy — generate up to six times more emissions than simpler subjects like high school history.

Reasoning-heavy models create an average of 543 tokens per query, compared to just 40 for concise-response models, resulting in a considerable increase in energy use and carbon dioxide output.

Accuracy also comes at a cost. The most precise model in the study, Cogito, reached 85% accuracy but emitted three times more carbon than similar-sized models that delivered shorter answers.

Researchers highlight a clear trade-off between sustainability and performance: none of the low-emission models achieved more than 80% accuracy across 1,000 benchmark questions.

To reduce environmental impact, researchers urge users to limit complex prompts and opt for concise queries when possible. They estimate that prompting DeepSeek R1 to answer 600,000 questions could generate as much carbon as a round-trip flight from London to New York.

In contrast, Alibaba Cloud’s Qwen 2.5 model answered three times as many queries with the same emissions output. The study calls for more responsible AI usage to manage the growing environmental footprint of large-scale language models.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!