unbias project key findings presented at our showcase event · this is illustrated in the...

12
UnBias project key findings Presented at our showcase event 1 st October 2018 at the Digital Catapult, London

Upload: others

Post on 14-Jul-2020

3 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: UnBias project key findings Presented at our showcase event · This is illustrated in the quantitative findings of the youth jury sessions: 93.8% of youth jury participants came up

UnBiasprojectkeyfindingsPresentedatourshowcaseevent

1stOctober2018attheDigitalCatapult,London

Page 2: UnBias project key findings Presented at our showcase event · This is illustrated in the quantitative findings of the youth jury sessions: 93.8% of youth jury participants came up

UnBiasprojectkeyfindings:presentedatShowcaseworkshopOctober1st2018

2

WhatistheUnBiasprojectabout?TheUnBias project seeks to promote fairness online.We live in an age of ubiquitous online datacollection,analysisandprocessing.Newsfeeds,searchengineresultsandproductrecommendationsincreasingly use personalisation algorithms to determine the information we see when browsingonline.Whilstthiscanhelpustocutthroughthemountainsofavailableinformationandfindthosebitsthataremostrelevanttous,howcanwebesurethattheyareoperatinginourbestinterests?Arealgorithmsever‘neutral’andhowcanwejudgethetrustworthinessandfairnessofsystemsthatheavilyrelyonalgorithms?

Ourprojectinvestigatestheuserexperienceofalgorithmdriveninternetservicesandtheprocessesofalgorithmdesign.Weareparticularlyinterestedincircumstancesinwhichalgorithmicprocessesmight(intentionallyorunintentionally)producebiasedorunfairoutcomes–forinstanceintheformofhelpingfakecontenttospreadonsocialmedia,producingsearchresultsthatreinforceprejudicedattitudes,ortheexcessivepersonalisationofcontentandcollectionofpersonaldata.

Mission:Todevelopco-designedrecommendationsandmaterialsfordesign,regulationandeducationtopromote‘fairness’inalgorithmicsystems.

Wefocusinparticularontheperspectivesofdifferentstakeholdersandcarryoutactivitiesthat:1)supportuserunderstandingaboutonlineenvironments,2)raiseawarenessamongonlineprovidersabouttheconcernsandrightsofinternetusers,and3)generatedebateaboutthe‘fair’operationofalgorithmsinmodernlife.UnBiasisacollaborationbetween:-HORIZONDigitalEconomyResearchinstituteattheUniversityofNottingham-HumanCentredComputingattheUniversityofOxford-CentreforIntelligentSystemsandtheirApplications(CISA)attheUniversityofEdinburgh.WealsoworkwithcreativestudioProboscis.Theproject is fundedundertheTrust, Identity,PrivacyandSecurity (TIPS)programmeof theUK’sEngineeringandPhysicalSciencesResearchCouncil(EPSRC).ProjectreferenceEP/N02785X/1TheprojectconsistsoffourinterconnectedWorkPackages(WPs):

Page 3: UnBias project key findings Presented at our showcase event · This is illustrated in the quantitative findings of the youth jury sessions: 93.8% of youth jury participants came up

UnBiasprojectkeyfindings:presentedatShowcaseworkshopOctober1st2018

3

WP1[ledbyNottingham]uses‘YouthJuries’workshopswith13-17year-old“digitalnatives”toco-produce citizen education materials on properties of information filtering/recommendationalgorithms;

WP2 [led by Edinburgh] uses co-design workshops and Hackathons to explore challenges in thealgorithmicsystemdesignprocessrelatedtobias/fairnessofalgorithmicdecisions,e.g.trade-offsinlimitedresourceallocationtasks;

WP3[ledbyOxford]exploresusers’experiencesofalgorithmdrivenonlineplatformsviaobservationsandotherfieldworkmethods;

WP4 [co-led by Oxford and Nottingham] explores policy dimensions, including information andeducationgovernanceframeworksforrespondingtothedemandsofthechanginglandscapeintheusageofalgorithmicdecisionandrecommendationsystems.Thesearedevelopedthroughbroad stakeholder focus groups including representatives of industry, regulators, third-sectororganizations,educators,lay-peopleandyoungpeople(a.k.a.“digitalnatives”).

