AI emissions vary with prompt and model of LLM

A new study by MIT reveals that the CO2 emissions from different AI models vary widely depending on model size, task complexity and the phrasing of users' prompts. Researchers compared 14 large language models (LLMs) and found that reasoning-intensive responses can produce up to 50 times more emissions per query than concise answers. There was a direct correlation between answer precision and emissions, showing the direct trade-off between sustainability and output quality. Subject matter impacts AI emissions too – abstract topics generate more CO2 than simpler ones. Users can reduce their carbon footprint by using lower-capacity models and requesting concise answers. Model and hardware choice also affect results significantly. (Frontiers) 

Why does this matter? AI is reshaping the physical as well as social world as we know it. The adoption of AI is driving massive energy demand, primarily  due power-hungry data centres. Projects such as the $500bn Stargate Initiative in the US – which exceeds the spend on the Apollo space programme – aim to build dozens of new data centres. Meanwhile, leading tech companies including  Meta and Microsoft are turning to nuclear power to power them. This is not typical tech evolution, AI is a seismic shift in global energy and infrastructure. 

For Rio Tinto, AI energy use has repercussions. The company mines critical minerals – such as copper and lithium – needed for AI infrastructure and renewable systems. AI advancement also offers efficiency gains across mining operations. However, without ethical guardrails, AI use risks fuelling unsustainable extraction and carbon emissions. Aligning AI with environmental and social responsibility is essential for Rio Tinto to lead in a just  energy transition. 

MIT's study evaluated LLMs with sizes ranging from seven to 72 billion parameters across 1,000 benchmark questions. Reasoning-enabled models, such as Cogito (70 billion parameters), averaged 543.5 "thinking" tokens per question and reached 84.9% accuracy. The model generated three times the CO2 emissions of similarly sized concise models. In contrast, concise models used just 37.7 tokens per question. Tasks involving abstract subjects such as philosophy emitted up to six times more CO2 than simpler topics such as high school history. For perspective, answering 600,000 queries with DeepSeek R1, also with 70 billion parameters, produces emissions equal to a round-trip flight from London to New York. 

AI companies are secretive about the emissions of their models, hence the significance of MIT’s study. Training models such as GPT-4 consumes tens of gigawatt-hours, but inference – responding to user queries – now represents up to 90% of AI's energy use. A single query may require little energy, but with billions occurring daily, the cumulative impact is vast. By 2028, AI could consume as much electricity as 22% of US households. With limited transparency, utilities risk offloading costs to consumers. 

Conversely, AI also offers promise for reducing global emissions in a systemic way. Another recent study by the Grantham Research Institute and Systemiq estimates that AI could help reduce global greenhouse gas emissions by 3.2 to 5.4 billion mt of CO2e annually by 2035. Published in npj Climate Action, the research highlights AI’s potential in five key areas –system transformation, resource efficiency, behaviour change, climate modelling and resilience. In the power sector, AI could boost solar and wind load factors by up to 20%. 

Research from both MIT and in the npj report stress that public investment and strong governance are essential to ensure equitable, sustainable deployment of the AI technology transforming the future.