歐盟人工智能法建立的人工智能系統(tǒng)的定義指南

ENEN

EUROPEAN

COMMISSION

Brussels,6.2.2025

C(2025)924finalANNEX

ANNEX

tothe

CommunicationtotheCommission

ApprovalofthecontentofthedraftCommunicationfromtheCommission-

CommissionGuidelinesonthedefintionofanartificialintelligencesystemestablished

byRegulation(EU)2024/1689(AIAct)

1

I.PurposeoftheGuidelines

(1)Regulation(EU)2024/1689oftheEuropeanParliamentandoftheCouncil(‘theAIAct’)1enteredintoforceon1August2024.TheAIActlaysdownharmonisedrulesforthedevelopment,placingonthemarket,puttingintoservice,anduseofartificialintelligence(‘AI’)intheUnion.2ItsaimistopromoteinnovationinandtheuptakeofAI,whileensuringahighlevelofprotectionofhealth,safety,andfundamentalrightsintheUnion,includingdemocracyandtheruleoflaw.

(2)TheAIActdoesnotapplytoallsystems,butonlytothosesystemsthatfulfilthedefinitionofan‘AIsystem’withinthemeaningofArticle3(1)AIAct.ThedefinitionofanAIsystemisthereforekeytounderstandingthescopeofapplicationoftheAIAct.

(3)Article96(1)(f)AIActrequirestheCommissiontodevelopguidelinesontheapplicationofthedefinitionofanAIsystemassetoutinArticle3(1)ofthatAct.ByissuingtheseGuidelines,theCommissionaimstoassistprovidersandotherrelevantpersons,includingmarketandinstitutionalstakeholders,indeterminingwhetherasystemconstitutesanAIsystemwithinthemeaningoftheAIAct,therebyfacilitatingtheeffectiveapplicationandenforcementofthatAct.

(4)ThedefinitionofanAIsystementeredintoapplicationon2February20253,togetherwithotherprovisionssetoutinChaptersIandIIAIAct,notablyArticle5AIActonprohibitedAIpractices.AsthedefinitionofanAIsystemisdecisivetounderstandingthescopeoftheAIActincludingtheprohibitedpractices,thepresentGuidelinesareadoptedinparalleltoCommissionguidelinesonprohibitedartificialintelligencepractices.

(5)TheseGuidelinestakeintoaccounttheoutcomeofastakeholderconsultationandtheconsultationoftheEuropeanArtificialIntelligenceBoard.

(6)ConsideringthewidevarietyofAIsystems,itisnotpossibletoprovideanexhaustivelistofallpotentialAIsystemsintheseGuidelines.Thisisinlinewithrecital12AIAct,whichclarifiesthatthenotionofan‘AIsystem’shouldbeclearlydefinedwhileproviding‘theflexibilitytoaccommodatetherapidtechnologicaldevelopmentsinthisfield’.ThedefinitionofanAIsystemshouldnotbeappliedmechanically;eachsystemmustbeassessedbasedonitsspecificcharacteristics.

(7)TheGuidelinesarenotbinding.AnyauthoritativeinterpretationoftheAIActmayultimatelyonlybegivenbytheCourtofJusticeoftheEuropeanUnion(CJEU).

II.ObjectiveandmainelementsoftheAIsystemdefinition

(8)Article3(1)oftheAIActdefinesanAIsystemasfollows:

1Regulation(EU)2024/1689.

2Article1AIAct.

3Article113,thirdparagraph,point(a).

2

“‘AIsystem’meansamachine-basedsystemthatisdesignedtooperatewithvaryinglevelsofautonomyandthatmayexhibitadaptivenessafterdeployment,andthat,forexplicitorimplicitobjectives,infers,fromtheinputitreceives,howtogenerateoutputssuchaspredictions,content,recommendations,ordecisionsthatcaninfluencephysicalorvirtualenvironments;”

(9)Thatdefinitioncomprisessevenmainelements:(1)amachine-basedsystem;(2)thatisdesignedtooperatewithvaryinglevelsofautonomy;(3)thatmayexhibitadaptivenessafterdeployment;(4)andthat,forexplicitorimplicitobjectives;(5)infers,fromtheinputitreceives,howtogenerateoutputs(6)suchaspredictions,content,recommendations,ordecisions(7)thatcaninfluencephysicalorvirtualenvironments.

