人工智能环境可持续性的标准化 Standardization for AI Environmental Sustainability

Standardizationfor

AIEnvironmental

Sustainability

TowardsacoordinatedglobalapproachUpdatedversion

February20th,2026

AnupdatepublishedfortheAIImpactSummit

COAI

ThisworkstemsfromaglobalinitiativelaunchedonOctober10,2024,atUNESCOheadquarters,bringingtogetherexpertsfromISO,ITUandIEEE,inpartnershipwiththeOECDandUNESCO.LedbytheFrenchMinistryinchargeofEnvironment,thisinitiativeledtothepublicationofafirstdocumenttoensurebettercoordinationbetweenstandardizationbodiesandoptimizeresourcesdedicatedtoassessingandreducingtheenvironmentalimpactofAI.ThefirstdocumentwaspublishedinthecontextoftheParisAIActionSummit(February10-11,2025).Thisnewversionisanupdateofthedocument,elaboratedthroughanewworksessionwithstandardizationorganizationsonDecember4th,2025,andsubsequentfeedbackgivenbyexperts.ItispublishedinthecontextoftheAIImpactSummitinIndia(February19-20,2026)toensurecontinuouscoordinationbetweenexperts.

Leadorganization

GOUVERNEMENT

COALITIONFOR

SUSTAINABLE

Partners

2

Peralphabeticalorderoflastnames

DerickOhmarAdil,ISO,GlobeTelecom,Philippines**

IsabelBarberá,Rhite,TheNetherlands*

AlexisBaria,ULStandards&Engagement,UnitedStates**

SylvainBaudoin,TheShiftProject,France*

BertrandBraunschweig,BiLaB,France*

NorbertBensalem,DirecteurStandardisationIBMFrance*

Jean-ManuelCanet,Orange,France&ITU-TSG5,Switzerland*

NathalieCharbonniaud,Orange,France*

HélèneCostadeBeauregard,Ecolab-FrenchMinistryofEnvironment,France**

VincentDanno,independantexpert,France*

RenaudDiFrancesco,EuropeTechnologyStandardsOffice,Italy-Inmemoriam*

HarmEllens,Independantexpert,Australia*

JulietteFropier,Ecolab-FrenchMinistryofEnvironment,France

CarolineGansCombe,OmnesEducation-INSEEC,France**

BorisGamazaychikov,Salesforce,USA/France*

ArtiGarg,AVEVA,US**

PaoloGiudici,UniversityofPavia,Italy*

ShubhamGupta,NoblesoftSolutionsInc.,UnitedStatesofAmerica**

AhmedHaddad,Arcep,France*

MikiHashimoto,MitsubishiElectric,Japan*

MathildeJay,Ecolab-FrenchMinistryofEnvironment,France**

YoungImCho,GachonUnivertiy,Korea*

SusannaKallio,Nokia,Finland

PolinaKoroleva,UNEP,Kenya

JacquesKluska,SchneiderElectric,France*

ValerieLivina,NationalPhysicalLaboratory,UnitedKingdom*

SashaLuccioni,HuggingFace,Canada*

NicolasMiailhe,GlobalPartnershiponAI(GPAI),France*

GritMunk,DanishAssociationofEngineers,Denmark*

SonNguyen,GreenTransformationandSustainabilityNetwork,Vietnam**

ArvinObnasca,BeEthical,Philippines*

EnricoPanai,AssociationofAIEthicists,France*

AaronPietzonka,Ecolab-FrenchMinistryofEnvironment,France*

VincentPoncet,Google,France*

PierreRiou,ACIMEOPresident,France*

Robert(Bob)Spence,I-Partnerships(IEEEP7100WG),UnitedKingdom

EmiliaTantar,BlackSwanLUX&CEN/CLCJTC21AIWG2,Luxembourg*

MarinaTrancheva,AMPECO,Bulgaria**

AuroreTual,Thales,France*

ReynaUbeda,TelecommunicationStandardizationBureau,InternationalTelecommunicationUnion,Switzerland

ArlettevanWissen,RoyalPhilips,TheNetherlands*

FrankWisselink,DeutscheTelekom,Germany*

DavidWotton,Independantexpert,Australia*

*Contributorofthe2025versiononly.

