歐盟EDPB《算法分?jǐn)?shù)提案（英文版）》

SUPPORTPOOL

OFEXPERTSPROGRAMME

AIAuditing

ProposalforAlgo-scores

byDr.GemmaGALDONCLAVELL

AIAuditing-ProposalforAlgo-scores

2

AspartoftheSPEprogramme,theEDPBmaycommissioncontractorstoprovidereportsandtoolsonspecifictopics.

TheviewsexpressedinthedeliverablesarethoseoftheirauthorsandtheydonotnecessarilyreflecttheofficialpositionoftheEDPB.TheEDPBdoesnotguaranteetheaccuracyoftheinformation

includedinthedeliverables.NeithertheEDPBnoranypersonactingontheEDPB’sbehalfmaybeheldresponsibleforanyusethatmaybemadeoftheinformationcontainedinthedeliverables.

Someexcerptsmayberedactedorremovedfromthedeliverablesastheirpublicationwould

underminetheprotectionoflegitimateinterests,including,interalia,theprivacyandintegrityofanindividualregardingtheprotectionofpersonaldatainaccordancewithRegulation(EU)2018/1725and/orthecommercialinterestsofanaturalorlegalperson.

AIAuditing-ProposalforAlgo-scores

3

TableofContents

1.Introduction 4

1.1.Nutriscore 4

1.2.A+++ 4

2.TowardsanAIalgo-score 5

2.1.Existinginitiatives 5

2.2.Proposalforanalgo-score 6

DocumentinitiallysubmittedinJanuary2023,updatedinJune2024

AIAuditing-ProposalforAlgo-scores

4

1.Introduction

AsAIsystemsproliferateandimpactonpeoples’life-chances,thereisaneedtofacilitatepublicawarenessandunderstandingofhowalgorithmicsystemswork.Theproposalbelowdrawsfromtwoexistingmethodologiesofsignificantsocietalimpact:theNutriscoreandA+++.

1.1.Nutriscore

NutriscorereliesonthecomputationofanutrientprofilingsystemderivedfromtheUnitedKingdomFoodStandardsAgencynutrientprofilingsystem(FSAscore).Thecalculationprocessconsistsofthreesteps.Inthefirststep,thenutritionalscoreofthefoodproductisassessed.Thenextstepis

thedeterminationofRayner'sscore,calculatedinthesamewayforallfoodproducts,withtheexceptionofcheese,vegetableandanimalfatsandoils,anddrinks.Thetwoscoresarethenusedtoclassifythefoodproductonthefive-levelNutri-Scorescale.

totalNscore-totalPscore=Nutritionalscore

Productingredientsnegatively(N)affectingtheNutri-Score

Productingredientspositively(P)affectingtheNutri-Score

Highenergydensityper100gorper100ml

Highsugarcontent

Highcontentofsaturatedfattyacids

Highsaltcontent

Contentoffruits,vegetables,nutsandlegumes

Fibercontent

Proteincontent

Rapeseed,walnutandoliveoilcontent

Basedonthisclassification,thesystemawards0to10pointsforenergyvalueandingredientsthatshouldbelimitedinthediet(saturatedfattyacids,sugarandsalt,etc.)and0to5pointsforbeneficialingredientswhoseconsumptionshouldbepromoted(fiber,protein,fruits,vegetables,legumes,nuts,andrapeseedoil).Todeterminethevalueofthelabelofagivenproduct,i.e.theletterA,B,C,DorE,thesumofpointsawardedforthebeneficialingredientsmustbesubtractedfromthesumofpointsgivenfortheingredientsthatneedtobelimited.Theproductisclassifiedinoneoffivevalueclasses(AtoE)basedonthefinalscore,whichmayvaryfrom-15to+40.Thelowerthescore,thebetterthenutritionalvalueoftheproduct.

1.2.A+++

In1992,theCounciloftheEuropeanCommunitiespublishedEUDirective92/75/EC“toenabletheharmonizationofnationalmeasuresonthepublication,particularlybymeansoflabellingandofproductinformation,ofinformationontheconsumptionofenergyandofotheressentialresources,

andadditionalinformationconcerningcertaintypesofhouseholdappliances,therebyallowingconsumerstochoosemoreenergy-efficientappliances.”Since2010,anewtypeoflabelexiststhatmakesuseofpictogramsratherthanwords,toallowmanufacturerstouseasinglelabelforproductssoldindifferentcountries.Updatedlabellingrequirementsenteredintoforcein2021(EU221/340).