(10)ThedefinitionofanAIsystemadoptsalifecycle-basedperspectiveencompassingtwomainphases:thepre-deploymentor‘building’phaseofthesystemandthepost-deploymentor‘use’phaseofthesystem4.Thesevenelementssetoutinthatdefinitionarenotrequiredtobepresentcontinuouslythroughoutbothphasesofthatlifecycle.Instead,thedefinitionacknowledgesthatspecificelementsmayappearatonephase,butmaynotpersistacrossbothphases.ThisapproachtodefineanAIsystemreflectsthecomplexityanddiversityofAIsystems,ensuringthatthedefinitionalignswiththeAIAct'sobjectivesbyaccommodatingawiderangeofAIsystems.

1.Machine-basedsystem

(11)Theterm‘machine-based’referstothefactthatAIsystemsaredevelopedwithandrunonmachines.Theterm‘machine’canbeunderstoodtoincludeboththehardwareandsoftwarecomponentsthatenabletheAIsystemtofunction.Thehardwarecomponentsrefertothephysicalelementsofthemachine,suchasprocessingunits,memory,storagedevices,networkingunits,andinput/outputinterfaces,whichprovidetheinfrastructureforcomputation.Thesoftwarecomponentsencompasscomputercode,instructions,programs,operatingsystems,andapplicationsthathandlehowthehardwareprocessesdataandperformstasks.

(12)AllAIsystemsaremachine-based,sincetheyrequiremachinestoenabletheirfunctioning,suchasmodeltraining,dataprocessing,predictivemodellingandlarge-scaleautomateddecisionmaking.TheentirelifecycleofadvancedAIsystemsreliesonmachinesthatcanincludemanyhardwareorsoftwarecomponents.Theelementof‘machine-based’inthedefinitionofAIsystemunderlinesthefactthatAIsystemsmustbecomputationallydrivenandbasedonmachineoperations.

(13)Theterm‘machine-based’coversawidevarietyofcomputationalsystems.Forexample,thecurrentlymostadvancedemergingquantumcomputingsystems,whichrepresentasignificantdeparturefromtraditionalcomputingsystems,constitutemachine-basedsystems,despitetheiruniqueoperationalprincipesanduseofquantum-mechanical

4ForoverviewoftheAIsystemphasesseetheOECD(2024),“ExplanatorymemorandumontheupdatedOECDdefinitionofanAIsystem”,OECDArtificialIntelligencePapers,No.8,OECDPublishing,Paris,

/10.1787/623da898-en,

p.7.

3

phenomena,asdobiologicalororganicsystemssolongastheyprovidecomputationalcapacity.

2.Autonomy

(14)Thesecondelementofthedefinitionreferstothesystembeing‘designedtooperatewithvaryinglevelsofautonomy’.Recital12oftheAIActclarifiesthattheterms‘varyinglevelsofautonomy’meanthatAIsystemsaredesignedtooperatewith‘somedegreeofindependenceofactionsfromhumaninvolvementandofcapabilitiestooperatewithouthumanintervention’.

(15)Thenotionsofautonomyandinferencegohandinhand:theinferencecapacityofanAIsystem(i.e.,itscapacitytogenerateoutputssuchaspredictions,content,recommendations,ordecisionsthatcaninfluencephysicalorvirtualenvironments)iskeytobringaboutitsautonomy.

(16)Centraltotheconceptofautonomyis‘humaninvolvement’and‘humanintervention’andthushuman-machineinteraction.Atoneextremeofpossiblehuman-machineinteractionaresystemswhicharedesignedtoperformalltasksthoughmanuallyoperatedfunctions.Attheotherextremearesystemsthatarecapabletooperatewithoutanyhumaninvolvementorintervention,i.e.fullyautonomously.