Projectteam

NottinghamProfessorDerekMcAuleyDrElviraPerezVallejosDrAnsgarKoeneDrLizDowthwaiteDrVirginiaPortilloDrHelenCreswickMonicaCanoOxfordProfessorMarinaJirotkaDrHelenaWebbDrMenishaPatelEdinburghProfessorMichaelRovatsosDrAlanDavoustDrSofiaCeppiProboscisGilesLane

Page 4: UnBias project key findings Presented at our showcase event · This is illustrated in the quantitative findings of the youth jury sessions: 93.8% of youth jury participants came up

UnBiasprojectkeyfindings:presentedatShowcaseworkshopOctober1st2018

4

KeyfindingsAstheUnBiasprojecthasdevelopedwehaveenjoyedagreatdealofcollaborationandcooperationacrosstheworkpackages.Wehavefoundvariouslinkagesbetweenourareasofworkandourresultsandoutputscomplementeachother.Acrosstheproject,wehaveidentified3keyfindings

1. The existence of concerns over the contemporary prevalence algorithm driven onlineplatforms

2. Adesireforchangetoimprovetheuserexperienceoftheseplatforms3. Opportunitiesforchange–wehaveidentifiedcapacitytoaddressunfairnessinalgorithmic

systemsthroughvariousmeans,inparticularviaeducation,engagementandpolicychange.The existence of concerns over the contemporary prevalence algorithm driven onlineplatformsInouryouthjurysessionsconductedwithyoungpeople,manyparticipantsrecognisedthebenefitsofusingpersonalisationandfilteringprocesses,believingthemtobeconvenientbymakingtheirinternetsearchesquicker,andusefulinshowinginformationthatwasnewandinterestingtothem.However,manyalsoexpressedconcernsthattheseprocessescanbeinaccurateandoftenmissthenuancesintheir searches. They also expressed worries about filter bubbles that limit the amount of otherinformationthattheymightseewhenonline.

Someyouthjurorsfeltthatthelevelofpersonalisationthattheywerecomfortablewithdependedonthecontext,forexamplemusicandentertainmentrecommendationswereseenasuseful,howeverthe personalisation of newsfeeds was met with concern. Some also expressed resignation andpowerlessnessthattheycoulddolittletochangeorinfluencethewaysthatthealgorithmsoperate.

Youthjuryparticipants“IwatchquitealotonNetflixatnight,andtheyrecommendyoufilmsthathavelikethesameactorinit.Andit’slikeyou’renotwatchingthefilmbecauseoftheactor.You’rewatchingthefilmbecauseofthefilm.”“Idon’treallyknowbecauseIfeellikesometimestheremightbeimportantstuffintheworldgoingonthatmightbefilteredoutformeforsomereasonwhenactuallyImightbequiteinterestedinit,soyeah.”“IfanalgorithmisworkinginawaythatifyouGoogled‘Trump’andthenitonlygaveyounegativeorpositivearticles,thenthat’swrongIthink.”“Itseemslikethey’recontrollingyouandtheyalreadyknowwhatyou’regoingtodo,soyoufollowwhattheyshowyou.”

Page 5: UnBias project key findings Presented at our showcase event · This is illustrated in the quantitative findings of the youth jury sessions: 93.8% of youth jury participants came up

UnBiasprojectkeyfindings:presentedatShowcaseworkshopOctober1st2018

5

Similar kinds of concernwere raised by the stakeholders on our panel. These professionals fromindustry,law,research,educationandNGOshighlightedthewaysinwhichthemechanismsthroughwhichindividualusersaccesscontentonlinecanbeproblematic.Forinstance,tailoredadvertisingcanhave very negative effects. They told us of the limitations of personalisation. It can be useful foradvertisingpurposesandofcoursehelpstokeepfreeaccesstoplatforms.Buttheassumptionsbeingmadeaboutuserscanbepatronisingandinaccurate.Furthermore,peoplecanbefearfulofinvasionsofprivacyandtheirvulnerabilitiesbeingtargeted.