Theenergylabelsareseparatedintoatleastfourcategories:1)Theappliance'sdetails:accordingtoeachappliance,specificdetails,ofthemodelanditsmaterials,2)Energyclass:acolorcodeassociatedtoaletter(fromAtoG)thatgivesanideaoftheappliance'selectricalconsumption,3)Consumption,efficiency,capacity,etc.:thissectiongivesinformationaccordingtoappliancetype,and4)Noise:thenoiseemittedbytheapplianceisdescribedindecibels.

AIAuditing-ProposalforAlgo-scores

5

Forinstance,lampsareclassifiedintoclassAif≤0.240.√Φ+0.0103.Φ,whereΦistheluminousfluxinlmandPisthepowerconsumptionofthelampinW.Calculationsandvaluesaredifferentforhomeappliancessuchasovens,fridgesandairconditioners,bulbs,carsortyres.

2.TowardsanAIalgo-score

Theaboveexamplesshowhowusingvisualcuestoinformcitizens/buyersoftheirchoicesisamethodthathasbeentriedandtestedinotherdomains,wheretheregulatorhasplayedasignificantroleinestablishingthemethodologyandsettingupfollow-upboardsthatcontinuouslyreviewthecalculationalgorithmsusedandproposeupdatestoit.

2.1.Existinginitiatives

2.1.1.AIethicslabel

1

ThepossibilityofanAlgo-scorebeingdevelopedtoguidecitizensandbuyersofAIsystemshasalreadybeenexplored.In2019,theAIEthicsImpactGroupatBertelsmannStiftungproposedanAIethicslabel.Theyreviewed100AIethicsguidelinesandproposedsixkeyvaluestoberated:transparency,accountability,privacy,justice,reliabilityandenvironmentalsustainability.ThescoringisobtainedfromtheVCIOmodel(values,criteria,indicatorsandobservables)thatisusedtodefinetherequirementsneededtoachieveacertainrating.

ThelabelincludesoneratingforeachofthevaluescapturedwiththeVCIO-approach.Inordertodeterminethelevelofcompliance(a,B,C,D,E,F,G),thesystemaggregatesobservablesonthebasisofminimumrequirements,thusmovingawayfromtheNutriscoreapproachwherebadpracticescanbeoffsetbygoodpractices.

Observablesarethenconvertedintoindicators.Thoseindicatorsareintheirturnconvertedintocriteria,andthesecriteriaarefinallyturnedintovalues,andviceversa.

2.1.2.DatasetNutritionLabel

TheDataNutritionProjecttakesinspirationfromnutritionallabelsusedinthefoodindustry.Itenhancescontext,contents,andlegibilityofdatasetsbyhighlightingthe‘ingredients’ofadatasettohelpshedlightonwhetherthedatasetis“healthy”foraparticularusecase.ItsformatisclosertoamedicalleafletthanalabelsuchastheA+++inthatitprovidesastructuredandvisualrepresentationofdatasetcharacteristics(includingdescription,date,author,content,sourceandforeseenusecases)andaseriesof“badges”(tovisualisewhetherasystemhasgonethroughethicsreview,ifithandlespersonalinformation,whetherthesystemisfor-profitornon-for-profit,etc.).Italsoincludesan“alert”scalethatisdesignedtohighlightissues,restrictions,andotherrelevantinformationaboutthedatathatmightnotbeobvioustosomeoneunfamiliarwiththedataset.Thesealerts

includeacolourcodethatshowswhetheraspecificrisk/alertcanbemitigated.

1

Fromprinciplestopractice:HowcanwemakeAIethicsmeasurable?(bertelsmann-stiftung.de)

AIAuditing-ProposalforAlgo-scores

6

2.1.3.Singlefocuslabels

Itisalsoworthmentioningtwoexamplesofsingle-focuslabels:theOpenEthicslabel,

2

whichproposesavoluntaryschemefocusedontheopennessofthetoolsandresultsofanAIsystemandsoassignsvaluetousingopenandnotproprietarytrainingdata,opensourcecode,andunrestricteddecisionspaces.Ontheotherhand,LelabelNumériqueResponsable,developedbytheInstituteforSustainableITinpartnershipwiththeFrenchMinistryofEcologicalTransition,ADEME,andtheWWF,isfocusedonissuesofsustainabilityanddigitalenvironmentalimpact.Ithasbeendesignedtobeusedbybothcommunitiesanddigitalservicesbusinessesanditisbasedon4themesand14broadprincipleswhichcoverdiversecommitmentsinareassuchasgovernance,participation,proactivity,inclusivityandsoftwaremanagement.