(17)Thereferenceto‘somedegreeofindependenceofaction’inrecital12AIActexcludessystemsthataredesignedtooperatesolelywithfullmanualhumaninvolvementandintervention.Humaninvolvementandhumaninterventioncanbeeitherdirect,e.g.throughmanualcontrols,orindirect,e.g.thoughautomatedsystems-basedcontrolswhichallowhumanstodelegateorsupervisesystemoperations.

(18)Forexample,asystemthatrequiresmanuallyprovidedinputstogenerateanoutputbyitselfisasystemwith‘somedegreeofindependenceofaction’,becausethesystemisdesignedwiththecapabilitytogenerateanoutputwithoutthisoutputbeingmanuallycontrolled,orexplicitlyandexactlyspecifiedbyahuman.Likewise,anexpertsystemfollowingadelegationofprocessautomationbyhumansthatiscapable,basedoninputprovidedbyahuman,toproduceanoutputonitsownsuchasarecommendationisasystemwith‘somedegreeofindependenceofaction’.

(19)ThereferenceinthedefinitionofanAIsysteminArticle3(1)AIActto‘machine-basedsystemthatisdesignedtooperatewiththevaryinglevelsofautonomy’underlinestheabilityofthesystemtointeractwithitsexternalenvironment,ratherthanachoiceofaspecifictechnique,suchasmachinelearning,ormodelarchitectureforthedevelopmentofthesystem.

(20)Therefore,thelevelofautonomyisanecessaryconditiontodeterminewhetherasystemqualifiesasanAIsystem.AllsystemsthataredesignedtooperatewithsomereasonabledegreeofindependenceofactionsfulfiltheconditionofautonomyinthedefinitionofanAIsystem.

4

(21)Systemsthathavethecapabilitytooperatewithlimitedornohumaninterventioninspecificusecontexts,suchasinthehigh-riskareasidentifiedinAnnexIandAnnexIIIAIAct,may,undercertainconditions,triggeradditionalpotentialrisksandhumanoversightconsiderations.Thelevelofautonomyisanimportantconsiderationforaproviderwhendevising,forexample,thesystem’shumanoversightorriskmitigationmeasuresinthecontextoftheintendedpurposeofasystem.

3.Adaptiveness

(22)ThethirdelementofthedefinitioninArticle3(1)AIActisthatthesystem‘mayexhibitadaptivenessafterdeployment’.Theconceptsofautonomyandadaptivenessaretwodistinctbutcloselyrelatedconcepts.TheyareoftendiscussedtogetherbuttheyrepresentdifferentdimensionsofanAIsystem’sfunctionality.Recital12AIActclarifiesthat‘a(chǎn)daptiveness’referstoself-learningcapabilities,allowingthebehaviourofthesystemtochangewhileinuse.Thenewbehaviouroftheadaptedsystemmayproducedifferentresultsfromtheprevioussystemforthesameinputs.

(23)Theuseoftheterm‘may’inrelationtothiselementofthedefinitionindicatesthatasystemmay,butdoesnotnecessarilyhaveto,possessadaptivenessorself-learningcapabilitiesafterdeploymenttoconstituteanAIsystem.Accordingly,asystem’sabilitytoautomaticallylearn,discovernewpatterns,oridentifyrelationshipsinthedatabeyondwhatitwasinitiallytrainedonisafacultativeandthusnotadecisiveconditionfordeterminingwhetherthesystemqualifiesasanAIsystem.

4.AIsystemobjectives

(24)ThefourthelementofthedefinitionisAIsystemobjectives.AIsystemsaredesignedtooperateaccordingtooneormoreobjectives.Theobjectivesofthesystemmaybeexplicitlyorimplicitlydefined.Explicitobjectivesrefertoclearlystatedgoalsthataredirectlyencodedbythedeveloperintothesystem.Forexample,theymaybespecifiedastheoptimisationofsomecostfunction,aprobability,oracumulativereward.Implicitobjectivesrefertogoalsthatarenotexplicitlystatedbutmaybededucedfromthebehaviourorunderlyingassumptionsofthesystem.TheseobjectivesmayarisefromthetrainingdataorfromtheinteractionoftheAIsystemwithitsenvironment.

