2025科学领域人工智能(AI)的环境影响研究报告 Considerations on the environmental impact of AI in science_第1页
2025科学领域人工智能(AI)的环境影响研究报告 Considerations on the environmental impact of AI in science_第2页
2025科学领域人工智能(AI)的环境影响研究报告 Considerations on the environmental impact of AI in science_第3页
2025科学领域人工智能(AI)的环境影响研究报告 Considerations on the environmental impact of AI in science_第4页
2025科学领域人工智能(AI)的环境影响研究报告 Considerations on the environmental impact of AI in science_第5页
已阅读5页,还剩55页未读 继续免费阅读

下载本文档

版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领

文档简介

Considerationsonthe

environmentalimpact

ofAIinscience

1

©InternationalScienceCouncil,2025

Tocitethisreport:

InternationalScienceCouncil(2025).ConsiderationsontheenvironmentalimpactofAIinscience.DOI:10.24948/2025.10

Authors:DenisseAlbornozandNataliaNorori

Reviewers:MarcelDorschandGongKe

Projectcoordination:DureenSamandarEweis,VanessaMcBride

Projectchair:DavidCastle

Fundingacknowledgement:Thisworkwascarriedoutwiththeaidofagrantfromthe

InternationalDevelopmentResearchCentre(IDRC),Ottawa,Canada.TheviewsexpressedhereindonotnecessarilyrepresentthoseofIDRCoritsBoardofGovernors.

Design:

MrClinton

Coverphoto:

FrankRamspott

AbouttheInternationalScienceCouncil

TheISCisaninternationalnon-governmentalorganizationwithauniqueglobalmembershipthatbringstogether250internationalscientificunionsandassociations,nationaland

regionalscientificorganizationsincludingscienceacademies,researchcouncils,regionalscientificorganizations,internationalfederationsandsocieties,andacademiesofyoungscientistsandassociations.

TheISCworksatthegloballeveltocatalysechangebyconveningscientificexpertise,adviceandinfluenceonissuesofmajorimportancetobothscienceandsociety.

Contents

KEYTAKEAWAYS

4

ABOUTTHISPAPER

5

INTRODUCTION

6

SECTION1:FOUNDATIONALCONCEPTSANDFRAMEWORKS8

1.1GreenAI 8

1.2SustainableAI 9

1.3EthicalandresponsibleAI 10

SECTION2:ENVIRONMENTALIMPACTOFAIACROSSTHELIFECYCLE11

2.1EnvironmentalimpactsofAI 11

2.2Alifecycleapproachtoassessingenvironmentalimpact 12

-

Softwarelayer12

-

Hardwarelayer13

2.3EstimatingdirectenvironmentalcostsacrosstheAIlifecycle13

-

Estimatingoperationalcosts13

-

Estimatingembodiedcosts14

2.4Transparencyandreporting14

SECTION3:STRATEGIESFORREDUCINGDIRECTENVIRONMENTAL16

COSTSOFAIAPPLICATIONSINSCIENTIFICRESEARCH

3.1Reducingenvironmentalcostsinthesoftwarelifecycle 16

-

Datamanagement16

-

Algorithmicoptimization17

-

CASESTUDY1:InkubaLM:Asmalllanguagemodelforlow-resourceAfrican18

languages

3.2Reducingenvironmentalcostsinthehardwarelifecycle18

-

Energy-efficienthardwareandcomputing18

-

CASESTUDY2:TinyMLenablesdataanalysisandresearchinlow-costdevices19

3.3Energy-efficientdatacentres19

-

CASESTUDY3:GreenAIinfrastructuresacrossregions20

3.4Decommissioningandend-of-lifemanagement21

CONCLUSION

22

REFERENCES

23

APPENDIX1:GLOSSARY28

3

Keytakeaways

•Thereislimitedawarenessandevidenceabouttheenvironmentalcostsofusing

artificialintelligence(AI)inscientificresearch.Thisarticleoffersframeworksand

toolsscientistsandresearchinstitutionscanconsidertoassesstheenvironmental

impactsoftheirresearchaspartofamoresustainable,ethicalandresponsibleuseofAIinscience.

