Should Researchers Consider Environmental and Resource Costs When Using AI at Scale?

When Does the Environmental Cost of AI Become a Research Decision?

A research team plans to process several million documents using a large language model. Another team intends to run hundreds of experiments comparing AI systems. Both have sufficient funding and access to the required computing infrastructure.

But should the availability of computational resources be enough to justify their use?

AI systems depend on physical infrastructure that consumes electricity, requires cooling, and incorporates hardware manufactured from material resources. When AI processing is repeated at scale, these demands can become substantial.

The question is not whether researchers should abandon computationally intensive methods. It is whether environmental and resource costs should influence methodological decisions when scientifically adequate alternatives may be available.

Yes, Environmental Costs Should Be Considered When AI Is Used at Scale

Understanding the Environmental and Resource Demands of Large-Scale AI Research

Why AI research has an environmental footprint

AI models are implemented through computing systems that require physical equipment and energy. Training, fine-tuning, and inference can involve processors, accelerators, memory, storage, and supporting data-center infrastructure.

Inference refers to using a trained model to produce predictions or outputs. Training refers to the process through which model parameters are learned or adjusted. Both can consume substantial resources, although their relative contributions depend on the system and workload.

The International Energy Agency's 2025 Energy and AI report estimated that data centers consumed approximately 415 terawatt-hours of electricity in 2024, representing around 1.5% of global electricity consumption. Its base-case projection placed data-center electricity consumption at approximately 945 terawatt-hours in 2030.

These estimates concern data centers as a whole, not AI research alone. Data centers also support cloud computing, storage, communications, and many other digital services. Consequently, their total electricity consumption should not be presented as the environmental footprint of AI research.

The figures nevertheless demonstrate why computational infrastructure has become relevant to broader energy planning.

What environmental impacts should researchers consider?

Electricity consumption is an important indicator, but it does not capture the complete environmental footprint of AI.

The OECD's analysis of AI's environmental impacts distinguishes direct impacts associated with computing infrastructure throughout its lifecycle from indirect impacts arising from AI applications. This distinction is useful because an AI system can create environmental burdens through its operation while potentially contributing to environmental benefits elsewhere.

Not every study can measure all these dimensions. Researchers should distinguish impacts they have measured, those they have estimated, and those they cannot quantify reliably.

Environmental reporting becomes misleading when a single convenient indicator is presented as a complete assessment.

Why AI processing scale changes the decision

A small number of AI queries may contribute little to a project's overall resource consumption. Large-scale workflows can involve millions of requests, extensive model comparisons, repeated processing, or computationally demanding experiments.

Even when individual operations are relatively efficient, their aggregate consumption may become substantial as the workload expands.

However, the environmental impact of AI cannot be inferred simply by multiplying an assumed universal cost per query by the number of requests.

Different models, hardware systems, processing configurations, and workloads can have different energy requirements. A short classification request is not necessarily comparable to extended reasoning, multimodal generation, or large-context document processing.

Consequently, the environmental impact of individual AI queries should be understood as workload-dependent rather than fixed.

Does using a larger model always consume more energy?

Not necessarily for every completed research task.

Larger models often require more memory and computational resources per operation under comparable conditions. However, actual energy consumption depends on model architecture, numerical precision, hardware efficiency, batching, software optimization, and the number of operations required.

A smaller model might consume less energy per request but require additional processing or repeated attempts to achieve acceptable results. Conversely, a larger model may use substantially more resources without providing a meaningful improvement for a narrow task.

Researchers should therefore compare resources consumed per scientifically acceptable result, not merely model size or theoretical computational demand.

This is particularly relevant when considering whether smaller AI models are scientifically sufficient for the intended task.

How can researchers estimate electricity consumption?

When reliable power measurements are available, electrical energy can be calculated from power consumption over time.

This calculation does not automatically include cooling, networking, storage, or other infrastructure outside the measurement boundary.

Where only accelerator power is measured, the result should be described as accelerator energy consumption rather than the complete energy footprint of the computing facility.

Actual measurements are preferable when feasible. Estimates based on nominal hardware specifications should be identified as estimates because rated power and actual operating power can differ.

How are electricity use and carbon emissions related?

Electricity consumption and greenhouse-gas emissions are related but distinct quantities.

Emissions associated with electricity use depend partly on how that electricity is generated and the accounting method applied. The same amount of electricity may have different associated emissions across locations and time periods.

Researchers should avoid assuming that a universal emissions factor applies to every AI workload.

Where appropriate electricity emissions factors are available, an illustrative operational estimate can be calculated.

The assumed factor is illustrative and does not represent a verified electricity grid or provider. The resulting estimate excludes emissions from hardware manufacturing and other lifecycle stages unless these are separately incorporated.

Researchers should identify the source, year, geographic relevance, and accounting basis of any emissions factor used in an actual study.

Why water consumption requires particular caution

Water may be used directly in some data-center cooling systems and indirectly in electricity generation or hardware manufacturing.

However, water withdrawal and water consumption are different measures. Withdrawn water may be returned to its source, while consumed water is not immediately returned to the same local water system.

Water-related impacts also depend on location, cooling technology, climate, and local water availability.

Consequently, applying a single water-per-query estimate to all AI services can create a misleading impression of precision.