**Contributorofthe2026updatedversiononly.

AIenvironmentalsustainabilitystandardization

Towardsacoordinatedglobalapproachto3

AIenvironmentalsustainabilitystandardization

Towardsacoordinatedglobalapproachto4

OBJECTIVE

FacedwiththerapidgrowthinAIuseandthegrowingawarenessofitsenvironmentalimpact,numerousinitiativesareunderwayaroundtheworldtobetterassessthisimpact,andtodevelopguidelinesandstandardsonhowtocalculate,report,reduceandpreventitsenvironmentalimpacts.

Besidesreasonsofcompliance,measuringtheenvironmentalimpactofAIsystems

shouldbeareflectionofenvironmental

sustainabilityasavaluethroughouttheAIlifecyclewhichisgloballyapplicable.

Theobjectiveofthisapproachisto

ensuretheefficientuseofresources,

enhanceclarity,promoteconsistencyinAIenvironmentalsustainability

standardization,andfacilitatethe

widespreadadoptionofbestpractices.

Theintentionofitscontributorsisto

worktowardsnon-conflictingstandardsandtofostercollaborationbetween

internationalstandardizationbodiestominimize,asmuchaspossible,their

duplication,contradictionandoverlap.

Theseinitiativesfaceseveraldifficulties,

includingalimitednumberofexpertswith

TARGETAUDIENCE

Thisdocumentisintendedfor

dualcompetencyinArtificialIntelligenceandenvironmentalsustainability,compoundedbyknowledgegapsduetothelackofrobust

qualitativeandquantitativedata.

policymakers,governmentofficials,

scientists,AIdevelopersandengineers,andindustryleadersworkingonorinterestedinAIenvironmentalsustainability,providing

themwithvisibilityintotheprogressmadebystandardizationorganizationsandthe

workthatstillliesahead.

Italsoservesasavaluableresourcefor

stakeholdersbroadlyinvolvedinAI.This

initiativeoffersanopportunitytoshowcasetheareasofworkofthestandardization

bodiesforgreatertransparencyandimprovedcollaboration.

Toavoidthedevelopmentofconflictingor

contradictingmethodologiesthatwould

undermineglobaleffortstoincreasethe

environmentalsustainabilityofAIsystems,andtomakethemostoftheexpertiseavailable

internationallyonthistopic,acommon

approachisessentialtobringvisibilitytothoseinitiativesthatalreadyexistandtodefine

collaborationopportunitiestoenhanceacommonapproachtoSustainableAI.

Partnersofcivilsociety,International

Organizations,administrationsandcompaniesaregatheringattheAIImpactSummitinIndiaon19-20February2026aroundthekeytopic

thedemonstrableimpactofAIforthePeople,PlanetandProgress.Thisdocumentisbuildingonthismomentumandshowingthe

engagementofexpertsandorganizationsto

thoroughlyandefficientlyadvanceonguidanceandstandardsaroundAIsustainability.

AIenvironmentalsustainabilitystandardization

Towardsacoordinatedglobalapproachto5

STATEOFTHEARTAND

NORMATIVEREFERENCES

AnumberofdocumentsrelatingtoAI

sustainability,includingspecificallywithin

theInformationandCommunications

Technologysector,havealreadybeen

publishedandcanserveasasolidbasisforfuturestandardsandguidance.Several

werepublishedin2025.(SeeAppendix2)

Additionally,severalprojectsarecurrentlyunderway.(SeeAppendix3)

Standardscanvaryintheexacttermsbutthesearelight,commonly-agreed

definitionstofacilitatetheunderstandingofthedocument.