•AddressingtheenvironmentalimpactofAIrequiresamulti-dimensionalapproach.ScientistsandresearcherswhoareplanningtoincorporateAIintotheirworkflowsneedtoassesstoolsinlightoftheirscientificvalue,socialequityandenvironmentalcostsacrosstheentireAIlifecycle,withattentiontoreboundeffectsandlong-termconsequences.

•Adoptingmoreresource-efficientAImodelshasenvironmentalandsocialbenefits.Smaller,localandfrugalapproachestoAIcanimproveaccessibility,affordability,

transparencyandsocialinclusionaroundtheuseofAI,especiallyindiverseandresource-constrainedresearchcontexts.

Keytakeaways

4

Aboutthispaper

Thispaperexaminestheenvironmentalimplicationsofapplyingartificialintelligence(AI)

inscientificresearch.Itservesasaprimerforscientists,researchinstitutionsandscience

policy-makerswhoseektounderstandvariousapproachestoaddressingtheenvironmentalimpactofAIinscience.Inaddition,itoffersguidanceonhowreducingenvironmental

costscancontributetothebroadergoalsofsustainabilityandethicalAIuseinresearchenvironments.

AlthoughevidenceonthespecificenvironmentalimpactsofAIinscientificresearchisstillemerging,thepaperprovidesconceptualframeworksandpracticaltoolstohelpassesstheenvironmentalimplicationsofthefullAIlifecyclewithinscientificprojects.Thefirstsectionintroduceskeyframeworksforunderstandingenvironmentalimpactsinaholisticway.

Thesecondsectionoutlinesanapproachfordefiningandmeasuringenvironmentalcosts

acrosstheAIlifecycle.Thethirdsectionpresentsconcretestrategiesforreducingthedirectenvironmentalfootprintofscientificprojectsthatuseordependonresource-intensiveAI

applications.

ThispaperreflectsonhowenvironmentalconsiderationsintersectwiththebroaderthemesofAIsustainabilityandethics,thoughanin-depthanalysisoftheseframeworksfallsoutsideitsscope.Similarly,whiletheroleofAIinadvancingtheSustainableDevelopmentGoalsisacknowledged,itisnotexploredindetail.SocialandeconomicimpactsofAIarediscussedonlyinsofarastheyrelatetoenvironmentalcosts.

ThepaperispartofaseriesofthreeprimersthatexplorevarioustechnicaldimensionsofAIanditsimpactonscience.Theotherprimersare“TypesofAIinscience”and“DataforAIinscience”.

Aboutthispaper

5

Introduction

Artificialintelligence(AI)isbecomingatransformativetoolacrossscientificfields(LeCun,

2025).Asystematicreviewofover70millionpapersfoundthat,since2015,theinfluenceanduseofAIhasspreadtonearlyeveryareaofnaturalsciences,impactingbiology,chemistry,

geologyandphysics(GaoandWang,2024).NotablebreakthroughsincludehighlyaccurateproteinstructurepredictionswithAlphaFold(Jumperetal.,2021);usingmachinelearningtogeneratethefirst-everphotoofasupermassiveblackhole(Brodericketal.,2022);and

improvedforecastingmodelsforextremeweatherevents(Bodnaretal.,2024).AdvancesinAIareimpactingsocialscienceresearchaswell,transformingmethodsinfieldssuchaslinguistics,politicalscienceandhistory(PennToday,2023).

Muchofthisprogresshasbeendrivenbyadvancesinmachinelearninganddeeplearning

modelsthatcanmoreaccuratelyidentifypatternsandmakepredictionsfromvast

unstructureddata(Choudharyetal.,2022).Inthecaseofsocialscience,generativeAI,

largelanguagemodelsandmultimodalAItools–allofwhicharepoweredbydeeplearning

architectures–arealsobecomingpowerfultoolsfordataanalysis,simulationandhypothesisgeneration(Grossmanetal.,2023).Thedemandforlargedatasets,specializedhardware

andhigh-performancecomputingresourcesrequiredforcomplexAIsystemshasalsoraisedquestionsforresearchinstitutionsaroundtheaccessibility,affordabilityandenvironmentalimpactofincorporatingAIacrossscientificdomains(LeCun,2025).

