人工智能與教學的未來 Artificial Intelligence and the Future of Teaching and Learning -Insights and Recommendations -美國教育部教育技術辦公室-2023.5

1

ArtificialIntelligenceandtheFutureofTeachingandLearning

InsightsandRecommendations

May2023

ArtificialIntelligenceandtheFutureofTeachingandLearning

MiguelA.Cardona,Ed.D.

Secretary,U.S.DepartmentofEducation

RobertoJ.Rodríguez

AssistantSecretary,OfficeofPlanning,Evaluation,andPolicyDevelopment

KristinaIshmael

DeputyDirector,OfficeofEducationalTechnology

May2023

ExamplesAreNotEndorsements

Thisdocumentcontainsexamplesandresourcematerialsthatareprovidedfortheuser’sconvenience.Theinclusionofanymaterialisnotintendedtoreflectitsimportancenorisitintendedtoendorseanyviewsexpressedorproductsorservicesoffered.Thesematerialsmaycontaintheviewsandrecommendationsofvarioussubjectmatterexpertsaswellashypertextlinks,contactaddresses,andwebsitestoinformationcreatedandmaintainedbyotherpublicandprivateorganizations.TheopinionsexpressedinanyofthesematerialsdonotnecessarilyreflectthepositionsorpoliciesoftheU.S.DepartmentofEducation.TheU.S.DepartmentofEducationdoesnotcontrolorguaranteetheaccuracy,relevance,timeliness,orcompletenessofanyinformationfromothersourcesthatareincludedinthesematerials.Otherthanstatutoryandregulatoryrequirementsincludedinthedocument,thecontentsofthisguidancedonothavetheforceandeffectoflawandarenotmeanttobindthepublic.

ContractsandProcurement

Thisdocumentisnotintendedtoprovidelegaladviceorapprovalofanypotentialfederalcontractor’sbusinessdecisionorstrategyinrelationtoanycurrentorfuturefederalprocurementand/orcontract.Further,thisdocumentisnotaninvitationforbid,requestforproposal,orothersolicitation.

LicensingandAvailability

ThisreportisinthepublicdomainandavailableontheU.S.DepartmentofEducation’s

(Department’s)websiteat

.

RequestsforalternateformatdocumentssuchasBrailleorlargeprintshouldbesubmittedtotheAlternateFormatCenterbycalling1-202-260-0852orbycontactingthe504coordinatorvia

email

atom_eeos@.

NoticetoLimitedEnglishProficientPersons

IfyouhavedifficultyunderstandingEnglish,youmayrequestlanguageassistanceservicesforDepartmentinformationthatisavailabletothepublic.Theselanguageassistanceservicesareavailablefreeofcharge.Ifyouneedmoreinformationaboutinterpretationortranslationservices,pleasecall1-800-USA-LEARN(1-800-872-5327)(TTY:1-800-437-0833);emailusat

Ed.Language.Assistance@

;orwritetoU.S.DepartmentofEducation,InformationResourceCenter,LBJEducationBuilding,400MarylandAve.SW,Washington,DC20202.

HowtoCite

Whilepermissiontoreprintthispublicationisnotnecessary,thesuggestedcitationisasfollows:

U.S.DepartmentofEducation,OfficeofEducationalTechnology,ArtificialIntelligenceandFutureofTeachingandLearning:InsightsandRecommendations,Washington,DC,2023.