Ourstakeholderstoldusthatlackofawarenessofalgorithmicprocessessuchassearchmechanismscancauseproblems.Forinstance,whereusersarenotawareofrankingprocessesoperatetheymayhaveafalseviewofhowfarthetopresultinasearchreflectsreality–asopposedtobeingtheresultofacalculationthatmightbeaffectedbyanumberoffactors.Similarly,usersmayhaveanoverhighfeelingoftrustinwhattheyseeonlineingeneral–expectingittobeauthoritativeandtruthful.Theactionsofplatformsinkeepingtheseprocesseslargelyopaqueandunclearwasalsohighlightedasaproblem.Aswasthelackofpluralismonline.Ouronlineexperienceisdominatedbyafewgiantcompaniesandweoftenhavelittleoptiontogoelsewhere;thiskindofenvironment,accordingtosomeofourstakeholders,allowsechochamberstothrive.Ourstakeholdersalsohighlightedtheparticularvulnerabilityofchildrenonline.Asdigitalnatives,theyhaveasocialimperativetousesocialmediabutlackprotection.Forinstance,termsandconditionsofplatforms are deliberately not written in ways they can understand and they may not yet havedevelopedthecriticalthinkingskillstoreadbetweenthelinesofwhattheyseeonlineorunderstandwhy theyarebeing shownparticularadvertsorbeingasked to inputparticular informationaboutthemselves.AdesireforchangetoimprovetheuserexperienceofalgorithmdrivenonlineplatformsOuryouthpanelparticipantswerehighlyarticulateinexpressingadesireforchange.Whilsttheydidattimesexpressresignationandfeelingsofpowerlessnesstheywerealsoclearatindicatingthatthey

Stakeholderparticipants“Soyou,yourFacebookandyourfriend’sFacebookmightbegettingverydifferentadstailoredtoyourusage,andwasveryveryinfluentialifweviewthereportfromCambridgeAnalyticaintermsofputtingourpersonalisedaddsthroughFacebookthatswayedpeople’sopinionsandthatwasseenascontent,sothatwasFakeNewsinthatsense.”“It’sverycondescendingisn’tit.Somesilicon-valleyfundamentalistdecidedtheyknowbetterwhataperson,thatandthatage,andthatandthatplaceshouldlike…whoarethey?”“Ithinkthatthemainproblemisthatconsumersexpecttheinternettotellthemthetruth.“IcanfindanythingIwantonline.AndIdoublecheckthenews,becauseIcanfindoutiftheyaretellingmetheyare not the truth, so false news, fakenews,whatever”.So they think theyare able to search forinformationproperly,butmaybetheyaren’tandtheydonotevenknowthattheycouldbebiasedintheirsearchbehaviour”.

Page 6: UnBias project key findings Presented at our showcase event · This is illustrated in the quantitative findings of the youth jury sessions: 93.8% of youth jury participants came up

UnBiasprojectkeyfindings:presentedatShowcaseworkshopOctober1st2018

6

wantedchangeandstatingwhatitcouldlooklike.Thisisillustratedinthequantitativefindingsoftheyouthjurysessions:

93.8%ofyouthjuryparticipantscameupwithsomeideasduringthesessionsabouthowtheInternetcouldbemadebetterforyoungpeople.