→ArtificialIntelligenceSystem:Machine-

basedsystemthatisdesignedtooperate

withvaryinglevelsofautonomyandthat

mayexhibitadaptivenessafterdeployment,andthat,forexplicitorimplicitobjectives,infers,fromtheinputitreceives,howto

generateoutputssuchaspredictions,

content,recommendations,ordecisionsthatcaninfluencephysicalorvirtual

environments.(EUAIActArticle3(1))

→Environmentalsustainability:Statein

whichtheecosystemanditsfunctionsaremaintainedforthepresentandfuture

generations.(ISO17889-1:2021,modified—generationmadeplural)

→Environment:Surroundingsinwhichanorganizationoperates,includingair,water,land,naturalresources,flora,fauna,

humansandtheirinterrelationships.(ISO14001:2015)

→Environmentalaspect:Elementofanorganization’sactivitiesorproductsorservicesthatcaninteractwiththe

environment.(ISO14001:2015)

→Environmentalimpact:Anychangetotheenvironment,whetheradverseorbeneficial,whollyorpartiallyresultingfroman

organization’senvironmentalaspects.(ISO14001:2015)

→GeneralPurposeAI:AnAImodel,

includingwheresuchanAImodelistrainedwithalargeamountofdatausingself-

supervisionatscale,thatdisplays

significantgeneralityandiscapableof

competentlyperformingawiderangeof

distincttasksregardlessofthewaythe

modelisplacedonthemarketandthatcanbeintegratedintoavarietyofdownstreamsystemsorapplications.(EUAIActArticle363))

→AICompute:Thecomputational

resources,includinghardwareandsoftwareinfrastructure,requiredduringtraining,

inference,validation,ordeploymentofAI

models.Thisencompassestheunderlyingelectricalgridwithitsfuelmixof

generationthatdefinescarbonintensity,

energysystems,datacenterinfrastructure,andsupplychainsthatprovidethepowerandcoolingnecessarytosustainthese

operations.

→EnvironmentalLifeCycleAssessment

(LCA):Compilationandevaluationofthe

inputs,outputsandthepotential

environmentalimpactsofaproductsystemthroughoutitslifecycle.(ISO14040:2006)

AIenvironmentalsustainabilitystandardization

Towardsacoordinatedglobalapproachto6

→LifecycleofAISystems(ISO/IEC5338:2023):

oInception

Designanddevelopment

Verificationandvalidation

Deployment

Operationandmonitoring

Continuousvalidation

Re-evaluation

Retirement

→LifecycleofAIComputeResources:

Rawmaterialacquisition

Production

TransportationoUse

End-of-life

→Second-ordereffect:Theindirectimpactcreatedbytheuseandapplicationof

InformationandCommunication

Technologies(ICTs),whichincludeschangesofenvironmentalloadduetotheuseof

ICTsthatcouldbepositiveornegative.(ITU-TL.1480)

→Higher-ordereffect:Theindirecteffect(includingbutnotlimitedtorebound

effects)otherthanfirstandsecondordereffectsoccurringthroughchangesin

consumptionpatterns,lifestylesandvaluesystems.(ITU-TL.1480)

→Reboundeffect:Increasesin

consumptionduetoenvironmental

efficiencyinterventionsthatcanoccur

throughapricereductionorother

mechanismincludingbehavioralresponses(i.e.,anefficientproductbeingcheaperorinotherwaysmoreconvenientandhence

beingconsumedtoagreaterextent).(ITU-TL.1480)

→Scope1/2/3emissions:refertothe

lastestversionoftheGreenhouseGas

Protocoldefinitionofscope1/2/3emissions.

AIenvironmentalsustainabilitystandardization

Towardsacoordinatedglobalapproachto7

EXTENTOFTHE

UPCOMINGWORK

AnumberofnormativegapsrelatingtoAIsustainabilityhavebeenidentified.Asafirststepoffuturestandardization

efforts,commonstructureacross

existingmethodologieshavebeen

identified,helpingtoestablishwhich

approachesarebestsuitedtospecific

contexts.Basedonthoseidentified

gaps,collaborationbetweenexpertscantakeplaceacrossorganizations.Those

areasofworkareidentifiedbelow.Theyhavebeenupdatedsincethefirst

publicationoftheglobalapproach.