Thisreportisavailableat

TableofContents

Introduction1

RisingInterestinAIinEducation 1

ThreeReasonstoAddressAIinEducationNow 2

TowardPoliciesforAIinEducation 3

BuildingEthical,EquitablePoliciesTogether6

GuidingQuestions 6

Foundation1:CenterPeople(Parents,Educators,andStudents) 6

Foundation2:AdvanceEquity7

Foundation3:EnsureSafety,Ethics,andEffectiveness8

Foundation4:PromoteTransparency9

OverviewofDocument10

WhatisAI?11

Perspective:Human-LikeReasoning 12

Perspective:AnAlgorithmthatPursuesaGoal 12

Perspective:IntelligenceAugmentation14

Definitionof“Model”14

Insight:AISystemsEnableNewFormsofInteraction15

KeyRecommendation:HumanintheLoopAI16

Learning18

Insight:AIEnablesAdaptivityinLearning 18

IntelligentTutoringSystems:AnExampleofAIModels 19

ImportantDirectionsforExpandingAI-BasedAdaptivity20

ADuality:LearningWithandAboutAI22

AChallenge:SystemsThinkingAboutAIinEducation22

OpenQuestionsAboutAIforLearning23

KeyRecommendation:SeekAIModelsAlignedtoaVisionforLearning24

Teaching25

AlwaysCenterEducatorsinInstructionalLoops 25

Insight:UsingAItoImproveTeachingJobs 26

PreparingandSupportingTeachersinPlanningandReflecting29

Designing,Selecting,andEvaluatingAITools30

Challenge:BalancingHumanandComputerDecision-Making30

Challenge:MakingTeachingJobsEasierWhileAvoidingSurveillance31

Challenge:RespondingtoStudents’StrengthsWhileProtectingTheirPrivacy32

QuestionsWorthAskingAboutAIforTeaching34

KeyRecommendation:Inspectable,Explainable,OverridableAI34

FormativeAssessment37

BuildingonBestPractices 37

ImplicationsforTeachingandLearning 38

Insight:AICanEnhanceFeedbackLoops39

AnExample:AutomatedEssayScoring40

KeyOpportunitiesforAIinFormativeAssessment41

KeyRecommendation:HarnessAssessmentExpertisetoReduceBias42

RelatedQuestions43

ResearchandDevelopment44

Insight:ResearchCanStrengthentheRoleofContextinAI 44

AttentiontotheLongTailofLearnerVariability 46

PartnershipinDesign-BasedResearch47

Re-thinkingTeacherProfessionalDevelopment48

ConnectingwithPublicPolicy49

KeyRecommendation:FocusR&DonAddressingContext50

OngoingQuestionsforResearchers50

DesiredNationalR&DObjectives51

Recommendations52

Insight:AligningAItoPolicyObjectives 52

CallingEducationLeaderstoAction 53

Recommendation#1:EmphasizeHumansintheLoop53

Recommendation#2:AlignAIModelstoaSharedVisionforEducation54

Recommendation#3:DesignUsingModernLearningPrinciples56

Recommendation#4:PrioritizeStrengtheningTrust57

Recommendation#5:InformandInvolveEducators57

Recommendation#6:FocusR&DonAddressingContextandEnhancingTrustandSafety59

Recommendation#7:DevelopEducation-SpecificGuidelinesandGuardrails60

NextSteps60

CommonAcronymsandAbbreviations62

Acknowledgements63

References64

1

Introduction

TheU.S.DepartmentofEducation(Department)iscommittedtosupportingtheuseoftechnologytoimproveteachingandlearningandtosupportinnovationthroughouteducationalsystems.Thisreportaddressestheclearneedforsharingknowledgeanddevelopingpoliciesfor“ArtificialIntelligence,”arapidlyadvancingclassoffoundationalcapabilitieswhichareincreasinglyembeddedinalltypesofeducationaltechnologysystemsandarealsoavailabletothepublic.Wewillconsider“educationaltechnology”(edtech)toincludeboth(a)technologiesspecificallydesignedforeducationaluse,aswellas(b)generaltechnologiesthatarewidelyusedineducationalsettings.Recommendationsinthisreportseektoengageteachers,educationalleaders,policymakers,researchers,andeducationaltechnologyinnovatorsandprovidersastheyworktogetheronpressingpolicyissuesthatariseasArtificialIntelligence(AI)isusedineducation.

AIcanbedefinedas“automationbasedonassociations.”Whencomputersautomatereasoningbasedonassociationsindata(orassociationsdeducedfromexpertknowledge),twoshiftsfundamentaltoAIoccurandshiftcomputingbeyondconventionaledtech:(1)fromcapturingdatatodetectingpatternsindataand(2)fromprovidingaccesstoinstructionalresourcestoautomatingdecisionsaboutinstructionandothereducationalprocesses.Detectingpatternsandautomatingdecisionsareleapsinthelevelofresponsibilitiesthatcanbedelegatedtoacomputersystem.TheprocessofdevelopinganAIsystemmayleadtobiasinhowpatternsaredetectedandunfairnessinhowdecisionsareautomated.Thus,educationalsystemsmustgoverntheiruseofAIsystems.ThisreportdescribesopportunitiesforusingAItoimproveeducation,recognizeschallengesthatwillarise,anddevelopsrecommendationstoguidefurtherpolicydevelopment.