(25)Recital12AIActclarifiesthat,‘theobjectivesoftheAIsystemmaybedifferentfromtheintendedpurposeoftheAIsysteminaspecificcontext’.TheobjectivesofanAIsystemareinternaltothesystem,referringtothegoalsofthetaskstobeperformedandtheirresults.Forinstance,acorporatevirtualAIassistantsystemmayhaveobjectivestoansweruserquestionsonasetofdocumentswithhighaccuracyinandlowrateoffailures.Incontrast,theintendedpurposeisexternallyorientedandincludesthecontextinwhichthesystemisdesignedtobedeployedandhowitmustbeoperated.Indeed,accordingtoArticle3(12)AIAct,theintendedpurposeofanAIsystemreferstothe‘use

5

forwhichanAIsystemisintendedbytheprovider’.Forexample,inthecaseofacorporatevirtualAIassistantsystem,theintendedpurposemightbetoassistacertaindepartmentofacompanytocarryoutcertaintasks.Thismightrequirethatthedocumentsthatthevirtualassistantusescomplywithcertainrequirements(e.g.length,formatting)andthattheuserquestionsarelimitedtothedomaininwhichthesystemisintendedtooperate.Thisintendedpurposeisfulfillednotonlythroughthesystem'sinternaloperationtoachieveitsobjectives,butalsothroughotherfactors,suchastheintegrationofthesystemintoabroadercustomerserviceworkflow,thedatathatisusedbythesystem,orinstructionsforuse.

5.InferencinghowtogenerateoutputsusingAItechniques

(26)ThefifthelementofanAIsystemisthatitmustbeabletoinfer,fromtheinputitreceives,howtogenerateoutputs.Recital12AIActclarifiesthat“[a]keycharacteristicofAIsystemsistheircapabilitytoinfer.”Asfurtherexplainedinthatrecital,AIsystemsshouldbedistinguishedfrom“simplertraditionalsoftwaresystemsorprogrammingapproachesandshouldnotcoversystemsthatarebasedontherulesdefinedsolelybynaturalpersonstoautomaticallyexecuteoperations.”Thiscapabilitytoinferisthereforeakey,indispensableconditionthatdistinguishesAIsystemsfromothertypesofsystems.

(27)Recital12alsoexplainsthat‘[t]hiscapabilitytoinferreferstotheprocessofobtainingtheoutputs,suchaspredictions,content,recommendations,ordecisions,whichcaninfluencephysicalandvirtualenvironments,andtoacapabilityofAIsystemstoderivemodelsoralgorithms,orboth,frominputsordata.’Thisunderstandingoftheconceptof‘inference’doesnotcontradicttheISO/IEC22989standard,whichdefinesinference‘a(chǎn)sreasoningbywhichconclusionsarederivedfromknownpremises’andthisstandardincludesanAI-specificnotestating:‘[i]nAI,apremiseiseitherafact,arule,amodel,afeatureorrawdata.”5.

(28)The‘processofobtainingtheoutputs,suchaspredictions,content,recommendations,ordecisions,whichcaninfluencephysicalandvirtualenvironments’,referstotheabilityoftheAIsystem,predominantlyinthe‘usephase’,togenerateoutputsbasedoninputs.A‘capabilityofAIsystemstoderivemodelsoralgorithms,orboth,frominputsordata’refersprimarily,butisnotlimitedto,the‘buildingphase’ofthesystemandunderlinestherelevanceofthetechniquesusedforbuildingasystem.

(29)Theterms‘inferhowto’,usedinArticle3(1)andclarifiedinrecital12AIAct,isbroaderthan,andnotlimitedonlyto,anarrowunderstandingoftheconceptofinferenceasanabilityofasystemtoderiveoutputsfromgiveninputs,andthusinfertheresult.Accordingly,theformulationusedinArticle3(1)AIAct,i.e.‘infers,howtogenerateoutputs’,shouldbeunderstoodasreferringtothebuildingphase,wherebyasystemderivesoutputsthroughAItechniquesenablinginferencing.