1.Definingtransparentandcommon

indicators,andareportingframework

ThefirstobjectiveofstandardizationforAIsustainabilitywillbetodevelopcommon

environmentalindicatorsthatare

measurableorcanbeestimatedforeachlifecyclestageofAIsystemresources.

Theseindicatorsmustberelevantto

specificwell-definedperimeters

(organizationalperimeter,serviceperimeter,etc.)thataresharedacrossorganizations

(offeringandconsumingAIservices).

Reportingontheindicatorsshouldbedoneinauniform,formalizedandtransparent

waytoenablemeaningfulcomparisonsbetweendifferentassessments(for

differentorganizationsorbetweenupdates).

Incorporateorganizations,accordingto

theirroleinthevaluechain,thiscouldforexamplebepartoftheScope1/2/3

reportingintheCorporateSocialResponsibility(CSR)strategy.

Theproposedindicatorsshouldalsoalignwithexistingenvironmentalreporting

frameworksliketheGlobalReporting

Initiative(GRI),GHGProtocolorISO14000standardsthatcompaniescommonlyuse.

Expertsinenvironmentalsciences,data

(e.g.,datascientists,statisticalexperts,anddatagovernancepractitioners),andAI

lifecycle,attheminimum,areneededinthedevelopmentofrobustenvironmental

indicatorstoensuredataquality,

implementationfeasibility,indicator

validation,riskmanagementand

consistencyacrossdifferentcountriesandregions.

2.Assessingdirectenvironmentalimpacts

TomanagetheenvironmentalimpactofAI

andtomakeinformeddecisions,thesecondobjectiveofstandardizationwillbeto

establishmethodologiesfortheassessmentoftheindicators,includingLifeCycle

AssessmentforAIsystemsandAIservices.Existingassessmentmethodologiesforthedigitalsector(seeAppendix1)could

possiblybeadaptedforAIsystems.

Thesemethodologiesshouldbegeneric

enoughtobeapplicabletothewidevarietyofAIsystems(fromgeneralpurposeto

domain-specificAI,etc.).Theymayincludeseveralscopes:AIsystem,AIservicebasedonseveralAIsystems,useofAIatthelevelofanorganization,useofAIatthe

territoriallevel(communities,countries,

etc.),anddifferentimplementationand

servicemodels,suchascloud-based,on-premise,onedge,etc.andwhetheritisanembeddedsystemorgeneral-purpose

operatingsystem.Thescopesofthe

differentevaluationmethodsmustrelyonthesamesetofmetrics.

AIenvironmentalsustainabilitystandardization

Towardsacoordinatedglobalapproachto8

Furthermore,theperimeterofthis

assessmentshouldbeascomprehensiveaspossible,coveringtheentirelifecycleofAIsystems.Thisincludesdesign,

inferenceandtuningphases,aswellas

embodiedimpacts,e.g.,productionand

end-of-lifeimpactsofthehardwareusedtorunthevariousphasesmentioned

earlier.

3.IdentifyingbestpracticesformitigationoftheenvironmentalimpactofAI

Thethirdobjectiveofstandardizationistoidentifystrategiestoreducethe

environmentalimpactofAlsystemsonatleastoneoftheindicators.

Thestrategiescanbeidentifiedfor

different‘actiondimensions’(like

challengingtherelevanceofusingAl,

infrastructuresizing,modeloptimization,implementationefficiency,etc.),and

impactdrivers(e.g.,usinglessresources,usinglow-carbonresources,etc.).

Thesestrategies,accompaniedbytheiradvantagesanddisadvantages,

implementationcontexts,keysuccessfactors(orconditionsofrelevance)andassociatedtracking(orfollow-up)

indicators,canbeidentifiedasbestpracticessharedbyallinvolved

stakeholders.

Strategiescanbeaccompaniedwith

guidelinesonhowtofacilitatestakeholderengagementandcollaboration.

Technicalstandardsforemerging

technologiesthatcanimprove

environmentalimpact(immersivecooling,automateddatacollection,server

virtualization,efficientmodel

architecturesandtrainingmethods,etc.)willalsohelpencouragetheshared

adoptionofbestpractices.