RisingInterestinAIinEducation

Today,manyprioritiesforimprovementstoteachingandlearningareunmet.Educatorsseektechnology-enhancedapproachesaddressingtheseprioritiesthatwouldbesafe,effective,andscalable.Naturally,educatorswonderiftherapidadvancesintechnologyineverydaylivescouldhelp.Likeallofus,educatorsuseAI-poweredservicesintheireverydaylives,suchasvoiceassistantsintheirhomes;toolsthatcancorrectgrammar,completesentences,andwriteessays;andautomatedtripplanningontheirphones.ManyeducatorsareactivelyexploringAItoolsastheyarenewlyreleasedtothepublic1.EducatorsseeopportunitiestouseAI-poweredcapabilitieslikespeechrecognitiontoincreasethesupportavailabletostudentswithdisabilities,multilinguallearners,andotherswhocouldbenefitfromgreateradaptivityandpersonalizationindigitaltoolsforlearning.TheyareexploringhowAIcanenablewritingorimprovinglessons,aswellastheirprocessforfinding,choosing,andadaptingmaterialforuseintheirlessons.

Educatorsarealsoawareofnewrisks.Useful,powerfulfunctionalitycanalsobeaccompaniedwithnewdataprivacyandsecurityrisks.EducatorsrecognizethatAIcanautomaticallyproduceoutputthatisinappropriateorwrong.TheyarewarythattheassociationsorautomationscreatedbyAImayamplifyunwantedbiases.Theyhavenotednewwaysinwhichstudentsmay

1WaltonFamilyFoundation(March1,2023).TeachersandstudentsembraceChatGPTforeducation.

/learning/teachers-and-students-embrace-chatgpt-for-education

2

representothers’workastheirown.Theyarewell-awareof“teachablemoments”andpedagogicalstrategiesthatahumanteachercanaddressbutareundetectedormisunderstoodbyAImodels.Theyworrywhetherrecommendationssuggestedbyanalgorithmwouldbefair.Educators’concernsaremanifold.EveryoneineducationhasaresponsibilitytoharnessthegoodtoserveeducationalprioritieswhilealsoprotectingagainstthedangersthatmayariseasaresultofAIbeingintegratedinedtech.

Todevelopguidanceforedtech,theDepartmentworkscloselywitheducationalconstituents.Theseconstituentsincludeeducationalleaders—teachers,faculty,supportstaff,andothereducators—researchers;policymakers;advocatesandfunders;technologydevelopers;communitymembersandorganizations;and,aboveall,learnersandtheirfamilies/caregivers.Recently,throughitsactivitieswithconstituents,theDepartmentnoticedasharpriseininterestandconcernaboutAI.Forexample,a2021fieldscanfoundthatdevelopersofallkindsoftechnologysystems—forstudentinformation,classroominstruction,schoollogistics,parent-teachercommunication,andmore—expecttoaddAIcapabilitiestotheirsystems.ThroughaseriesoffourlisteningsessionsconductedinJuneandAugust2022andattendedbymorethan700attendees,itbecameclearthatconstituentsbelievethatactionisrequirednowinordertogetaheadoftheexpectedincreaseofAIineducationtechnology—andtheywanttorolluptheirsleevesandstartworkingtogether.Inlate2022andearly2023,thepublicbecameawareofnewgenerativeAIchatbotsandbegantoexplorehowAIcouldbeusedtowriteessays,createlessonplans,produceimages,createpersonalizedassignmentsforstudents,andmore.Frompublicexpressioninsocialmedia,atconferences,andinnewsmedia,theDepartmentlearnedmoreaboutrisksandbenefitsofAI-enabledchatbots.AndyetthisreportwillnotfocusonaspecificAItool,service,orannouncement,becauseAI-enabledsystemsevolverapidly.Finally,theDepartmentengagedtheeducationalpolicyexpertiseavailableinternallyandinitsrelationshipswithAIpolicyexpertstoshapethefindingsandrecommendationsinthisreport.

