A recent study published in Frontiers in Communication by German researchers reveals that large language models (LLMs) delivering higher accuracy also generate significantly higher carbon emissions. The findings suggest a direct trade-off between AI performance and environmental impact.
Analysing 14 open-source LLMs using 1,000 test questions, the researchers found that larger, more capable models—such as DeepSeek—produced the highest emissions. Models that perform multi-step “reasoning” also consumed more energy than simpler models.
While some exceptions emerged, such as Cogito 70B achieving high accuracy with a slightly lower carbon footprint, the overall trend was clear: better performance typically comes with a greater environmental cost.
Lead author Maximilian Dauner emphasised the need for task-specific AI usage. “We don’t always need the biggest, most heavily trained model to answer simple questions,” he said. “Smaller models can handle many tasks efficiently.”
The study raises broader concerns about the growing energy demand of AI systems, particularly as their use expands across everyday digital services.