Additionally,they: -wantedmoreinfluenceonhowdigitaltechnologiesandservicesarerun(65.6%agree,10.3%

neutralagree) -wantedmorecontrolovertheirpersonaldataonline(84.3%agree,3.5%neutralagree) -feltthatyoungpeoplewouldbelistenedtoabouttheinternettheywant(44.8%agree,19.4%

neutralagree)However,theyfeltthattheycouldn’tcontrolwhathappenstothemandtheirpersonaldatawhentheyusedigitaltechnologies(40.7%agree,11.9%neutralagree)

Similarly,ourprofessionalstakeholderswereclearthatchange isdesirableanddescribeddifferentways itcouldbeachieved.Theypointedtochanges inareasaroundeducation,regulationandthepractices of the platform themselves. They highlighted the importance of key principles: thetransparency of algorithms, the accountability of platforms that use algorithmic systems, theauditabilityofthealgorithms.Ourstakeholderswereverykeenthatthiswasacomplexmatterandcannotberesolvedinsimpleterms.Wemayneedtomaketrade-offs:ifwewanttokeepourfreeaccesstoplatformswemighthavetogiveupsomedataandprivacy,or takeresponsibility forourownbehavioursandchoices.However,theydidalsopointoutthatwecanlearnfromothersectorsandbringinprocessesthathaveworkedelsewhere.

Youthpanelparticipants“MorecontrolisdefinitelynicelikeforexamplemyFacebook,wheneverIopenmyaccountit’slikefullofjunkIdon’twanttosee!Thisalgorithmbasicallyfailsme.Idon’twanttoseelikewhoisinloveorsomethinglikethat!Itjustfailsme.IthinktoagreatextentIagreethatweneedmore control, so actually we can control how the algorithm works like disable those filteroptions”.“Ithinkthatanissuethatneedstoberaisedinparliamentorwhateverisweshouldhavemoreknowledgeoraccessibilityofknowingwho’sgottheinformationandwhytheyhaveitandIfwecanmanagethat.”“Ithinktheyshouldhavesomethinglikeinschoolsandotherprojectsoutsideofschoolsforotherpeopleaswell,theyshouldteachthembecauseIknowinsomeschoolstheydocitizenshipbuttheretheyspeakmoreaboutmoralissuesthatmightnotnecessarilyaffecteverybody,well,acoupleinaclass,butiftheydiddosomethingonliketermsandconditionsandhowtheinternetisuseditwouldchange.”

Page 7: UnBias project key findings Presented at our showcase event · This is illustrated in the quantitative findings of the youth jury sessions: 93.8% of youth jury participants came up

UnBiasprojectkeyfindings:presentedatShowcaseworkshopOctober1st2018

7

OpportunitiesforchangeInourprojectworkweidentifiedtheexistenceofconcernsabouttheoperationofalgorithmdrivenonlineplatforms,andtheircapacitytointroduceunjustifiedbiasandproduceunfairoutcomes.Wealsoidentifiedadesireforchangetoaddressthisunfairness.Ournextstepasaprojectwastoconsideropportunitiesforpositivechangetooccur.Whatmechanismsmightmitigateunfairnessandpromotea fairer user experience online? In this final section of our key findings report we characterise 3importantthemesinrelationtopromotingpositivechange:

o anychangemustacknowledgeandaddressthecomplexityoftheissuesinvolvedo valueofeducationandengagemento specificUnBiasoutputsandrecommendations

Anychangemustacknowledgeandaddressthecomplexityoftheissuesinvolvedinalgorithmicbias,‘fairness’andregulation.Weneedtorecognise(asidentifiedinourWorkpackage3userexperienceobservations) that users employ their own agency and sense making when browsing online.Algorithmsdoshapetheuserbrowsingexperiencebutthisisalsofilteredthroughusers’ownpracticalreasoningandisnotstraightforwardlydeterministic.Methodologicallyitisalsoverydifficulttotraceexactlyhowusersareaffectedbyalgorithmicprocessesandwhatconsequencesthishasforindividualandgroupbehaviours.Furthermore,wealsoneedtorecognisethatthe‘problems’ofalgorithmicbiasandunfairnesshavemultiplerootsandcauses,andthattherearecompetingperspectivesoverhowtheycanberesolved.Itmight be that we need to think on a case by case basis andmake trade-offs to findworkablesolutions.Asourparticipantsremindus,thesequestionsaregroundedinwidersocietalissues.

