A recent study by German researchers, published in Frontiers in Communication and highlighted via SciTechDaily, revealed that AI models designed for detailed “step‑by‑step” reasoning—aptly dubbed reasoning‑enabled models—can emit up to 50 times more CO₂ per question than models giving concise answers, all without necessarily delivering better accuracy. Testing 14 large language models (7 to 72 billion parameters) against 1,000 standardized prompts, researchers noted reasoning models generated an average of 543.5 “thinking” tokens, compared to just 37.7 for concise models, substantially hiking emissions. Although the Cogito model (70B) was the most accurate at 84.9%, it emitted three times more CO₂ than similarly sized concise models, illustrating a clear accuracy‑sustainability trade‑off. The findings carry broader implications, pointing to an environmental footprint that could rival air travel depending on prompt complexity.
Sources: AP News, OECD.ai, SciTech Daily
Key Takeaways
– Efficiency Matters: AI systems that “think through” answers in detail emit dramatically more CO₂—sometimes up to 50×—without guaranteeing improved performance.
– Wider Environmental Context: The hidden carbon cost of AI, especially reasoning‑heavy models, contributes to growing concern over tech infrastructure’s climate impact.
– Path Forward Exists: Experts suggest ways to reduce AI’s footprint, such as prompt brevity and greener data‑center choices, indicating a pragmatic path toward sustainability.
In-Depth
AI’s rise may be smart—but not always green. German research showcased in SciTechDaily warns that “thinking” AI models—those that generate internal reasoning chains—can guzzle up to 50 times more carbon than leaner, direct‑answer models, despite often tying on accuracy.
By evaluating 14 models across 1,000 questions, the study found that reasoning models created roughly 543 extra tokens, compared to just 38 for straightforward models—a direct driver of increased emissions. Notably, the most accurate model, Cogito (70B parameters), reached 84.9% correctness but discharged triple the CO₂ of similarly sized concise models—highlighting a tangible trade‑off between precision and sustainability.
Beyond AI, reporting by the Associated Press underscores how data centers—especially those supporting AI—are quietly draining energy and water, often relying on fossil fuels and cooling systems that consume millions of gallons per day.
Yet, there’s promise: an OECD study shows that optimizing across what it calls the “4 Ms” (Model, Machine, Mechanization, Map) can slash emissions dramatically—even by hundreds‑fold over time if models are efficient, hardware is optimized, and data centers run on clean power
OECD AI
.
The environmental cost of AI can’t be ignored. But with mindful prompting, smarter model selection, and greener infrastructure, it’s possible to balance innovation with responsible stewardship—so we can use AI’s power wisely, without leaving behind too much carbon in its wake.