5ISO/IEC22989:2022,Informationtechnology—Artificialintelligence—Artificialintelligenceconceptsandterminology.

6

5.1.AItechniquesthatenableinference

(30)FocusingspecificallyonthebuildingphaseoftheAIsystem,recital12AIActfurtherclarifiesthat‘[t]hetechniquesthatenableinferencewhilebuildinganAIsystemincludemachinelearningapproachesthatlearnfromdatahowtoachievecertainobjectives,andlogic-andknowledge-basedapproachesthatinferfromencodedknowledgeorsymbolicrepresentationofthetasktobesolved.’Thosetechniquesshouldbeunderstoodas‘AItechniques’.

(31)Thisclarificationexplicitlyunderlinesthattheconceptof‘inference’shouldbeunderstoodinabroadersenseasencompassingthe‘buildingphase’oftheAIsystem.Recital12AIActthenprovidesfurtherguidanceontechniquesthatenablethisabilityofanAIsystemtoinferhowtogenerateoutputs.Accordingly,thetechniquesthatmaybeusedtoenableinferenceinclude‘machinelearningapproachesthatlearnfromdatahowtoachievecertainobjectivesandlogic-andknowledge-basedapproachesthatinferfromencodedknowledgeorsymbolicrepresentationofthetasktobesolved.’

(32)ThefirstcategoryofAItechniquesmentionedinrecital12AIActis‘machinelearningapproachesthatlearnfromdatahowtoachievecertainobjectives’.Thatcategoryincludesalargevarietyofapproachesenablingasystemto‘learn’,suchassupervisedlearning,unsupervisedlearning,self-supervisedlearningandreinforcementlearning.

(33)Inthecaseofsupervisedlearning,theAIsystemlearnsfromannotations(labelleddata),wherebytheinputdataispairedwiththecorrectoutput.Thesystemusesthoseannotationstolearnamappingfrominputstooutputsandthengeneralisesthistonew,unseendata.AnAI-enablede-mailspamdetectionsystemisanexampleofasupervisedlearningsystem.Duringitsbuildingphase,thesystemistrainedonadatasetcontainingemailsthathumanshavelabelledas‘spam’or‘notspam’tolearnpatternsfromthefeaturesofthelabellede-mails.Oncetrainedandinuse,thesystemcananalysenewe-mailsandclassifythemasspamornotspambasedonthepatternsithaslearnedfromthelabelleddata.

(34)OtherexamplesofAIsystemsbasedonsupervisedlearningincludeimageclassificationsystemstrainedonadatasetofimages,wherebyeachimageislabelledwithasetoflabels(e.g.objectssuchascars),medicaldevicediagnosticsystemstrainedonmedicalimaginglabelledbyhumanexperts,andfrauddetectionsystemsthataretrainedonlabelledtransactiondata.

(35)Inthecaseofunsupervisedlearning,theAIsystemlearnsfromdatathathasnotbeenlabelled.Themodelistrainedondatawithoutanypredefinedlabelsoroutputs.Usingdifferenttechniques,suchasclustering,dimensionalityreduction,associationrulelearning,anomalitydetection,orgenerativemodels,thesystemistrainedtofindpatters,structuresorrelationshipsinthedatawithoutexplicitguidanceonwhattheoutcomeshouldbe.AIsystemsusedfordrugdiscoverybypharmaceuticalcompaniesisan

7

exampleofunsupervisedlearning.AIsystemsuseunsupervisedlearning(e.g.clusteringoranomalitydetection)togroupchemicalcompoundsandpredictpotentialnewtreatmentsfordiseasesbasedontheirsimilaritiestoexistingdrugs.

