Source-linked AI summary
Unraveling the Hidden Environmental Impacts of AI Solutions for Environment
Anne-Laure Ligozat, Julien Lefèvre, Aurélie Bugeau, Jacques Combaz
TL;DR
The paper addresses the gap in assessing the complete environmental impacts of AI for Green, beyond energy use and GHG emissions. It reviews impact-assessment approaches, presents life cycle assessment for AI services, and discusses evaluating environmental usefulness while identifying limitations in existing work.
Problem
Existing AI for Green work does not quantify all environmental costs, while AI methods’ energy use and associated GHG emissions represent only part of their impacts.
Method
The paper reviews AI’s environmental impacts and assessment methodologies, applies life cycle assessment to evaluate direct impacts, and proposes comparing a reference solution with an AI solution.
Results
The proposed methodology can comprehensively evaluate direct environmental impacts through life cycle assessment, but it shows only a service’s technical potential.
Takeaways & Limitations
Assessing AI for Green requires considering multiple direct environmental impacts and comparing AI services with reference solutions rather than focusing only on GHG emissions.
Takeaways & Limitations
Existing AI literature often addresses only a small part of direct impacts and neglects production and end of life; large-scale deployment may also require consequential rather than attributional LCA.
Abstract
from arXiv · showhide
In the past ten years, artificial intelligence has encountered such dramatic progress that it is now seen as a tool of choice to solve environmental issues and in the first place greenhouse gas emissions (GHG). At the same time the deep learning community began to realize that training models with more and more parameters requires a lot of energy and as a consequence GHG emissions. To our knowledge, questioning the complete net environmental impacts of AI solutions for the environment (AI for Green), and not only GHG, has never been addressed directly. In this article, we propose to study the possible negative impacts of AI for Green. First, we review the different types of AI impacts, then we present the different methodologies used to assess those impacts, and show how to apply life cycle assessment to AI services. Finally, we discuss how to assess the environmental usefulness of a general AI service, and point out the limitations of existing work in AI for Green.
1 Introduction
AI is increasingly presented as a solution to environmental problems, while its own impacts extend beyond energy use and GHG emissions. The paper reviews these impacts and proposes ways to assess them comprehensively.
- Environmental impacts: Energy consumption and associated GHG emissions capture only part of AI methods’ complete environmental impacts, including indirect effects on the global digital sector.The paper emphasizes that wider impacts should be considered alongside commonly reported energy and carbon measures.
- Environmental impacts: Deep learning methods require large quantities of data whose acquisition, transfer, storage, and processing consume equipment and energy.The number of required devices can vary substantially across applications such as surveillance satellites and smart buildings.
- Environmental impacts: Training deep neural models requires substantial computation time and resources because the model learns a comprehensive representation of the data.Continuous learning can increase computation cost further.
- Motivation: AI for Green applications may create negative environmental impacts, including rebound effects that can increase global GHG emissions despite unitary efficiency gains.Assessing actual impacts therefore requires considering both positive and negative effects.
- Contributions: The paper reviews environmental-impact assessment work, presents life cycle assessment for direct impacts, and discusses evaluating the environmental value of AI services.It also argues that the methodology shows technical potential that may not fully materialize in real-life contexts.
2 Related work
Existing AI impact assessments often focus on energy or carbon during use, whereas life cycle assessment can cover multiple impacts and life cycle stages. The paper reviews available tools and highlights gaps in evaluating AI for Green trade-offs.
- Assessment tools: Integrated tools measure energy consumption and associated carbon footprint, while online tools estimate impacts from parameters such as duration, material, and location but are less accurate.Examples include Experiment Impact Tracker, Carbon Tracker, CodeCarbon, Green Algorithms, and ML CO2 Impact.
- Carbon footprint of AI: AI literature mostly addresses a small part of direct impacts and neglects production and end of life.A cited life cycle study estimated manufacturing at about 75% of total emissions for Apple or an iPhone 5.
- Assessment tools: Existing work has revealed substantial variation in GHG impact: considering data-center location and network sparsity reduced estimated impact by a factor of 100 in one study.The cited study evaluates operating computers and data centers, rather than all environmental impacts.
- AI for Green evaluation: Meaningful AI for Green cost-benefit assessments should consider complete AI impacts, while many proposed environmental solutions lack rigorous evaluation of their cost-benefit balance.A cited framework addresses environmental and societal trade-offs of AI foundation models.
- Life cycle assessment: Life cycle assessment quantifies multiple environmental criteria across production, use, and end-of-life stages, using a functional unit and life cycle inventory.Its system perspective is intended to reduce problem shifting between environmental impacts and life cycle phases.
- Life cycle assessment: LCA has well-known data and complexity limitations, including difficulty assigning reliable values to life cycle inventory flows such as GPU manufacturing impacts.The paper notes that LCA has rarely been applied to AI services.
3 Life cycle assessment of an AI solution
The paper adapts life cycle assessment to AI services by defining system boundaries across software tasks, equipment, life cycle stages, and environmental criteria. It emphasizes that meaningful assessment must include all devices and relevant impacts, not only operational energy.