2.2.Proposalforanalgo-score

Theanalysisaboveshowsthepossibilitiesandlimitationsofexistinglabels,andpointtotheissuesthatneedtobetakenintoaccounttodevelopalabelforAI-relatedtechnicalsystems.

AscoringmethodologybasedontheNutriscoremodelwouldnotworkwhenrankingtechnologicaldevelopmentsasitimpliesthatbadpracticescanbe“compensated”bygoodpractices.Thiswouldprobablybeunacceptableformanystakeholders.AsystembasedontheA+++modelisalsonotanimmediatelyusefulreferenceasitisbasedonsharedandagreedonvaluesofimpact.Namely,electricityconsumptionandnoiseindicators.

AsforthemethodsdevelopedspecificallyforAIsystems,whichhavenotyetbeenimplemented,theyposeseveralchallenges.TheDatasetNutritionLabel,ontheonehand,ismoreofatransparencytoolthanamethodtocomparebetweensystems.Asitoffersmultiplepossibilitiesforgraphicallyrepresentingtechnicalsystems,andasthosesystemsmaybeverydiverse,itishighlylikelythatsystemswouldusenon-comparableindicators.Thisdoesnotmeanthatthemethodproposeddoesnothavevalue,butitsusefulnessdeviatesfromasystemusedtofacilitateconsumerchoice.

TheAIethicslabelisbyfarthemostdevelopedmethodtoassesscompliance.However,itdoesnotlooklikeithasmanagedtogettractionsinceitwasreleasedin2019.Someofitsshortcomingsareprobablylinkedtotheincentivestoadoptsuchamodel,butalsotothefactthatitissomewhatremovedfromlegalcompliancerequirements-eventhoughthesecouldbeemphasized.

Takingthelessonsfromthesepreviousefforts,andthenameandimagerecognitionofexistingsystems,theproposalbelowbeginstosketchthemethodologythatcouldhelppushforwardapracticalproposalforavisualsystemdesignedtopromotetransparency,accountabilityandconsumerchoice.

TheproposaltakesasareferencetheA+++method,andspecifically3ofitsattributes:

-Itisimmediatelyrecognizable

-Itallowsforamainmeasureofspecificvariables,butalsoothericonsthataredeemedrelevant

-Itisamethodthatgeneratesincentivestoimproveinthealphabeticalscale

Additionally,initsuseinappliances,theA+++labelhasproventobeabletoincorporatedifferentmeasuresfordifferentappliancetypes(fromfridgestotyres).Thisopensthepossibilityfordifferenttechnicalsystemstouseslightlydifferentmethods,andthustheuseofthesamelabelandrecognizableimageforrecommendersystems,socialmediaplatforms,LLMsandbiometricssystems.

2

OpenEthicsLabel–OpenEthicsInitiative

AIAuditing-ProposalforAlgo-scores

7

Basedontheabove,weproposealabelthatcapturesthree“l(fā)evels”ofanalysis:

-AIgovernance:appointmentoftherelevantgovernancepositions(DPO,AIofficer,etc),compliancewithrelevantstandards(ISO/IEC42001,23894,23053,24027,etc),documentationofkeyriskassessmentdocuments(DPIA,conformityassessment,environmentalimpactassessment,etc.),transparencydocuments(AIleaflet,etc.),dataprovenancecertifications,andlinkstoopendata/sources.

-Modelfairnessandperformance:DatasheetsforDatasets,Modelcard/s,keymetrics(fairness,robustness,security,accuracy…,etc),keyattributes,experimentaltesting(redteaming,etc.)

-Post-marketmonitoringandauditing:regularbiasaudits,incidentreporting,effectiverecourseandredressmechanismsandmetrics

AnissuethatisnoteasytoexportformtheA+++methodologytoAIistheuseofanagreed-uponmeasureofcalculationsuchaswatts.Therefore,theproposalisforthesystemtoassesstheexistence/complianceofdifferentexercisesandsupportingdocumentation.Havingcompletedspecifictaskswouldallowthosesubjectedtothelabelassessmentt