A data center's water consumption in a water-stressed region may raise different concerns from an equivalent volume consumed where water availability is less constrained.

Researchers should therefore avoid presenting generic water-footprint estimates as measurements of their own workload unless the underlying assumptions and system boundaries are defensible.

What is the difference between operational and lifecycle environmental impacts?

Operational impacts arise while the computing system is being used. Lifecycle impacts extend to the manufacture, transportation, maintenance, and disposal of the equipment and infrastructure.

Both perspectives can be useful, but they answer different questions.

A study that reports only measured electricity consumption during model inference should not describe that figure as the complete environmental footprint of the AI system.

The OECD recommends looking beyond operational electricity and emissions when assessing AI's environmental impacts.

Should researchers account for environmental benefits of AI?

Yes, when those benefits are relevant and supported by evidence.

AI may support environmental research, improve the analysis of energy systems, or assist with identifying opportunities for resource efficiency. Such applications may generate benefits that justify substantial computational expenditure.

However, anticipated benefits should not be treated as proven offsets against the resources consumed by the research.

For example, developing an AI model intended to improve energy efficiency does not establish that its deployment will produce net emissions reductions. That claim requires an appropriate comparison against a credible alternative and consideration of relevant indirect effects.

Researchers should distinguish potential benefits, measured benefits, and net environmental effects.

How can research teams reduce unnecessary AI processing?

Environmental considerations can be incorporated into methodological planning without compromising scientific standards.

Possible measures include using representative pilot datasets before full-scale processing, avoiding redundant experiments, reusing validated intermediate outputs when methodologically appropriate, and selecting computational methods based on demonstrated task requirements.

Efficient batching, suitable hardware, and appropriate model configurations may also reduce resource consumption, although the effects depend on the system.

Importantly, computational efficiency should not become an excuse for inadequate validation. Repeated experiments may be necessary to assess reliability, quantify uncertainty, or investigate consequential errors.

The goal is to remove unnecessary consumption, not scientifically necessary evidence.

Researchers comparing methods should consider the relationship between AI performance and computational cost rather than assuming that the most resource-intensive model is the most defensible choice.

What can researchers know about cloud-based AI environmental impacts?

Cloud services can make computational infrastructure accessible without requiring researchers to purchase or operate their own hardware.

However, commercial providers may not disclose all information needed to estimate workload-specific energy use, cooling demand, hardware allocation, or emissions.

Provider sustainability reports can offer useful organizational information, but company-wide figures generally cannot establish the environmental footprint of an individual research workload.

Where direct measurements are unavailable, researchers should document the accessible workload information, such as model identity, usage volume, processing duration, and relevant deployment details.

They should then explain which environmental impacts cannot be estimated reliably.

Transparent uncertainty is preferable to an apparently precise calculation built on unsupported assumptions.

Why environmental responsibility is not the same as minimizing computation at all costs

Scientific research can produce benefits that justify resource consumption. Computationally demanding methods may be necessary to investigate complex problems, evaluate safety, or develop reliable scientific tools.

Nevertheless, scientific value does not automatically justify unlimited computational expenditure.

The relevant judgment concerns proportionality: whether the resources consumed are reasonable in relation to the research objectives, expected benefits, available alternatives, and methodological requirements.

The UNESCO Recommendation on the Ethics of Artificial Intelligence recognizes environmental and ecosystem considerations as part of responsible AI governance. The OECD similarly emphasizes the need for improved measurement and transparency rather than relying on simplistic claims about AI's environmental effects.

Environmental assessment should therefore inform scientific judgment, not replace it.

Comparing the Resource Requirements of Two AI Research Workflows

The comparison does not establish that Model A has one-quarter of Model B's total lifecycle environmental footprint. It measures only the electricity consumption represented by the specified system boundary.

It also assumes that pilot measurements can be reasonably extrapolated to the complete workload. In actual research, that assumption should be checked because throughput, batching, and system utilization may change at scale.

If Model B provided a necessary improvement in accuracy or safety, the resource difference would need to be weighed against that benefit rather than treated as an automatic reason for rejection.

Common Misconceptions About AI's Environmental Costs

How to Incorporate Environmental Considerations Into AI Research Planning

Researchers should begin by asking whether the intended AI method is necessary for the scientific objective and whether alternatives could provide adequate performance.

When AI is justified, resource efficiency can become one criterion for choosing among acceptable methods. The importance of this criterion should increase when the workload is extensive, the environmental burden is potentially substantial, or comparable methods differ markedly in resource consumption.

For large projects, researchers should also consider whether institutional computing policies, funder expectations, or disciplinary reporting practices address sustainability and resource use.

Maintaining a resource log can support both budget management and environmental assessment. The process of reporting computational costs in AI-assisted research provides a useful starting point, even when direct environmental measurements are not available.

The broader objective is responsible methodological choice. Environmental considerations should help researchers identify unnecessary resource use while preserving the quality of evidence needed to answer the research question.

Before Launching a Large-Scale AI Research Workflow

Questions About Environmental Responsibility in AI Research

Environmental Responsibility Should Inform AI Research Design

How to Cite This Guide

This guide is intended to be read, shared, and used in research, teaching, and academic work. If you draw on its ideas, explanations, or other content, please acknowledge the source by citing the guide. Doing so gives appropriate credit and helps your readers locate the original resource.