Stakeholderparticipants“I think having some sort of regulatory framework and some tests that all algorithmsmust besubmitted towhere…thealgorithmcanbeassessed to seewhether it representsa fairdecisionbasedonthedata”.“…butasa societyweneedtogetready,weneedtounderstand,weneedtoappreciatethat inordertogetstuffforourpersonalisationanddiversitywemightneedtopayforit.”“Socasinos,whethertheyareonlineorintherealworldareobligedtomonitortheircustomerstoseeiftheyexhibitthetraitsofaddictiveproblemgamblers…SothereisregulationonthatandthatmightbeamodelthatcouldpotentiallybeappliedtosomethinglikeFacebook”.“Soonepossiblesolutionwastochangethelegalframeworkandtothinkofplatformsunderthepublic servicebroadcastmodel inwhichtheyhavetotakeresponsibilityfor someofthecontentthat is broadcast on their platforms. Or we could have things such as a kite mark for trustedcontent…”

Page 8: UnBias project key findings Presented at our showcase event · This is illustrated in the quantitative findings of the youth jury sessions: 93.8% of youth jury participants came up

UnBiasprojectkeyfindings:presentedatShowcaseworkshopOctober1st2018

8

Furthermore,theyremindusthatwhilsttheaimoftheUnBiasprojectistopromotefairnessonline,identifyingexactlywhat fairness is, is a complex taskWe conducted two roundsof a survey1 thatpresentedparticipantswithascenarioabouttheallocationof limitedresourceswithinaparticularcontext.Wealsogavethemasetof5algorithmsthatcoulddistributethelimitedresourcesandaskedthemtoselecttheirpreferredandleastpreferredone.Wewantedtotestoutwhetherwepresentalgorithmstolayaudiencesinawaythattheywouldfindmeaningfulandbeabletomakeaninformedchoiceabout.Wewereabletoachievethis.Ourparticipantsengagedenthusiasticallywiththetaskwegavethemanddisplayed an understanding of the how the algorithmswould operate in the context and theimplicationstheymighthave.Wealsowantedtoseewhetherourparticipantswouldselectthesamealgorithms as most or least preferred: there was no consensus and selections of most and leastpreferredwerespreadacrossall5algorithms.Allofourparticipantsspokeofwantingtoselectanalgorithm thatwas fairbutdrewondifferentunderstandingsofwhatwas fairwhenmaking theirselections–sometimesitwasaboutgivingeveryonemoreorlessthesame,sometimesitwasaboutmakingthemostpeopleaspossiblehappy,sometimes itwasaboutfindingwaystomeasurewhatpeopledeservedandsoon.Itwasmucheasierforourparticipantstoarticulatewhatwasnotfairthanwhatwasfair.Theirselectionswerealsohighlycontextdependentwithchangesinselectionbeingmade when the scenario changed slightly. Some of our results suggested that participants’interpretationsofwhatwasfairinacontexthadsystematicconnectionstotheculturalbackgroundtheparticipantswerein–thisissomethingwearekeentoassessfurther.Education and engagement are valuable mechanisms to prompt and promote change. This issomethingwehavebeenobservingandpractisingacrosstheproject.Wehavedevelopededucationalmaterialsincludingavideoanimationtohelpusersunderstandalgorithmicprocessesandtheeffectsofpersonaldatacollectionetc.Thesehaveoftenbeencocreatedwithendusers–inparticularyoungpeople–todrawontheirunderstandingsandperspectives,andmeettheirneeds.Wehaverunactivitieswithuniversitystudents–suchashackathoncompetitionsandgroupprojects–thatchallengedthemto(re)designfairsocialnetworksandbuildsystemsthatcanempowerusersonline.Wehave takenpart invariouspublicengagementactivitiesandcreateda setof toolsandactivitiesthathelponlineusersofallagestoincreasetheirunderstandingandreflectontheirown

