Each time I interact with ChatGPT, I consume energy—what does that really mean? A new study has highlighted the environmental costs of using large-scale language models (LLMs) and provided insights on how users can minimize their carbon footprints.

German researchers evaluated 14 open-source LLMs, ranging from 14 to 72 billion parameters, administering 1,000 benchmark questions to assess the CO2 emissions generated in response to each.

They discovered that utilizing internal reasoning to formulate answers can result in emissions up to 50 times greater than those generated by a brief response.

Conversely, models with a higher number of parameters—typically more accurate—also emit more carbon.

Nonetheless, the model isn’t the only factor; user interaction plays a significant role as well.

“When people use friendly phrases like ‘please’ and ‘thank you,’ LLMs tend to generate longer answers,” explained Maximilian Dorner, a researcher from Hochschule München Applied Sciences University and the lead author of the study, to BBC Science Focus.

“This results in the production of more words, which leads to longer processing times for the model.

The extra words don’t enhance the utility of the answer, yet they significantly increase the environmental impact.

“Whether the model generates 10,000 words of highly useful content or 10,000 words of gibberish, the emissions remain the same,” said Dorner.

This indicates that users can help reduce emissions by encouraging succinct responses from AI models, such as asking for bullet points instead of detailed paragraphs. Casual requests for images, jokes, or essays when unnecessary can also contribute to climate costs.

The study revealed that questions demanding more in-depth reasoning—like topics in philosophy or abstract algebra—yield significantly higher emissions compared to simpler subjects like history.

Researchers tested smaller models that could operate locally, yet Dorner noted that larger models like ChatGPT, which possess more than 10 times the parameters, likely exhibit even worse patterns of energy consumption.

“The primary difference between the models I evaluated and those powering Microsoft Copilot or ChatGPT is the parameter count,” Dorner stated. These commonly used models have nearly tenfold the parameters, which equates to a tenfold rise in CO2 emissions.

Dorner encourages not only individual users to be mindful but also highlights that organizations behind LLMs have a role to play. For instance, he suggests that they could mitigate unnecessary emissions by creating systems that select the smallest model necessary for accurately answering each question.

“I’m a big supporter of these tools,” he remarked. “I utilize them daily. The key is to engage with them concisely and understand the implications.”

read more:

About our experts

Maximilian Dorner, PhD candidate at Hochschule München Applied Sciences University.