OneofthekeychallengesistheenvironmentalcostofAIapplications(Luccionietal.,2024b).AgrowingbodyofevidencehasdemonstratedthattrainingcomplexAImodels,suchasdeeplearningorgenerativeAIapplications,involvessubstantialresourceuseandgreenhousegasemissions(OrganisationforEconomicCo-operationandDevelopment,2022).Onestudy

suggeststhattrainingonelargelanguagemodelcangenerateapproximately300,000kgofcarbondioxideemissions,equivalentto125round-tripflightsbetweenNewYorkandBeijing(Dhar,2020).

ThecomputationaldemandsoftrainingAIaredoublingapproximatelyevery100days,withprojectionssuggestingthatby2028,AIcouldconsumemoreelectricitythanIcelanddid

in2021(WorldEconomicForum2024).Reproducibilityalsoremainsachallenge,asmany

deeplearningmodelsareopaqueanddifficulttoreplicate.Thiscreatessituationsinwhich,duetothelackofreportingandtransparency,therearemanyunknownssurroundingthe

trueenvironmentalcostofAItoolsusedforscientificresearch,whichcanalsoleadto

unnecessaryenergyusewhenmodelscannotbereusedorrequirerepeatedexperimentstoreproduceresults(Lannelongueetal.,2023).

Introduction

ThecostsofAIextendbeyondthetrainingphase.Althoughoftenperceivedasanintangibledigitalsystem,AIdependsonextensivephysicalinfrastructurethatconsumesenergyandnaturalresources(OrganizationforEconomicCo-operationandDevelopment,2022).Theproductionofspecializedhardware,suchasgraphicsprocessingunits,whichareessentialfortrainingandrunningAImodels,requirestheextractionandprocessingofrareearth

mineralsandmetals.Thisprocesscontributestoenvironmentaldegradationthrough

6

increasedgreenhousegasemissions,addedstrainonoverburdenedwatersystems,and

serioushumancosts,withworkersinvolvedinthesesupplychainsfacingunsafeconditions,exposuretotoxicchemicalsandviolationsoftheirhumanrights(Nayar,2021).

Similarly,theenergydemandofdatacentres,whichpowerAIsystemsbystoringand

processingvastvolumesofdata,isdoublingapproximatelyeveryfouryears(Thangametal.,2024).Thesefacilitiesconsumesubstantialamountsofenergy,whethersourcedfromfossilfuelsorrenewablesources,andrequirelargequantitiesofwaterforcoolingandmaintainingoptimalhardwareperformance.Thisplacesfurtherpressureonlocalenergygridsand

watersupplies,particularlyinregionsfacingresourceconstraints(Mytton,2021).AsAIusecontinuestogrow,theincreasingdemandfordatacentreinfrastructureislikelytoplaceanevengreaterburdenoncountriesintheGlobalSouth,wherechallengingclimateconditions,waterscarcity,limitedconnectivityandfrequentpoweroutagesareongoingconcerns

(UnitedNationsConferenceonTradeandDevelopment,2024).

ThesedevelopmentsraisequestionsabouttheenvironmentalsustainabilityofusingAIforscientificresearch(SamuelandLucassen,2022).AwarenessofAI’senvironmentalcosts

varieswidelyamongresearchers,particularlyinacademicsettingswherecomputational

resourcesmayseem‘free’andtheirenvironmentalimpactlargely‘invisible’(Lannelongueetal.,2021b).Emergingeffortsarebeginningtoquantifythecarbonfootprintofdata-drivenmethodsacrossdisciplinessuchasastrophysics,proteinscience,healthresearchand

computing(LannelongueandInouye,2023;Jahnkeetal.,2020;Lannelongueetal.,2023).Forinstance,astudyfoundthattraininglargedeeplearningmodelssuchasAlphaFoldandESMFoldcanhaveacarbonfootprintofover100tonnes(LannelongueandInouye,2023),theequivalentofpoweringover20homesintheUnitedStatesforayear,demonstrating

thatevenscientificbreakthroughsasimpactfulasproteinpredictioncancomeatsteepenvironmentalcosts.