1 Round2ofthesurveycanbeseenhere:https://oxford.onlinesurveys.ac.uk/unbias-preference-survey

Stakeholderparticipants“Wewantouralgorithmsinasensetofollowhighervalues,moralvaluesthatwethinkaremore important than giving an exact reflection of the world…algorithms are inherentlypoliticised.”“Wellifyouwanttohaveadiscussionaboutfairness,youhavetogointoacausalanalysistotryandworkoutyourcontextandwhatactuallyisimportant.Youcannothaveaonesizefitsallrule.”

Page 9: UnBias project key findings Presented at our showcase event · This is illustrated in the quantitative findings of the youth jury sessions: 93.8% of youth jury participants came up

UnBiasprojectkeyfindings:presentedatShowcaseworkshopOctober1st2018

9

perspectivesregardingalgorithmicbiasandfairness.Moredetailsaboutallofourengagementandeducationmaterialsareonourprojectwebsite:http://unbias.wp.horizon.ac.uk/Theothergreatbenefitofourengagementactivitiesintheprojectisthattheyprovideaground-upapproach to identifying specific opportunities for change. We have been able to elicit from ourparticipants the forms of improvement theywould like to see – for instance in relation tomoreaccessibleplatform termsand conditions, the incorporationof regulatorymechanisms fromotherindustries, and the requirement formeaningful transparency rather than ‘just’making algorithmsvisible.Bringingtogetherstakeholdersfromdifferentsectorshasbeenparticularlyusefulinbuildingupanuancedperspectiveonwhatchangesmightbepossibleandfeasible.Theseperspectivesarethenincorporatedintoourfurtherprojectoutputs.Wehaveproducedawiderangeofprojectoutputs,including:

- Contributionstosetsofpolicyrecommendations:EuropeanParliamentScienceTechnologyOptionsAssessment;UKICOAgeAppropriateDesignCode;

- Contributions toStandards:Chairingdevelopmentof IEEEP7003Standard forAlgorithmicBiasConsiderations;Participation(throughBSI)inworkonStandardsforArtificialIntelligencebyISO/IEC

- 24+Publicengagementengagementevents- Education materials: Online Educational Resources; Fairness Toolkit; Video animation;

Outreachsessionsandeventswithschools.- Contributionto7Governmentenquiries: Impactofsocialmediaandscreenuseonyoung

people’s health; House of Lords Artificial Intelligence committee; Algorithms in decision-making;FakeNews;TheInternet:toregulateofnottoregulate?;ChildrenandtheInternet;RoleandprioritiesofUKRIinterimchair

- Academicpublicationsinjournalsrelatingtothesocialsciences,ethicsincomputing,humanfactorsincomputing,socialmedia,AI

- Media engagement: television and radio interviews, podcasts, online articles in TheConversation,mentionsinprintmedia

- 45+Workshopsandconferences- 8Collaborationsandfollowonactivities:5Rights(ICOCode);DefendDigitalMe(evaluationof

educational apps); HumanRightsNetwork/RightToPrivacy.org (Telling Tales of Engagementvideos); ThirdWolrdNetwork (eCommerce trade negotiations – algorithm briefings); ISOC-England(UserTrust);IEEEGlobalInitiativeonEthicsofAutonomousandIntelligentSystems(P7003 standards + ad hoc presentations); RevealingReality (ICO AgeAppropriateDesignCode);ImpactExplorationGrant(UnBiasFairnessToolkit-UsageandImpact);

More details about all of these outputs can be found on our project websitehttp://unbias.wp.horizon.ac.uk/ or by contacting [email protected] [email protected]

Page 10: UnBias project key findings Presented at our showcase event · This is illustrated in the quantitative findings of the youth jury sessions: 93.8% of youth jury participants came up

UnBiasprojectkeyfindings:presentedatShowcaseworkshopOctober1st2018

10

InthefinalpartofthisreportwehighlighttwoparticularUnBiasoutputs.