- Scope and dimensions: AI-impact quantification requires a defined system and commonly focuses too narrowly on the equipment use phase.The proposed framework adapts LCA to deep-learning code used alone or within a larger application.
- First-order impacts: First-order impacts cover raw-material extraction, manufacturing, transport, use, and end of life; the paper merges the first three into production.These impacts concern the equipment involved in the AI service.
- Scope and dimensions: LCA evaluates life-cycle phases across multiple environmental criteria, including GHG emissions, water footprint, human toxicity, and abiotic resource depletion.The selected criteria should reflect the study goal and the environmental issue targeted by the AI solution.
- AI-service inventory: The AI service includes all equipment supporting tasks from data acquisition through inference, while the application phase specifically denotes inference.Figure 2 links software tasks to devices and then to each device’s hardware life cycle.
- System boundaries: System boundaries span terminals, networks, and data-center or server infrastructure, with production, use, and end-of-life stages assessed for relevant unit processes.Dedicated equipment receives full-stage treatment, while shared equipment requires allocated production and use impacts.
- Allocation: Shared infrastructure impacts are allocated using operational measures such as execution time, dynamic energy consumption, or a fraction of static consumption.For simultaneously running programs, the paper gives 1/n as an example allocation for static consumption.
4 Assessing the usefulness of an AI for Green service
The paper evaluates AI for Green services by comparing reference and AI-enhanced applications through life cycle assessment, while distinguishing direct, application-level, and societal impacts. Its literature review finds that environmental evaluations are often absent or incomplete, and argues that technical benefits may not fully materialize in practice.
- 4 Assessing the usefulness of an AI for Green service: The proposed framework compares a reference application without AI against an AI-enhanced application using life cycle assessment across relevant environmental impacts.The comparison includes the AI service and associated equipment rather than limiting assessment to model operation.
- 4.1 Theoretical aspects: The net environmental value should account for positive gains alongside negative impacts from energy use, chips, waste, biodiversity risks, and other effects.The paper argues that the relevant environmental criteria should match the problem the AI service is intended to address.
- 4.1 Theoretical aspects: First-order impacts arise from equipment life cycles, second-order impacts from the application, and third-order impacts from technology or societal changes.The methodology focuses on first- and second-order impacts; third-order effects are beyond its scope.
- 4 Assessing the usefulness of an AI for Green service: The methodology can show a service’s technical environmental potential, but that potential may not be fully realized in real-life contexts.Third-order effects, including rebound effects, are among the broader consequences discussed outside the methodology’s main scope.
- 4.1 Theoretical aspects: The paper prefers the LCA-difference formulation because benefit-cost aggregation can mix diverse impacts and lacks a clear practical computation method.It notes that reducing multiple environmental impacts to one score involves arbitrary value choices and may dilute the criterion of interest.
- 4.2 Case studies: In the reviewed AI for Green literature, about half of the papers included no environmental evaluation, while many others used unquantified descriptions or distant proxies.A few citations evaluated environmental gain, mostly energy gain, but none considered the AI service’s own impacts.
5 Discussion
The paper finds that AI for Green evaluations capture only a small share of direct environmental impacts and proposes more complete assessment through life cycle assessment. It also cautions that LCA-based benefits assume unchanged surrounding applications and may miss broader societal effects, especially at large scale.
- AI for Green papers account for only a small part of direct environmental impacts.
- Attention to AI’s GHG emissions has focused on electricity consumption, while material flows receive less attention.
- Life cycle assessment evaluates global warming potential and other direct impacts across production, use, and end of life.Reliable life-cycle inventory data remain difficult to obtain, including manufacturing impacts for GPUs.
- The proposed LCA methodology assumes AI enhances or replaces existing applications while other things remain equal.
- Socio-technical concerns and third-order effects require evaluation because AI deployment can transform systems beyond technical efficiency gains.Autonomous vehicles illustrate how potential environmental benefits may coexist with non-ecological mobility transformations.
- Large-scale deployment may reorganize society and increase demand for materials or energy in nonlinear ways.
Authors contribution
The authors collectively contributed to conceptualization, methodology, validation, formal analysis, investigation, writing, and supervision, with specified authors leading data curation, visualization, and project administration.
- All authors contributed to conceptualization, methodology, validation, formal analysis, and manuscript writing.
- J.L. and A.-L. L. handled investigation and data curation, while A.B. and A.-L. L. handled visualization.
- A.-L. L. provided supervision and project administration.
Abbreviations
The manuscript defines abbreviations for artificial intelligence, deep learning, environmental assessment, computing hardware, and related technical terms.
- AI means Artificial Intelligence, while DL means Deep Learning and ML means Machine Learning.
- GHG means Greenhouse Gas, and LCA means Life Cycle Assessment or Analysis.
- GPU, CPU, and TPU denote Graphics Processing Unit, Central Processing Unit, and Tensor Processing Unit.
- ICT means Information and Communications Technology, while NLP means Natural Language Processing and CNN means Convolutional Neural Network.