ThreeReasonstoAddressAIinEducationNow

“IstronglybelieveintheneedforstakeholderstounderstandthecyclicaleffectsofAIandeducation.Byunderstandinghowdifferentactivitiesaccrue,wehavetheabilitytosupportvirtuouscycles.Otherwise,wewilllikelyallowviciouscyclestoperpetuate.”

—LydiaLiu

Duringthelisteningsessions,constituentsarticulatedthreereasonstoaddressAInow:

First,AImayenableachievingeducationalprioritiesinbetterways,atscale,andwithlowercosts.Addressingvariedunfinishedlearningofstudentsduetothepandemicisapolicypriority,andAImayimprovetheadaptivityoflearningresourcestostudents’strengthsandneeds.Improvingteachingjobsisapriority,andviaautomatedassistantsorothertools,AImayprovideteachersgreatersupport.AImayalsoenableteacherstoextendthesupporttheyoffertoindividualstudentswhentheyrunoutoftime.Developingresourcesthatareresponsivetotheknowledgeandexperiencesstudentsbringtotheirlearning—theircommunityandculturalassets—isapriority,andAImayenablegreatercustomizabilityofcurricularresourcestomeetlocalneeds.

3

Asseeninvoiceassistants,mappingtools,shoppingrecommendations,essay-writingcapabilities,andotherfamiliarapplications,AImayenhanceeducationalservices.

Second,urgencyandimportancearisethroughawarenessofsystem-levelrisksandanxietyaboutpotentialfuturerisks.Forexample,studentsmaybecomesubjecttogreatersurveillance.Someteachersworrythattheymaybereplaced—tothecontrary,theDepartmentfirmlyrejectstheideathatAIcouldreplaceteachers.Examplesofdiscriminationfromalgorithmicbiasareonthepublic’smind,suchasavoicerecognitionsystemthatdoesn’tworkaswellwithregionaldialects,oranexammonitoringsystemthatmayunfairlyidentifysomegroupsofstudentsfordisciplinaryaction.SomeusesofAImaybeinfrastructuralandinvisible,whichcreatesconcernsabouttransparencyandtrust.AIoftenarrivesinnewapplicationswiththeauraofmagic,buteducatorsandprocurementpoliciesrequirethatedtechshowefficacy.AImayprovideinformationthatappearsauthentic,butactuallyisinaccurateorlackingabasisinreality.Ofthehighestimportance,AIbringsnewrisksinadditiontothewell-knowndataprivacyanddatasecurityrisks,suchastheriskofscalingpatterndetectorsandautomationsthatresultin“algorithmicdiscrimination”(e.g.,systematicunfairnessinthelearningopportunitiesorresourcesrecommendedtosomepopulationsofstudents).

Third,urgencyarisesbecauseofthescaleofpossibleunintendedorunexpectedconsequences.WhenAIenablesinstructionaldecisionstobeautomatedatscale,educatorsmaydiscoverunwantedconsequences.Inasimpleexample,ifAIadaptsbyspeedingcurricularpaceforsomestudentsandbyslowingthepaceforotherstudents(basedonincompletedata,poortheories,orbiasedassumptionsaboutlearning),achievementgapscouldwiden.Insomecases,thequalityofavailabledatamayproduceunexpectedresults.Forexample,anAI-enabledteacherhiringsystemmightbeassumedtobemoreobjectivethanhuman-basedrésuméscoring.Yet,iftheAIsystemreliesonpoorqualityhistoricaldata,itmightde-prioritizecandidateswhocouldbringbothdiversityandtalenttoaschool’steachingworkforce.

Insummary,itisimperativetoaddressAIineducationnowtorealizekeyopportunities,preventandmitigateemergentrisks,andtackleunintendedconsequences.

