Ilaria Rosella Pagliaro Prompting AI creates emissions. A new study reveals that complex queries can pollute 50x more than simple ones, challenging our assumptions about accuracy and sustainability.
Every time we ask an AI something—even the simplest query—it begins to “think.” But that thinking, as invisible as it may seem, has a carbon footprint.
Behind every AI-generated answer lies a mechanical process powered not by magic, but by data tokens—units of text, like words or fragments, transformed into numeric strings that language models can interpret. This transformation, and the computations it triggers, require energy. And energy, of course, means CO₂ emissions.
Recently, a group of German researchers set out to measure just how much pollution these invisible processes create. The findings, published in Frontiers in Communication, may force us to rethink how we use AI: certain types of prompts can generate up to 50 times more carbon emissions than others.
Reasoning costs more—literally
The study, led by Maximilian Dauner from Hochschule München University of Applied Sciences, compared various large language models (LLMs)—from the fast-and-blunt types to those that unravel complex logic webs. It turns out, the latter—models that perform explicit reasoning steps—burn through significantly more energy than those that cut straight to the chase.
To test this, the researchers fed 1,000 questions to 14 different LLMs, ranging in size from 7 to 72 billion parameters. (For context, parameters are the building blocks that help AI “learn” from data.)
The results? Models that engaged in reasoning produced, on average, 543.5 “thinking tokens” per question. Models that skipped that step? Just 37.7 tokens. Those extra tokens—each representing a mental detour or intermediate step—translate into higher electricity use, and therefore more emissions.
And here’s the kicker: more tokens don’t always mean more accurate answers.
Accuracy vs sustainability: the trade-off
The most accurate model in the study, Cogito, came with 70 billion parameters and an impressive 84.9% accuracy rate. But its answers also generated triple the emissions compared to equally large models that kept things brief. As Dauner put it:
“We clearly see a trade-off between accuracy and sustainability.”
In practical terms, no model that kept its CO₂ emissions below 500 grams (1.1 lbs) per 1,000 questions managed to exceed an 80% accuracy rate. The researchers used CO₂ equivalent (CO₂e) to quantify the climate impact—this metric rolls all greenhouse gas emissions into a single standardized unit.
But the model alone isn’t to blame. The nature of the question matters, too. Asking an AI to solve abstract algebra or interpret a philosophical dilemma can generate six times the emissions of simpler queries, like high school history trivia.
How to use AI responsibly
So, should we ditch the smarter models? Not exactly. The researchers aren’t calling for AI abstinence—they’re advocating for smarter usage. In Dauner’s words:
“Significant reductions in emissions can be achieved simply by requiring concise responses or by reserving the most powerful models only for tasks that justify their use.
“
He’s got a point. Running 600,000 queries through DeepSeek R1 (70B parameters) would generate the same carbon emissions as a round-trip flight from London to New York. Yet Qwen 2.5 (72B parameters) can answer nearly 1.9 million questions with comparable accuracy, and the same climate cost.
This difference highlights how some models are more efficient, not just more powerful. But before we get carried away, it’s worth noting that factors like hardware type, electricity grid efficiency, and data center location also affect the outcome. So, don’t treat these numbers as gospel—they’re best seen as strong indicators, not absolute truths.
And if you’re thinking, “Well, it’s just a meme generator…”—think again. As Dauner warns:
“If users were fully aware of the environmental cost of AI-generated responses—even for trivial uses like transforming into a video game character—they would likely take a more selective and responsible approach to using these technologies.”
So maybe, next time you ask your chatbot a ridiculous question, consider: is this worth the emissions of a lightbulb running for hours?
Source: Frontiers in Communication
