ClearFactAll topics

Does large-scale AI model training significantly increase carbon emissions and environmental impact at a rate that outpaces its potential benefits?

5 peer-reviewed studies for, 5 against

Key takeaways

Stronger evidence: Against

Verdict

The proposition makes a comparative claim: that AI training's environmental costs grow faster than its benefits. The evidence shows real and rising costs, but it does not show that they outpace the benefits, and the stronger material points the other way.

The case for the proposition has substance. Strubell et al. estimated about 284 tonnes of CO2 for an extreme training run, and Green AI describes a roughly 300,000-fold rise in training compute between 2012 and 2018. The Joule study projects AI electricity use of 85 to 134 TWh per year by 2027 and warns that efficiency gains could increase total demand. The water-footprint work adds a dimension beyond carbon. Still, much of this evidence is dated, projective, or aimed at worst-case scenarios. The 284-tonne figure comes from an unusual architecture-search setup, and the Google researchers argue that such estimates overstated typical training emissions. The studies also document costs without weighing them against benefits. Power Hungry Processing, for example, concerns deployment more than training and calls for more careful weighing, not a conclusion that costs dominate.

The evidence against speaks more directly to the question of rate. The Science recalibration found that data center computing grew about sixfold from 2010 to 2018 while energy use rose only about 6%, which shows efficiency can offset demand growth. The Google authors report that four practices can cut training energy up to 100 times and emissions up to 1,000 times, and that ML was under 15% of Google's energy use. The cloud-intensity study shows emissions vary widely by region and timing, so the footprint can be managed. On benefits, the climate-ML survey identifies concrete mitigation applications, and the Scientific Reports study finds AI emits far less CO2 than humans for comparable writing and illustration tasks, although it excludes rebound effects.

Some caveats apply. Some of the against evidence comes from industry authors, and efficiency gains may not continue. Rebound effects remain uncertain. Even so, the evidence more strongly supports the view that AI's footprint is significant but manageable, and that it has not been shown to outpace the benefits.

Strongest study on each side

Ranked by citation count.

Strongest study for

Energy and Policy Considerations for Deep Learning in NLP (Proceedings of ACL, 2019)

6,321 citations

Quantified the energy use, carbon emissions and cloud cost of training recent language models. The most extreme case, training a large Transformer with neural architecture search, was estimated to emit about 284 tonnes of CO2, roughly five times the lifetime emissions of an average car including its fuel. The authors warn that accuracy gains increasingly depend on huge computational resources and recommend reporting training costs and prioritizing efficient methods.

Strongest study against

Tackling Climate Change with Machine Learning (ACM Computing Surveys, 2022)

2,313 citations

A survey by more than 20 researchers identifying high-impact ways machine learning can reduce greenhouse gas emissions and help societies adapt, including smart electricity grids, forecasting renewable power, more efficient buildings, transport and industry, monitoring land use and deforestation, climate modelling and disaster management. Argues machine learning can be a powerful tool against climate change when applied in collaboration with other fields.

Open the interactive page →

The evidence

For (5)

Energy and Policy Considerations for Deep Learning in NLP (Proceedings of ACL, 2019)

6,321 citations

Quantified the energy use, carbon emissions and cloud cost of training recent language models. The most extreme case, training a large Transformer with neural architecture search, was estimated to emit about 284 tonnes of CO2, roughly five times the lifetime emissions of an average car including its fuel. The authors warn that accuracy gains increasingly depend on huge computational resources and recommend reporting training costs and prioritizing efficient methods.

Green AI (Communications of the ACM, 2020)

2,935 citations

Notes that the computation used to train leading deep learning models increased about 300,000-fold between 2012 and 2018, and argues the field's focus on accuracy at any cost carries a rising carbon footprint and excludes researchers without large budgets. Proposes that efficiency be reported and treated as an evaluation criterion alongside accuracy.

The growing energy footprint of artificial intelligence (Joule, 2023)

1,007 citations

Based on projected shipments of AI servers, estimates that AI could consume 85 to 134 terawatt-hours of electricity per year by 2027, comparable to the annual electricity use of a country such as the Netherlands, Argentina or Sweden. Notes that after training, everyday use of AI models may account for most of their energy use, and cautions that efficiency gains could increase total demand rather than reduce it.

Making AI Less 'Thirsty': Uncovering and Addressing the Secret Water Footprint of AI Models (Communications of the ACM, 2025)

993 citations

Estimates that training GPT-3 in Microsoft's US data centers could have directly evaporated around 700,000 liters of clean freshwater for cooling, and that global AI demand could account for 4.2 to 6.6 billion cubic meters of water withdrawal in 2027. The authors argue AI's environmental impact goes beyond carbon, and call for water footprint to be reported and managed alongside emissions.