TowardPoliciesforAIinEducation

The2023

AIIndexReport

fromtheStanfordInstituteforHuman-CenteredAIhasdocumentednotableaccelerationofinvestmentinAIaswellasanincreaseofresearchonethics,includingissuesoffairnessandtransparency.2Ofcourse,researchontopicslikeethicsisincreasingbecauseproblemsareobserved.Ethicalproblemswilloccurineducation,too.3Thereportfoundastrikinginterestin25countriesinthenumberoflegislativeproposalsthatspecificallyincludeAI.IntheUnitedStates,multipleexecutiveordersarefocusedonensuringAIistrustworthyandequitable,andtheWhiteHouseOfficeofScienceandTechnologyPolicyhasintroduceda

2Maslej,N.,Fattorini,L.,BrynjolfssonE.,Etchemendy,J.,Ligett,K.,Lyons,T.,Manyika,J.,Ngo,H.,Niebles,J.C.,Parli,V.,Shoham,Y.,Wald,R.,Clark,J.andPerrault,R.,(2023).TheAIindex2023annualreport.StanfordUniversity:AIIndexSteeringCommittee,InstituteforHuman-CenteredAI.

3Holmes,W.&Porayska-Pomsta,K.(Eds.)(2022).Theethicsofartificialintelligenceineducation.Routledge.ISBN978-0367349721

4

BlueprintforanAIBillofRights

(Blueprint)4thatprovidesprinciplesandpracticesthathelpachievethisgoal.Theseinitiatives,alongwithotherAI-relatedpolicyactivitiesoccurringinboththeexecutiveandlegislativebranches,willguidetheuseofAIthroughoutallsectorsofsociety.

InEurope,theEuropeanCommissionrecentlyreleased

Ethicalguidelinesontheuseofartificial

intelligence(AI)anddatainteachingandlearningforeducators

.5

AIismovingfastandheraldingsocietalchangesthatrequireanationalpolicyresponse.Inadditiontobroadpoliciesforallsectorsofsociety,education-specificpoliciesareneededtoaddressnewopportunitiesandchallengeswithinexistingframeworksthattakeintoconsiderationfederalstudentprivacylaws(suchasthe

FamilyEducationalRightsandPrivacy

Act

,orFERPA),aswellassimilarstaterelatedlaws.AIalsomakesrecommendationsandtakesactionsautomaticallyinsupportofstudentlearning,andthuseducatorswillneedtoconsider

howsuchrecommendationsandactionscancomplywithlawssuchasthe

Individualswith

DisabilitiesEducationAct

(IDEA).Wediscussspecificpoliciesintheconcludingsection.

Figure1:ResearchaboutAIisgrowingrapidly.Otherindicators,suchasdollarsinvestedandnumberofpeopleemployed,showsimilartrends.

AIisadvancingexponentially(seeFigure1),withpowerfulnewAIfeaturesforgeneratingimagesandtextbecomingavailabletothepublic,andleadingtochangesinhowpeoplecreatetextand

4WhiteHouseOfficeofScienceandTechnologyPolicy(October2022),BlueprintforanAIbillofrights:MakingautomatedsystemsworkfortheAmericanpeople.TheWhiteHouseOfficeofScienceandTechnologyPolicy.

/ostp/ai-bill-of-rights/

5EuropeanCommission,Directorate-GeneralforEducation,Youth,SportandCulture.(2022).Ethicalguidelinesontheuseofartificialintelligence(AI)anddatainteachingandlearningforeducators,PublicationsOfficeoftheEuropean

Union.

https://data.europa.eu/doi/10.2766/153756

5

images6.TheadvancesinAIarenotonlyhappeninginresearchlabsbutalsoaremakingnewsinmainstreammediaandineducational-specificpublications.

ResearchershavearticulatedarangeofconceptsandframeworksforethicalAI7,aswellasforrelatedconceptssuchasequitable,responsible,andhuman-centeredAI.Listeningsessionparticipantscalledforbuildingontheseconceptsandframeworksbutalsorecognizedtheneedtodomore;participantsnotedapressingneedforguardrailsandguidelinesthatmakeeducationaluseofAIadvancessafe,especiallygiventhisacceleratingpaceofincorporationofAIintomainstreamtechnologies.Aspolicydevelopmenttakestime,policymakersandeducationalconstituentstogetherneedtostartnowtospecifytherequirements,disclosures,regulations,andotherstructuresthatcanshapeapositiveandsafefutureforallconstituents—especiallystudentsandteachers.