1) AgovernanceframeworkforalgorithmicaccountabilityandtransparencyAs part of our policy and stakeholder engagement work, members of the UnBias team are co-authoring2aScienceTechnologyOptionsandAssessmentreportonalgorithmicaccountabilityandtransparencyfortheEuropeanParliament. Inthisreportweprovideanoverviewofpossible legal,technical and operational frameworks for governing algorithmic accountability and transparency.Beyond providing an up-to-date and systematic review of current and potential governanceframeworks for algorithmic accountability and transparency, the report aims to providerecommendationsforanalgorithmicimpactassessmentmetricforassessingthedegreeofregulatoryscrutinyofanalgorithmicsystemthatwouldbeappropriate/necessaryforaparticularcontextofuse.Thereportisstillbeingfinalisedbutourrecommendationswillinclude:Awarenessraising:education,watchdogsandwhistleblowers

- The provision of “algorithmic literacy” that teaches core concepts such as: computationalthinking,theroleofdataandtheimportanceofoptimisationcriteria.

- Theintroductionofstandardisednotificationpracticestocommunicatethetypeanddegreeofalgorithmicprocessinginvolvedindecisions.

- Theprovisionofcomputationalinfrastructureandaccesstotechnicalexpertstosupportthedata analysis and algorithm reverse engineering efforts of “algorithmic accountabilityjournalists”.

- Whistleblower protection (expanding the current EC proposal to include any violation ofHuman Rights) and protection against prosecution on grounds of breaching copyright orTermsofServicewhendoingsoservedthepublicinterest.

Accountabilityinpublicsectoruseofalgorithmicdecisionmaking- MandatingofAlgorithmicImpactAssessmentsaspartofpublicsectoruseandprocurement

ofalgorithmicsystems,involving:o Publicationofpublicauthority’sdefinitionof“algorithmicsystem”.o Publicdisclosureofpurpose,scope,intendeduseandassociatedpolicies/practices,

self-assessment timeline/process and potential implementation timeline of thesystem.

2FulllistofauthorsAnsgarKoene, UniversityofNottinghamNozhaBoujemaa, InriaChrisClifton, PurdueUniversityYohkoHatada, EMLSRIJacobLaViolette UniversityofOxfordCaioMachado UniversityofOxfordMenishaPatel, UniversityofOxfordRashidaRichardsonAINowInstituteDillonReisman AINowInstituteHelenaWebb UniversityofOxfordDavidWeinberger HarvardUniversity

Page 11: UnBias project key findings Presented at our showcase event · This is illustrated in the quantitative findings of the youth jury sessions: 93.8% of youth jury participants came up

UnBiasprojectkeyfindings:presentedatShowcaseworkshopOctober1st2018

11

o Performing and publishing of self-assessment of the system with focus oninaccuracies,bias,harmstoaffectedcommunities,anddescribesmitigationplansforpotentialimpacts.

o Publicationofplanformeaningful,ongoingaccesstoexternalresearcherstoreviewthesystemonceitisdeployed.

o Publicparticipationperiodo Publication of final Algorithmic Impact Assessment, once issues raise in public

participationhavebeenaddressed.o RenewalofAIAsonaregulartimelineo Opportunity for public to challenge failure to mitigate issues raised in the public

participationperiodorforeseeableoutcomes.

RegulatoryoversightandLegalliability- Thecreationofaregulatoryagencyforalgorithmicdecisionmakingtaskedwith:

o Establishingariskassessmentmatrixforclassifyingalgorithmtypesandapplicationdomainsaccordingtopotentialforsignificantnegativeimpactoncitizens.

o Investigatingtheuseofalgorithmicsystemswherethereisasuspicion(e.g.evidenceprovidedbyawhistleblower)ofinfringementofHumanRights.

o Advisingotherregulatoryagenciesregardingalgorithmicsystemsastheyapplytotheremitofthoseagencies.