4.Definingrelevantmanagementsystems

Thefourthobjectiveistobeabletomakedecisionsontheinitialization,continuityorretirementofanAlsystem,takinginto

accountalltheAlsystemsinan

organization,theirbenefitsandcostfortheenvironment.

Withthisholisticviewinmind,some

guidanceshouldbegivenonhowto

prioritizedifferentmitigations,trade-offs(e.g.,betweenenergyconsumptionand

improvingaccuracyortestingfor

robustness)andalternatives(including

non-Al)thatneedtobetakenintoaccount.

Fororganizations,guidanceshouldbe

elaboratedontherelevantmanagement

systemstosystematicallysupportthe

environmentalsustainabilityofAland

balancingcompetingprioritiesatthelevelofanorganization.

5.Assessingindirectenvironmentalimpacts

Thefifthobjectiveistofurtherdevelop

methodologiesforassessingthesecondandhigher-orderedeffectsofAlsystems.lt

includesbothindirectpositiveeffects(e.g.electrictyconsumptionavoidedthankstoanAlsystem,reuseoftheheatproducedbydatacenters,decarbonizationof

industrialprocess...),aswellasnegativeeffectssuchasthereboundeffect(i.e.efficientgainsdrivingincreaseusage).

Theseassessmentsshouldemployidenticalmetricstothoseusedinevaluatingthe

directenvironmentalimpactofAlsystems.

AIenvironmentalsustainabilitystandardization

Towardsacoordinatedglobalapproachto9

6.Standardizingnewenvironmentalindicators

Thesixthobjectiveistodevelop

methodologiesforadditionalenvironmentalindicators,drawinguponthematurityoftheacademicliteratureandindustry

experiences.

Areassuchasbiodiversityimpactandnoisepollutionrepresentpotentialdomainsfor

exploration,contingentupontheirassessedlevelofmethodologicalmaturity.

7.EstablishingstandardsforAIsustainabilityliteracy

Theseventhobjectiveistofocuson

establishingcomprehensivestandardsfor

sustainableAIliteracyacrossstakeholder

communities.Thesestandardswoulddefinecorecompetenciesneededtounderstand,

evaluate,andcommunicateaboutAIsystems'environmentaldimensionsthroughouttheirlifecycle.

Suchliteracyframeworksshouldaddress

theknowledgerequirementsfordiverse

audiences,fromAIdevelopersand

deployerstopolicymakers,endusersand

thegeneralpublic,ensuringallparticipantsintheAIecosystemcanmakeinformed

decisionsregardingenvironmental

implications.Thestandardsshouldevolvealongsidetechnicalmethodologies,

maintainingalignmentwithcurrent

environmentalassessmentpractices.

AIenvironmentalsustainabilitystandardization

Towardsacoordinatedglobalapproachto10

1.Indicatorsfortheenvironmental

assessmentofAIsystemsinclude,butarenotlimitedto,globalwarmingpotential(kgCO2eq),energyconsumption(kWhorMJ),

waterconsumptionandwithdrawal(m3orL),andrawmaterialconsumption(kg).Itisessentialtotakeintoaccounttheenergy

gridinterconnectionandnationalfuel

mixestoquantifythecarbonfootprintfromtheelectricitygenerationfortheAIneeds.

2.Thetargetoftheenvironmental

assessmentshouldbethattheentire

lifecycleoftheAIsystemmustbe

subjectedtoanenvironmentallifecycleassessment.

Thisshouldincludeevaluatingthe

environmentalimpactofthe“training”

phase(inception/design&development/

verification&validation/deployment),whileconsideringthegenealogyofmodelsthat

mayhaveservedinthestepsofpre-training,post-training,fine-tuning,

instructiontuninganddistillation.

Thisshouldincludeaswelltheassessmentoftheenvironmentalimpactofthe“use”or“inference”phase(operationand

monitoring/continuousvalidation/re-evaluation).