Policiesareurgentlyneededtoimplementthefollowing:

1.leverageautomationtoadvancelearningoutcomeswhileprotectinghumandecisionmakingandjudgment;

2.interrogatetheunderlyingdataqualityinAImodelstoensurefairandunbiasedpatternrecognitionanddecisionmakingineducationalapplications,basedonaccurateinformationappropriatetothepedagogicalsituation;

3.enableexaminationofhowparticularAItechnologies,aspartoflargeredtechoreducationalsystems,mayincreaseorundermineequityforstudents;and

4.takestepstosafeguardandadvanceequity,includingprovidingforhumanchecksandbalancesandlimitinganyAIsystemsandtoolsthatundermineequity.

6Sharples,M.&PérezyPérez,R.(2022).Storymachines:Howcomputershavebecomecreativewriters.Routledge.ISBN9780367751951

7Akgun,S.,Greenhow,C.(2022).Artificialintelligenceineducation:AddressingethicalchallengesinK-12settings.AIEthics,2,431–440./10.1007/s43681-021-00096-7

6

BuildingEthical,Equitable

PoliciesTogether

Inthisreport,weaimtobuildonthelisteningsessionstheDepartmenthostedtoengageandinformallconstituentsinvolvedinmakingeducationaldecisionssotheycanprepareforandmakebetterdecisionsabouttheroleofAIinteachingandlearning.AIisacomplexandbroadtopic,andwearenotabletocovereverythingnorresolveissuesthatstillrequiremoreconstituentengagement.Thisreportisintendedtobeastartingpoint.

TheopportunitiesandissuesofAIineducationareequallyimportantinK-12,highereducation,andworkforcelearning.Duetoscopelimitations,theexamplesinthisreportwillfocusonK-12education.Theimplicationsaresimilaratalllevelsofeducation,andtheDepartmentintendsfurtheractivitiesin2023toengageconstituentsbeyondK-12schools.

GuidingQuestions

UnderstandingthatAIincreasesautomationandallowsmachinestodosometasksthatonlypeopledidinthepastleadsustoapairofbold,overarchingquestions:

1.Whatisourcollectivevisionofadesirableandachievableeducationalsystemthatleveragesautomationtoadvancelearningwhileprotectingandcenteringhumanagency?

2.Howandonwhattimelinewillwebereadywithnecessaryguidelinesandguardrails,aswellasconvincingevidenceofpositiveimpacts,sothatconstituentscanethicallyandequitablyimplementthisvisionwidely?

IntheLearning,Teaching,andAssessmentsectionsofthisreport,weelaborateonelementsofaneducationalvisiongroundedinwhattoday’slearners,teachers,andeducationalsystemsneed,andwedescribekeyinsightsandnextstepsrequired.Below,wearticulatefourkeyfoundationsforframingthesethemes.Thesefoundationsarisefromwhatweknowabouttheeffectiveuseofeducationaltechnologytoimproveopportunity,equity,andoutcomesforstudentsandalsorelatetothenewBlueprint.

Foundation1:CenterPeople(Parents,Educators,andStudents)

Education-focusedAIpoliciesatthefederal,state,anddistrictlevelswillbeneededtoguideandempowerlocalandindividualdecisionsaboutwhichtechnologiestoadoptanduseinschoolsandclassrooms.Considerwhatishappeningineverydaylives.ManyofususeAI-enabledproductsbecausetheyareoftenbetterandmoreconvenient.Forexample,fewpeoplewanttousepapermapsanymore;peoplefindthattechnologyhelpsusplanthebestroutetoadestinationmoreefficientlyandconveniently.Andyet,peopleoftendonotrealizehowmuchprivacytheyaregivingupwhentheyacceptAI-enabledsystemsintotheirlives.AIwillbringprivacyandotherrisksthatarehardtoaddressonlyviaindividualdecisionmaking;additionalprotectionswillbeneeded.

7

Thereshouldbeclearlimitsontheabilitytocollect,use,transfer,andmaintainourpersonaldata,includinglimitsontargetedadvertising.Theselimitsshouldputtheburdenonplatformstominimize