Theseearlyfindingsprovideevidenceonhow‘sciencecan,andfrequentlydoes,impact

theenvironment’(Lannelongueetal.,2023),raisingquestionsabout‘whoshouldbe

responsiblefortheseimpactsandhowtheseresponsibilitiesshouldbedistributed’(SamuelandLucassen,2022).AsAIandotherdata-intensivetechnologiesbecomemoreembeddedinscientificworkflows,researchersandfundingbodiesareincreasinglyintegrating

sustainabilityandenvironmentalconsiderationsintoresearchethicsframeworks,exploringwaystosupportmoreenvironmentallyresponsiblescientificpractices(Murrayetal.,2023).

Introduction

7

Section1:Foundationalconceptsandframeworks

AddressingtheenvironmentalimpactofAIrequiresamulti-dimensionalapproachthat

integratestheprinciplesofsustainableAI,greenAIandresponsibleAI.ThisinvolvesnotonlyreducingthedirectenvironmentalcostsofAItechnologies,suchasenergyconsumption

andcarbonemissions,butalsodevelopingholisticframeworksthatassesstheirlong-termsustainability.Theseeffortsmustbealignedwithbroadercommitmentstoresponsible

andethicalAI,ensuringthattheuseofAIinknowledgeproductionandsocietycontributespositivelyandequitablytohumanandenvironmentalwellbeing.

ThissectionexamineskeyconceptsandframeworkstoconsiderwhenaddressingtheenvironmentalimpactofAIinscientificresearch.

1.1GreenAI

GreenAIisafielddedicatedtounderstandingandreducingtheecologicalfootprintof

computinganddata-driventechnologies.Itpromotesenergyefficiencyandsustainable

computingpracticesthroughouttheentireAIdevelopmentlifecycle(Verdecchiaetal.,

2023).ItaimstolowercarbonemissionsandenvironmentalcostswhilesupportingAI

adoptionthatisalsosociallyinclusive(Schwartzetal.,2020).AcentralconcernofGreenAIistheincreasingenergyuseinhigh-performance“state-of-the-art”research,alsoreferredtoasRedAI.

RedAIemphasizesperformanceandaccuracy,oftenrelyingonenergy-intensivemodelsthatrequirevastdatasets,substantialcomputationalresourcesandrepeatedtraining(Bolón-

Canedoetal.,2024).Thesepractices,largelydrivenbyBigTechcompanies,havefuelledthedevelopmentofincreasinglycomplexandresource-demandingmulti-purposeAIsystems.

Section1:Foundationalconceptsandframeworks

Inacompetitiveresearchlandscape,thepressuretoachievesuperiorresultshasfurther

intensifiedtheenvironmentalandfinancialcostsofAIdevelopment(Luccionietal.,2024a).

Incontrast,GreenAIencouragesmoreenergy-efficientalternativesthatpreservemodelperformancewhilereducingresourceuse(Schwartzetal.,2020).Thisincludesadopting

technicalsolutionstomakealgorithmdesign,hardwareanddatamanagementmore

energy-efficient.GreenAIalsosupportsthedevelopmentofnarrowerandsmallermodelsdesignedforspecifictasks,which,inadditiontoreducingenvironmentalcosts,canbemoreappropriateandeffectiveformanyresearchtasks,domainsandcontexts.ThesepracticesarediscussedinmoredetailinSection3.

Theemphasisonsmaller,task-specificandpurpose-drivenAIsolutionsreflectstheprincipleofenergysufficiency,whichinvolvesintentionallyreducingenergyconsumptionthrough

energy-consciousdesign.Thisleadstomoreaffordableandaccessiblearchitectures

suchasFrugalAI,whichsupportstheuseofAIinsettingswithlimitedcomputingpowerandinfrastructure,whichworkwellinenvironmentswithscarceresources(Yamada,2024).