Bothphasesshouldbereported.

Thesephasesmightbeattributed

differentlyacrossentitiesororganizations

involvedinthedevelopmentanduseofAI.Differentscopesmaybeusefulfordifferentstakeholders:

AreportingoftheAIsystemsthatcouldbeaggregatedforcorporatereporting,

likeemissionsperyear;

Areportingperunitofwork(pertoken

inthecontextofLLMs,orperaspecificsizeofimageforanimageclassifier),foruserstoconsiderthecriteriatochoose

asystemandtobeabletoincludetheiruseofthesystemintheirowncorporateinventory.

3.ThedatalifecycleofanAIsystemcan

bringsignificantaddedenvironmentalcostsforcollection,pre-processing,transfer,

update,andstorage.Thesecostsattributedtotraining,testing,inputoroutputdata

shouldbeincludedintheassessment,fortherelevantlifecyclestages.

4.Indirecteffectsincludesecond-order

andhigher-ordereffects(asdefinedintheTerminologyandconceptssection),for

examplereboundeffects.Theindirect

effectsofanAIsystemshouldbeassessedatleastqualitatively.

Ifquantitativeassessmentisnotfeasible,ajustificationmustbeprovided.The

assessmentofindirecteffectsshouldbeseparatedfromtheassessmentofdirect,firstordereffects.

AIenvironmentalsustainabilitystandardization

Towardsacoordinatedglobalapproachto11

5.AllequipmentusedthroughoutthelifecycleoftheAIsystemshouldbedocumented.

Thisincludes,butisnotlimitedto,the

equipmentdedicatedto:computing

infrastructure,datacollectiondevices,

storagesystems,userdevices(e.g.,robots,smartdevices),andnetworkequipment.

Sincethesephysicaldevicesmightserve

multipleAIsystemsorotherservices,awaytoallocateenvironmentalimpactofthe

equipment–particularlyitsproductionandend-of-life–toaspecificsystemneedstobedefined.

Forexample,theenvironmentalimpact

couldbeproportionallyattributedbasedonthedurationoftheequipment’susebythesystemoveritstotallifecycle.

6.Bestpracticesshouldbedeveloped

acrossseveralareas:equipment,data

management,modelperformance,hardwareutilization,measurementandcommon

metrics,includingkeyperformanceindicators.

Keyinitialbestpractices,suchasdynamicsizingofcomputingresourcesorgreen

coding,arealreadysharedacrossthedigitalsectorandneedtobeimplementedat

scale.

Bestpracticesaroundorganizational

governancearealsoneededtoensurethatthecorrectmonitoringandmitigationoftheenvironmentalimpactsofAIisputforward,consideringcomplementaryincentiveslikeperformanceandmonetarycost.

Choicesonwhatthepriorityisforsystem

optimizationshouldbedocumented,giventhattherearemanytrade-offsbetween

differenttypesofperformancemetricsthatcan‘interfere’withsustainability,like

optimizingforfairness,robustness,or

privacy.EvaluatingtherelevanceofAIasasolutioncomparedtolessenvironmentallycostlyalternativesisabestpracticein

itself.

7.Thesestandardsandbestpractices

shouldsupportcodesofconductof

companiesandinstitutionsmakingthemmorerobust,reliable,comparableand

compatible.

AIenvironmentalsustainabilitystandardization

Towardsacoordinatedglobalapproachto12

On-goingandfuturestandardizationeffortsstillfaceanumberofchallenges:

RapidadvancementofAI:The

developmentsinAIsystemsare

advancingatanextremelyrapidpace

andmethodologiesandbestpracticesmustremainadaptabletoemerging

technologiescomingupinthefollowingyears.Expertsinstandardizationwill

needtomonitortheadaptationofstandardstocurrentstate-of-art.

Raisingawareness:Thepublicationofastandarddoesnotguaranteeits

practicalimplementationduetobarrierssuchasindustryreluctance,feasibility

ofdatacollection,costimplications,oralackofenforcementmechanisms.

