AI chatbots consume 50 times more energy for long answers

Many people fire off rapid‑fire queries at the latest AI chatbots without a second thought. Each answer taps electricity, and a new study shows that some replies raise the meter by a factor of fifty.

The analysis covered 14 large language models (LLMs). The researchers found that the most verbose engines release as much heat‑trapping gas per answer as running a midsize laptop for an hour, while concise models sip power.

The study was led by Maximilian Dauner at the Hochschule München University of Applied Sciences.

AI answers carry a carbon cost

Every interaction begins with tokens, the word fragments a model turns into numbers so its circuits can reason. More tokens mean more processor cycles and, ultimately, more electricity.

The link between long output and higher emissions is not new. A 2019 study estimated that training one natural‑language model could emit as much carbon as five cross‑country flights.

Recent projections suggest generative AI as a whole could consume 29.3 terawatt‑hours per year, rivaling Ireland’s entire grid use.

Power plants that supply that demand release hundreds of grams of carbon dioxide equivalent for every kilowatt‑hour they generate, with the International Energy Agency placing the world average at about 480 g.

Testing shows big emission differences

Dauner’s team put seven‑ to 72‑billion‑parameter systems through 1,000 standardized questions across philosophy, history, law, abstract algebra, and mathematics. The researchers counted the electricity drawn by an NVIDIA A100 GPU during each run.

Models designed to “think out loud” produced an average of 543.5 thinking tokens before answering, compared with 37.7 tokens for brisk responders. That extra chatter translated into as much as 50 times more emissions per question.

“The environmental impact of questioning trained LLMs is strongly determined by their reasoning approach, with explicit reasoning processes significantly driving up energy consumption and carbon emissions,” said Dauner.

Long answers in algebra caused six times more emissions than short history replies. It was the symbolic reasoning in AI’s long answers, not just length, that raised energy use.

Greater accuracy means more emissions

The 70‑billion‑parameter Cogito model hit 84.9 percent accuracy but released three times more carbon than similarly sized concise systems.

No model that stayed below 500 g of emissions cracked the 80 percent accuracy ceiling. That ceiling also shifted by topic. History questions were easy wins, while algebra stumped nearly every contestant despite burning far more power. 

“Currently, we see a clear accuracy‑sustainability trade‑off inherent in LLM technologies,” Dauner noted.

Size alone was not destiny. Qwen 2.5, a 72‑billion model that keeps its answers short, handled almost two million questions for the same emissions budget that DeepSeek R1 burned on 600,000, with only a small accuracy gap.

Environmental burden of AI use

To put the numbers in context, generating answers with some reasoning-heavy AI models emits as much CO₂ as driving a car for several miles.

For example, DeepSeek R1 (70B) emits over 2,000 grams of CO₂ to answer 1,000 questions, which is roughly the same as burning a quarter of a gallon of gasoline.

Over time, those numbers multiply. Running a reasoning-enabled model to answer questions non-stop over a full day could produce as much carbon as running a refrigerator for two weeks.

This comparison helps highlight the hidden but accumulating environmental burden of everyday AI use.

Avoiding long AI answers

Individual habits matter. A quick setting that asks for a concise answer slices token counts dramatically.

Saving the heavyweight models for code reviews or legal briefs and using lighter variants for trivia cuts emissions further.

Cloud providers can help by surfacing real‑time energy dashboards so users see the carbon cost of every query. Transparent metrics nudge people toward greener defaults without sacrificing capability when it truly counts.

Future AI must limit long answers

Engineers are racing to squeeze more work from fewer tokens, trimming redundant reasoning steps and caching intermediate thoughts. Hardware makers also chase efficiency, but clean grids remain essential.

Data‑center growth is already prompting utilities to burn more fossil fuel, with a recent investigation warning of billions of dollars in public‑health impacts tied to AI power demand.

Policy teams are weighing disclosure rules, minimum efficiency standards, and incentives for renewable‑powered compute zones. Without them, the next wave of models could push the sector’s footprint beyond that of entire mid‑size nations.

Incremental gains in software design, smarter user choices, and cleaner electricity can bend the curve. The challenge is aligning all three before the growth of digital brains outpaces the planet’s patience.

The study is published in the journal Frontiers in Communication.

—–

Like what you read? Subscribe to our newsletter for engaging articles, exclusive content, and the latest updates. 

Check us out on EarthSnap, a free app brought to you by Eric Ralls and Earth.com.

—–