8

Relatedmethods,suchasEdgeAIanddistributedcomputing,prioritizelocaldataprocessingandofferpromisingpathwaysfordevelopingAItechnologiesthatareenvironmentallysustainable,privacy-consciousandcost-effective.

1.2SustainableAI

ThegrowingenvironmentalcostsofAIareoftentiedtoconcernsaboutsustainability.ThefieldofsustainableAIconsiderssustainabilityfromtwoperspectives(VanWynsberghe,2021).

AIforSustainabilityreferstotheapplicationofAIsystemstoaddressSustainable

DevelopmentGoals,includingenvironmentalsustainability,anditsapplicationsinrelevantfieldssuchasclimatescience,environmentalmonitoringornatureprotection.Thesecond,SustainabilityofAI,referstothe‘sustainabledevelopment’ofAIsystems,emphasizingpracticesthatreducetheirenvironmental,economicandsocialcosts.

BothapproachesaregroundedintheUnitedNationsBrundtlandCommission’sdefinitionofsustainability,whichemphasizes‘meetingtheneedsofthepresentwithoutcompromisingtheabilityoffuturegenerationstomeettheirownneeds’(TheWorldCommissionon

EnvironmentandDevelopment,1987).ItrecognizestheinherenttensionofpursuingtheinnovationandefficiencyaffordedbyAIsystems,whileaddressingtheirenvironmental,economicandsocialimpacts.

ApplyingasustainabilitylensrequiresresearcherstoholisticallyaddressdimensionsinAIdevelopmentbeyondperformanceandaccuracy(VanWynsberghe,2021).Areviewof100highlycitedmachinelearningpapersfoundthatonly15percentaddressedsocietalneeds,andjust1percentdiscussednegativeimpacts.Mostpapersemphasizedvaluessuchas

performance,efficiencyandscalabilityasopposedtoaddressingimpactsforunderservedcommunitiesorthereductionofenvironmentalharm(Birhaneetal.,2022).

TherapidlygrowingecologicalfootprintofAItechnologieshasalsoresultedincoordinatedactiontowardsunderstandingandaddressingthesustainabilityofAI(Ramanetal.,2024).

Theseinclude,butarenotlimitedto:

CoalitionforsustainableAI

:AninitiativebytheFrenchgovernment,the

UN

EnvironmentProgramme

andthe

InternationalTelecommunicationUnion

,aimedat

promotingtheresponsibledevelopmentanduseofAIforsustainabledevelopment,withanemphasisonthedevelopmentofAIthatisrespectfulofplanetaryboundaries.

CoalitionforDigitalEnvironmentalSustainability(CODES)

1:Aglobalhubforpolicy-makers,academics,technologycompanies,andNGOstoleadandcontributetodigitalsustainabilityinitiatives,includingmitigatingthenegativeenvironmentalandsocial

impactsofdigitalization.

SustAIn:SustainabilityindexforAI

:Aframeworkandpracticalchecklistsdevelopedtosupportorganizationsinadvancingtheenvironmental,socialandeconomicdimensionsof‘sustainableAI’,understoodasthedevelopmentanddeploymentofAIsystemsthat

respectplanetaryboundaries,donotreinforceproblematiceconomicdynamics,anddonotendangersocialcohesion.

Section1:Foundationalconceptsandframeworks

1TheInternationalScienceCouncilisamemberofCODES.

9

Together,theseeffortsreflectagrowingrecognitionthataddressingtheenvironmentalimpactofAIrequiresaholisticandinterdisciplinaryapproach.

1.3EthicalandresponsibleAI

ReducingtheenvironmentalimpactofAIdevelopmentalsorequiresattentiontoitsethicaldimensions.ThegrowingadoptionofethicalAIandresponsibleAIframeworkswithinscientificinstitutions(Banoetal.,2025)reflectsarecognitionthatprinciplessuchas

fairness,accountabilityandtransparencyarekeytobuildingtrustinAIsystems(ISO,n.d.).Thisshiftisalsoevidentacrossdiversecontexts,includingcountriessuchasAustralia,

Uruguay,ChinaandMalaysia,wheretheresponsibleuseofAIinscienceisatthecentreofnationalAIstrategies(Castleetal.,2024).