- Systemsclassifiedascausingpotentiallyseverenon-reversibleimpactarerequiredtoproduceanAlgorithmicImpactAssessment,similartopublicsectorapplications.

- Systemswithmediumseveritynon-reversibleimpactrequiretheserviceprovidertoacceptstricttortliability,withapossibilityofreducingtheliabilitybyhavingthesystemcertifiedascompliantwith(asyettobedetermined)impactmitigatingstandards.

Globaldimensionofalgorithmicgovernance- Astrongpositionintradenegotiationstoprotectregulatoryabilitytoinvestigatealgorithmic

systemsandholdpartiesaccountableforviolationsofEuropeanlawsandHumanRights.- Establishingapermanentforumformulti-stakeholderdialogontheethicaldevelopmentof

algorithmicsystems,basedontheprinciplesofResponsibleResearchandInnovation.

FormoredetailsabouttheupcomingEPSTOAreport,[email protected]

2FairnessToolkit

Page 12: UnBias project key findings Presented at our showcase event · This is illustrated in the quantitative findings of the youth jury sessions: 93.8% of youth jury participants came up

UnBiasprojectkeyfindings:presentedatShowcaseworkshopOctober1st2018

12

OurFairnessToolkitwhichaimstopromoteawarenessandstimulateapubliccivicdialogueabouthowalgorithmsshapeonlineexperiencesandtoreflectonpossiblechangestoaddressissuesofonlineunfairness. The tools are not just for critical thinking, but for civic thinking – supporting a morecollectiveapproachto imaginingthe future incontrast tothe individualatomisingeffect thatsuchtechnologiesoftencause.Thetoolkithasbeenco-createdwithyoungpeopleandstakeholdersandconsistsofthreemainparts:1.AwarenessCards–adeckofcardsdesignedtohelpyoungpeopledeviseandexplorescenariosthatillustratehowbiasinalgorithmicsystemscanaffectthem.Thecardsarea“peertopeer”tool,enablingyoungpeopletocollaborativelyexploretheissuesofdataprivacyandprotection,onlinesafetyandsocial justice to create compelling stories and scenarios that help communicate and developawareness.Thecardsaredesignedtobeusedinbothfacilitatedenvironments,suchasschoolsandyouthgroups,butalsoasacivicthinkingtoolforanyonetoinvestigatetherelationshipsbetweendata,rights,values,processesandthefactorswhichaffect theiroperationforusas individualsandasasociety.2.TrustScapes–aposterforyoungpeopletovisualisetheirperceptionstheissuesofalgorithmicbias,dataprotectionandonlinesafetyandwhattheywouldliketoseedonetomaketheonlineworldfairandtrustworthy.Designedtocaptureboththeirfeelingsaboutthecurrentsituationandtheirdreamsandidealsforwhattheinternetcouldorshouldbeinadynamicandvisualway.TheTrustScapesformthefirstelementinthepubliccivicdialoguethatUnBiaswillinitiate:imagesoftheTrustScapeswillbesharedviaUnBiassocialmediaaccountstoarticulateyoungpeople’svisionsforthefutureinternetandamplifytheirvoiceinthedebateontrustandfairness.3.MetaMaps–aposterforstakeholdersintheICTindustry,policymaking,regulation,publicsectorand research to respond to the young people’s TrustScapes. By selecting and incorporating aTrustScapefromthosesharedonline,stakeholderscanrespondtotheyoungpeople’sperceptions.MetaMapswillalsobecapturedandsharedonlineviaUnBiassocialmediatoenhancethepubliccivicdialogue,anddemonstratethevalueofparticipationtoyoungpeopleinhavingtheirvoicelistenedandrepliedto.TheFairnessToolkitalsoincludes“valueperception”worksheetstohelpparticipantsandstakeholderscriticallyassessandevaluatethevalueofparticipation.FormoreinformationabouttheToolkit,[email protected]