Expertswillneedtoraiseawareness

throughworkshops,policybriefs,or

industrypartnershipsontheavailabilityandimplementationofthestandards,

encouragethepublicationofdata,andfacilitateguidanceforthe

implementation.

ComplexityofAIsystemdevelopment:

AIsystemsarenotstaticproducts.Overtheirlifetime,theywillberepeatedly

usedaswellasretrained,whichadds

complexitytodefiningthescopeoftheperimeterofanenvironmental

evaluation.Furthermore,the

environmentalcostofexperimentation,includingfailed,intermediateor

incompletetrainingruns,ishardtoattributetoaspecificAIsystem.

Exhaustivereporting:Guidelinesfor

reportingtheenvironmentalsustainability(includingimpact)ofAISystemsshould

identifythestakeholdersresponsibleforthisreportingthroughoutthevaluechain,andencouragecommunicationalongthe

valuechain.WhilesomeAI-dedicated

computationalfacilitiesexistandcanbe

directlymonitoredforelectricityand

waterconsumption,mostAItrainingandinferenceactivitiesoccurinmixed-use

facilities.Thesefacilitieshandlediverse

processesacrosssharedhardware

resourcessuchasGPUsandCPUs,makingitchallengingtoisolatetheenvironmentalimpactofspecificAIoperations.This

complexityrequiresthecontinuous

developmentofmethodologiestofairly

andtransparentlyallocateenvironmentalcostsamongcoexistingprocesseswithinmixed-useenvironments.Real-time

monitoringofresourceusewithinsharedfacilitiesprovidesanimportantbasisforallocationandestimation.Whilepreciseattributionmaybedifficult,the

environmentalimpactsassociatedwithsharedinfrastructureshallbeaccountedforratherthanexcluded,andthe

assumptionsanduncertaintiesinvolvedshallbedocumented.

AIenvironmentalsustainabilitystandardization

Towardsacoordinatedglobalapproachto13

Accesstoenvironmentaldata:Thelackofcollectedand/orsharedrobustdataon

keyparametersforcalculatingthe

environmentalimpactofAIsystemsposeschallengesfortestingamethodology

acrossmultiplesystemsandindicators.

Giventhatenergyconsumptiondataare

morereadilyavailable,adirect

environmentalassessmentfirstlimitedtoGHGemissionscouldbeconsidered.Inthelongerterm,addingindicatorssuchas

waterconsumption,materialconsumption,etc.,mustbeapriority.

Fordevelopingmoreaccurateassessment,itisencouragedtoshareinformationabout

embodiedemissionsofcommonhardware,foundationalmodels,etc.

Withthisproposedapproach,expertsfrommultilateralorganizations,companies,andadministrationscalltoactionandexpresstheirsharedcommitmenttocollaborateinensuringthatorganizationscanrapidly,

efficiently,andaccuratelyadoptstandardsforimprovedAIsustainability.

Asexperts,theywishtoreconvenebeforefutureAIsummitsorotherinternational

eventsofinteresttomonitor

implementationandtoupdatethisdocument.

Appendix1:

Diagramforpublishedandin-developmentstandards

assessmentfor

productsorat

organizationlevel

(generalguidelines)

ISO14040:2006,

ISO14090:2019

ISO14064-1:2018

ISO14068-1:2023

ISO59020:2024

Managementsystems

(forenvironmentalmanagement)ISO59020:2024

ISO14001:2015PRFIWA48

(forAI)

ISO/IEC42001:2023

ISO/IEC42005:2025

IEEEP2863

PublishedStandard

Approvedproject,indevelopment

StandardsspecifictoAI&Environmental

sustainability

Broaderstandards

Bestpracticesformitigationofenvironmentalimpact

IEEE1924.1(digitalarchitectures)ISO/IECTS8236-1and2(datacenter

managament)

ISO/IECAWITS42111(lightweightsystems)

ISO/IECAWITS42112(trainingefficiency)

IEEEP1927.1(virtualizednetwork)

IEEEP3403(multipledatacenters)

ISO/IECTS20125(Ecodesign)

ITU-TL.