However,todate,manyoftheseframeworksfailtoframesustainabilityasacomponent

ofethicalandresponsibleAIpractice(Luccionietal.,2025).A2019globalreviewfound

thatonly14of84AIguidelinesmentionedsustainability(Jobinetal.,2019),including

the

EuropeanGuidelinesforTrustworthyAI

and

theUNESCORecommendationontheEthics

ofAI

,consideringbothenvironmentalsustainabilityandsocietalwellbeing.Inresponsetogrowingevidence,AIresearchersarecallingforathirdwaveofAIethicsthatdirectly

addressestheenvironmentalcrisisandplacessustainabledevelopmentatitscentre(Luccionietal.,2025).

Recentinitiativeshavebeguntoaddressthisgap.TheWorkingGrouponResponsibleAIoftheGlobalPartnershiponArtificialIntelligencehasdevelopeda‘

ResponsibleAIStrategyfor

theEnvironment

’tohelpMembercountriesassesstheenvironmentalfootprintofAImodelsandapplications.Researchfundingbodies,suchasthe

WellcomeTrust

,areintegrating

environmentalconsiderationsintofundingrequirements.Newresearchethicsframeworks,suchasenvironmentalresponsibility,arealsoemergingtoguideresearchersinadoptingmoreecologicallyconsciousapproacheswhenconceiving,planning,conductingand

concludingresearch(Murrayetal.,2023).

Applyinganethicallenstosustainabilitycanalsohelpresearchersunderstandthe

implicationsofenvironmentalharmforvulnerablepopulations.ResearchersfromHuggingFaceandtheDistributedAIResearchInstitutehavehighlightedhowtheunevenglobal

distributionofAI’secologicalimpactscanexacerbateexistingsocio-economicdisparitiesandsocial-ecologicaljusticeconcerns(Luccionietal.,2025).Examplesincludehowtheconstructionofdatacentrescontributestowaterscarcityindrought-proneregionsandto

theproliferationofelectronicwasteinnearbycommunities(Barrattetal.,2025).AddressingenvironmentalequitycontributestothedevelopmentofAIsystemsthatarenotonlyenergy-efficientbutalsofair,justandequitable(Lietal.,2023).

Section1:Foundationalconceptsandframeworks

10

Section2:EnvironmentalimpactofAIacrossthelifecycle

AssessingthesustainabilityofAIrequiresaholistic,interdisciplinaryapproachthataccountsforenvironmental,economicandsocialcoststhroughouttheAIlifecycle–fromhardware

manufacturingandmodeltrainingtodeploymentanddecommissioning(Luccionietal.,

2024b).Althoughtherearestillnostandardizedmethodsforevaluatingsustainabilityacrossallstages,significantprogresshasbeenmadeinmeasuringenvironmentalimpacts.

2.1EnvironmentalimpactsofAI

Researchersdistinguishbetweentwotypesofenvironmentalimpacts(OrganizationforEconomicCo-operationandDevelopment,2022):

1.Directimpactsarefirst-orderimpactsfromtheAIlifecycleanddescribe‘computing-related’consumptionofresources:energy,waterandmineralresources,aswellas

emissionsande-wastegeneratedfromoperatingAIsystems.Evidenceindicatesthatthesearepredominantlyharmfultotheenvironmentandecosystems.

2.IndirectimpactsareeffectsdrivenbytheimmediateapplicationofAIacrosssectors.Theyaredifficulttopredictandcanbeeitherbeneficialorharmful.Forexample,AImayimproveenergyefficiencyinsmartgridsbutcanalsodriveunsustainableconsumptionpatternsthroughautomation.