Power Hungry Processing: Watts Driving the Cost of AI Deployment? (ACM FAccT, 2024)

918 citations

The first systematic comparison of the energy and carbon cost of running different AI models. General-purpose generative models were orders of magnitude more energy-intensive than task-specific models doing the same job, even at similar model size. The authors caution that the growing use of multi-purpose generative AI should be weighed more carefully against its energy and emissions costs.

Against (5)

Tackling Climate Change with Machine Learning (ACM Computing Surveys, 2022)

2,313 citations

A survey by more than 20 researchers identifying high-impact ways machine learning can reduce greenhouse gas emissions and help societies adapt, including smart electricity grids, forecasting renewable power, more efficient buildings, transport and industry, monitoring land use and deforestation, climate modelling and disaster management. Argues machine learning can be a powerful tool against climate change when applied in collaboration with other fields.

Recalibrating global data center energy-use estimates (Science, 2020)

2,139 citations

Found that between 2010 and 2018 the amount of computing done in data centers grew about sixfold, yet their energy use rose only about 6%, to roughly 1% of global electricity, because of large efficiency gains in servers, storage and cooling. The authors argue earlier forecasts of runaway growth were overstated, while noting that continued efficiency policies are needed to keep up with demand.

The Carbon Footprint of Machine Learning Training Will Plateau, Then Shrink (IEEE Computer, 2022)

797 citations

Google researchers report that four practices (efficient model architectures, specialized processors, cloud data centers and locations with clean energy) can cut training energy by up to 100 times and emissions by up to 1,000 times. Machine learning accounted for under 15% of Google's total energy use from 2019 to 2021, and some widely cited earlier estimates substantially overstated training emissions. They predict total training emissions will decline by 2030 if the field adopts these practices.

Measuring the Carbon Intensity of AI in Cloud Instances (ACM FAccT, 2022)

501 citations

Measured the emissions of training a range of AI models in the cloud and found that the same training job's emissions vary greatly depending on the region and time of day it runs, because the carbon intensity of electricity grids differs. Choosing cleaner regions and shifting or pausing jobs to low-carbon hours can substantially cut emissions, showing AI's footprint is largely a manageable choice rather than fixed.

The debate

Priya: AI training has a serious environmental cost. Strubell and colleagues estimated that one large language model trained with architecture search could emit as much CO2 as five cars over their entire lifetimes.

Daniel: That number became famous, but Patterson and colleagues at Google showed the estimate was substantially overstated. With efficient models, specialized chips, cloud data centers and clean energy, training emissions can be cut by up to a thousand times.

Priya: Efficiency isn't keeping up with scale. Schwartz's Green AI paper notes training compute grew about 300,000-fold in six years. De Vries projects AI could use as much electricity as the Netherlands by 2027.

Daniel: We've heard runaway forecasts before. Masanet's team found in Science that data center computing grew sixfold from 2010 to 2018 while energy use rose only 6 percent. It's still about 1 percent of global electricity.

Priya: That was before generative AI. Luccioni, Jernite and Strubell found general-purpose generative models use orders of magnitude more energy than task-specific ones. And it's not just carbon: Li and colleagues estimate training GPT-3 evaporated around 700,000 liters of freshwater.

Daniel: But the footprint depends on choices. Dodge and colleagues showed the same training job's emissions vary hugely by region and time of day. Run it on a clean grid at the right hour and emissions drop sharply.

Priya: Choices that companies don't always make. And what benefits outweigh all this?

Daniel: Rolnick and over 20 researchers mapped how machine learning can cut emissions, from smarter grids to forecasting renewables to tracking deforestation. And Tomlinson's study found AI produces a page of text with 130 to 1,500 times less CO2 than a human writer.

Priya: Those benefits are potential. The emissions are happening now.

Daniel: Fair. The footprint is real and manageable, but only with clean energy, efficient models and honest reporting.

Frequently asked questions

How much carbon does training an AI model produce?

It varies widely. One 2019 study estimated about 284 tonnes of CO2 for an extreme case involving neural architecture search, but Google researchers argue such estimates overstated typical training emissions.

How much electricity could AI use in the future?

One projection based on expected AI server shipments estimates 85 to 134 terawatt-hours per year by 2027, comparable to the annual electricity use of a country such as the Netherlands, Argentina or Sweden. The same study cautions that efficiency gains could increase total demand rather than reduce it.

Can AI's carbon footprint be reduced?

Studies indicate it can. Running training jobs in cleaner regions or at low-carbon times of day can substantially cut emissions, and Google researchers report that efficient architectures, specialized processors, cloud data centers and clean energy can cut training energy by up to 100 times.

Does AI also have environmental benefits?

A survey of more than 20 researchers identifies uses of machine learning for smart grids, renewable power forecasting, efficient buildings and transport, and climate modelling. Another study found AI emitted far less CO2 than humans for comparable writing and illustration tasks, though it excluded rebound and social effects.

Related questions