(36)Self-supervisedlearningisasubcategoryofunsupervisedlearning,wherebytheAIsystemlearnsfromunlabelleddatainasupervisedfashion,usingthedataitselftocreateitsownlabelsorobjectives.AIsystemsbasedonself-supervisedlearningusevarioustechniques,suchasauto-encoders,generativeadversarialnetworks,orcontrastivelearning.AnimagerecognitionsystemthatlearnstorecogniseobjectsbypredictingmissingpixelsinanimageisanexampleofanAIsystembasedonself-supervisedlearning.Otherexamplesincludelanguagemodelsthatlearntopredictthenexttokeninasentenceorspeechrecognitionsystemsthatlearntorecognisespokenwordsbypredictingthenextacousticfeatureinanaudiosignal.

(37)AIsystemsbasedonreinforcementlearninglearnfromdatacollectedfromtheirownexperiencethrougha‘reward’function.UnlikeAIsystemsthatlearnfromlabelleddata(supervisedlearning)orthatlearnfrompatterns(unsupervisedlearning),AIsystemsbasedonreinforcementlearninglearnfromexperience.Thesystemisnotgivenexplicitlabelsbutinsteadlearnsbytrialanderror,refiningitsstrategybasedonthefeedbackitgetsfromtheenvironment.AnAI-enabledrobotarmthatcanperformtaskslikegraspingobjectsisanexampleofanAIsystembasedonreinforcementlearning.Reinforcementlearningcanbealsoused,forexample,tooptimisepersonalisedcontentrecommendationsinsearchenginesandtheperformanceofautonomousvehicles.

(38)Deeplearningisasubsetofmachinelearningthatutiliseslayeredarchitectures(neuralnetworks)forrepresentationlearning.AIsystemsbasedondeeplearningcanautomaticallylearnfeaturesfromrawdata,eliminatingtheneedformanualfeatureengineering.Duetothenumberoflayersandparameters,AIsystemsbasedondeeplearningtypicallyrequirelargeamountsofdatatotrain,butcanlearntorecognisepatternsandmakepredictionswithhighaccuracywhengivensufficientdata.AIsystemsbasedondeeplearningarewidelyused,anditisatechnologybehindmanyrecentbreakthroughsinAI.

(39)Inadditiontovariousmachinelearningapproachesdiscussedabove,thesecondcategoryoftechniquesmentionedinrecital12AIActare‘logic-andknowledge-basedapproachesthatinferfromencodedknowledgeorsymbolicrepresentationofthetasktobesolved’.Insteadoflearningfromdata,theseAIsystemslearnfromknowledgeincludingrules,factsandrelationshipsencodedbyhumanexperts.Basedonthehumanexpertsencodedknowledge,thesesystemscan‘reason’viadeductiveorinductiveenginesorusingoperationssuchassorting,searching,matching,chaining.Byusinglogicalinferencetodrawconclusions,suchsystemsapplyformallogic,predefinedrulesorontologiestonewsituations.Logic-andknowledge-basedapproachesincludeforinstance,knowledgerepresentation,inductive(logic)programming,knowledgebases,inferenceanddeductiveengines,(symbolic)reasoning,expertsystemsandsearchandoptimisationmethods.Forexample,classicallanguageprocessingmodelsbasedongrammaticalknowledgeandlogicalsemanticsrelyonthestructureoflanguage,

8

identifyingthesyntacticalandgrammaticalcomponentsofsentencestoextractthemeaningofagiventext.AnotherprominentexampleofAIsystemsbasedonlogicandknowledge-basedapproachesareearlygenerationexpertsystemsintendedformedicaldiagnosis,whicharedevelopedbyencodingknowledgeofarangeofmedicalexpertsandwhichareintendedtodrawconclusionsfromasetofsymptomsofagivenpatient.

5.2.SystemsoutsidethescopeoftheAIsystemdefinition

(40)Recital12alsoexplainsthattheAIsystemdefinitionshoulddistinguishAIsystemsfrom“simplertraditionalsoftwaresystemsorprogrammingapproachesandshouldnotcoversystemsthatarebasedontherulesdefinedsolelybynaturalpersonstoautomaticallyexecuteoperations.”