Therearetwotypesofindirectimpacts:second-orderimpactsandhigher-orderor

‘system-levelimpacts’.BothareindirecteffectsthatemergeovertimeasAIreshapes

SectionEnvironmentalimpact2:ofAIacrossthelifecycle

productionandconsumptionpatterns,withthelatterpotentiallydrivingstructuralchangesintheeconomyandsociety.Indirectimpactscouldultimatelyoutweightheefficiencies

obtainedbystrategiesfocusedonmitigatingfirst-orderimpacts.

Reboundeffectsareanadditionalcategorythatreferstosituationsinwhichefficiency

gainsleadtogreaterresourceusethatcanoffsetinitialenvironmentalbenefits.Forexample,moreefficientdatacentres,whileconsuminglessenergyperoperation,mightbeusedto

handlemoredataandpowermoreservices,ultimatelyincreasingtotalenergyconsumption.Anticipatingandmeasuringreboundeffectscanillustratehowinitialefficiencygainscanbeoutweighedbynegativeconsequences,challengingthenotionthatenergyefficiencyalonewilldeliverclimatebenefits.

Whilethereareincreasingmethodologiesandmetricstomeasuredirectenvironmental

impactsofAI,limitedconsensusremainsonhowtomeasureindirectsecond-orderand

higher-ordereffects(OECD,2022).GiventhegrowinguseofAIinscienceanditspotentiallarge-scaleeffectonplanetaryhealth,thereisaneedforaresearchagendathataddressestheindirecteffectsofAIsystemsonknowledgeproduction.

11

2.2Alifecycleapproachtoassessingenvironmentalimpact

ThissectionexamineskeyapproachestounderstandingandmeasuringdirectenvironmentalimpactsacrosstheAIlifecycle.

AIresearchersandinstitutions,suchastheOrganizationforEconomicCo-operationandDevelopment,theUnitedNationsEnvironmentProgramme,theUNCommissionforTradeandDevelopment(UNCTAD,2024),andtheInternationalTelecommunicationsUnion,

amongothers,proposeusingalifecycleorend-to-endapproachtoassessenvironmentalimpacts(OECD,2022;UnitedNationsEnvironmentProgramme,2024;International

TelecommunicationUnion,2024).ThemostestablishedmethodologyisLifeCycle

Assessment,whichevaluatestheenvironmentalburdensassociatedwithaproductorservice,fromtheextractionofrawmaterialstowasteremoval(Klöpffer,1997).

Alifecycleapproachcanhelpillustratethewiderscopeofdirectenvironmental

costsinvolvedindesigninganddeployingamodel,clarifywhereemissionsoriginate

andconcentrate,andhighlightwheretargetedinterventionsand“sustainableAIpractices”canreduceenvironmentalcosts.TheAIlifecycleistypicallydescribedintwolayers,or

“stacks”:softwareandhardware.

SOFTWARELAYER

Thefirstistheoperationalorsoftwarelayer(UNEP,2024),alsocalledthemodel

developmentcycle(Wuetal.,2022).ThisincludesalldecisionsregardingthedataandAI

modelarchitecturethatinfluenceenergyandresourceuse.Itcoversrelevantdatawork

(datacollectionandpreparation),algorithmicdesign(modelarchitecture),modeltrainingandmodeldeployment(includinginference)(Clemmetal.,2024).

Theenvironmentalcostsofthislayerarecalledoperationalcostsandarecorrelatedto

thesizeofdatasetsandthemodelarchitecture,whichdirectlyaffectcomputetimeandcarbonfootprint(Chenetal.,2022).Trainingandinferenceareusuallythemostsignificantcontributorstolifecycleenergyconsumption.Forexample,researchfoundthroughanAI

inferenceimpactassessmentthatgenerativearchitecturesaremoreenergy-intensivethantask-specificmodelsasaresultofinference(Luccionietal.,2024a).Duetothecommercialdeploymentoflargemulti-purposegenerativeAImodels,likeOpenAI’sGPT-4,inference

nowrepresentsanincreasingmajorityofAI’senergydemands,withanestimated80–90

percentofcomputingpowerinAItra

温馨提示

  • 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
  • 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
  • 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
  • 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
  • 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
  • 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
  • 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。

评论

0/150

提交评论