(41)SomesystemshavethecapacitytoinferinanarrowmannerbutmayneverthelessfalloutsideofthescopeoftheAIsystemdefinitionbecauseoftheirlimitedcapacitytoanalysepatternsandadjustautonomouslytheiroutput.Suchsystemsmayinclude:

Systemsforimprovingmathematicaloptimization

(42)Systemsusedtoimprovemathematicaloptimisationortoaccelerateandapproximatetraditional,wellestablishedoptimisationmethods,suchaslinearorlogisticregressionmethods,falloutsidethescopeoftheAIsystemdefinition.Thisisbecause,whilethosemodelshavethecapacitytoinfer,theydonottranscend‘basicdataprocessing’.Anindicationthatasystemdoesnottranscendbasicdataprocessingcouldbethatithasbeenusedinconsolidatedmannerformanyyears6.Thisincludes,forexample,machinelearning-basedmodelsthatapproximatefunctionsorparametersinoptimizationproblemswhilemaintainingperformance.Thesystemsaimtoimprovetheefficiencyofoptimisationalgorithmsusedincomputationalproblems.Forexample,theyhelptospeedupoptimisationtasksbyprovidinglearnedapproximations,heuristics,orsearchstrategies.

(43)Forexample,physics-basedsystemsmayusemachinelearningtechniquestoimprovecomputationalperformance,acceleratingtraditionalphysics-basedsimulationsorestimatingparameters,thatarethenfedintotheestablishedphysicsmodels.ThesesystemswouldfalloutsidethescopeoftheAIsystemdefinition.Inthisexample,machinelearningmodelsapproximatecomplexatmosphericprocesses,suchascloudmicrophysicsorturbulence,enablingfasterandmorecomputationallyefficientforecasts.

(44)Anotherexampleofasystemthatfallsoutsidethescopeofthedefinitionisasatellitetelecommunicationsystemtooptimizebandwidthallocationandresourcemanagement.Insatellitecommunication,traditionaloptimizationmethodsmaystrugglewithreal-timedemandsofnetworktraffic,especiallywhenadjustingforvaryinglevelsofuserdemandacrossdifferentregions.Machinelearningmodels,forinstance,canbeusedtopredict

6Inanycase,thesystemsthatarealreadyplacedonthemarketorputintoservicebefore2August2026benefitfrom‘grandfathering’clauseforeseeninArticle111(2)AIAct.

9

networktrafficandoptimizetheallocationofresourceslikepowerandbandwidthtosatellitetransponders,havingsimilarperformancetoestablishedmethodsinthefield.

(45)Whilstthesesystemsmayincorporateautomaticself-adjustments,theseadjustmentsareaddressedatoptimisingthefunctioningofthesystemsbyimprovingitscomputationalperformanceratherthan,forexample,atpermittingadjustmentsoftheirdecisionmakingmodelsinanintelligentway.UndertheseconditionstheymaybeexcludedfromtheAIsystemdefinition.

Basicdataprocessing

(46)Basicdataprocessingsystemreferstoasystemthatfollowspredefined,explicitinstructionsoroperations.Thesesystemsaredevelopedanddeployedtoexecutetasksbasedonmanualinputsorrules,withoutany‘learning,reasoningormodelling’atanystageofthesystemlifecycle.Theyoperatebasedonfixedhuman-programmedrules,withoutusingAItechniques,suchasmachinelearningorlogic-basedinference,togenerateoutputs.Thesebasicdataprocessingsystemsinclude,forexample,databasemanagementsystemsusedtosortorfilterdatabasedonspecificcriteria(e.g.‘findallcustomerswhopurchasedaspecificproductinthelastmonth’),standardspreadsheetsoftwareapplicationswhichdonotincorporateAIenabledfunctionalities,andsoftwarethatcalculatesapopulationaveragefromasurveythatislaterexploitedinageneralcontext.

(47)Alsosystemsthatsolelyintendedfordescriptiveanalysis,