PS68CH16-Metcalfe ARI 7September2016 12:2 E V I E R W S E I C N N A D V A Learning from Errors Janet Metcalfe DepartmentofPsychology,ColumbiaUniversity,NewYork,NewYork10027; email:[email protected] Annu.Rev.Psychol.2017.68:6.1–6.25 Keywords TheAnnualReviewofPsychologyisonlineat errorlesslearning,generationeffect,hypercorrectioneffect,feedback, psych.annualreviews.org after-actionreview(AAR),errormanagementtraining(EMT),formative Thisarticle’sdoi: assessment,reconsolidation,predictionerror 10.1146/annurev-psych-010416-044022 Copyright c 2017byAnnualReviews. Abstract ⃝ Allrightsreserved AlthougherroravoidanceduringlearningappearstobetheruleinAmeri- canclassrooms,laboratorystudiessuggestthatitmaybeacounterproductive strategy,atleastforneurologicallytypicalstudents.Experimentalinvestiga- tionsindicatethaterrorfullearningfollowedbycorrectivefeedbackisben- eficialtolearning.Interestingly,thebeneficialeffectsareparticularlysalient whenindividualsstronglybelievethattheirerroriscorrect:Errorscommit- ted with high confidence are corrected more readily than low-confidence errors.Correctivefeedback,includinganalysisofthereasoningleadingup tothe mistake, is crucial. Aside fromthe direct benefit tolearners, teach- ersgainvaluableinformationfromerrors,anderrortoleranceencourages students’active,exploratory,generativeengagement.Ifthegoalisoptimal performance in high-stakes situations, it may be worthwhile to allow and even encourage students to commit and correct errors while they are in low-stakeslearningsituationsratherthantoassiduouslyavoiderrorsatall costs. 6.1 PS68CH16-Metcalfe ARI 7September2016 12:2 Contents INTRODUCTION............................................................... 6.2 ENCOURAGINGVERSUSDISCOURAGINGERRORS INTHECLASSROOM........................................................ 6.3 ERRORGENERATIONANDMEMORYFORCORRECTRESPONSES INTHELAB.................................................................. 6.4 FeedbacktoErrors............................................................. 6.6 SourcesofErrors............................................................... 6.7 CONFIDENCEINERRORS..................................................... 6.8 ErrorsVersusGuesses.......................................................... 6.8 CorrectingHigh-ConfidenceVersusLow-ConfidenceErrors..................... 6.8 ExplanationsoftheHypercorrectionEffect...................................... 6.8 EXCEPTIONS................................................................... 6.11 AmnesicsandError-FreeLearning.............................................. 6.11 OlderAdultsandHypercorrection...............................................6.11 IMPLICATIONSOFTHEHYPERCORRECTIONEFFECT.................... 6.12 ScientificUnderstanding........................................................ 6.12 LexicalRepresentations......................................................... 6.12 CorrectingFalseInferencesinEpisodicMemory................................. 6.12 THEORIESOFWHYERRORSENHANCELEARNING....................... 6.13 Mediation...................................................................... 6.13 RecursiveReminding........................................................... 6.13 Reconsolidation................................................................ 6.14 PredictionError................................................................6.14 SECONDARYBENEFITSOFENCOURAGINGERRORS...................... 6.15 ActiveGenerationIsBeneficialtoMemoryforCorrectResponses................ 6.15 KnowledgeofStudentErrorsHelpstheTeacher.................................6.15 ConsiderationofErrorsOffsetsOverconfidence..................................6.16 ORIGINOFTHEIDEATHATERRORLESSLEARNING ISAGOODTHING.......................................................... 6.16 EMOTIONALCONSEQUENCESOFERRORS.................................6.17 Event-RelatedPotentialStudies................................................. 6.17 EmotionalConsequencesofMedicalandPoliceErrors........................... 6.18 PersonalityDifferencesinResponsivitytoErrors.................................6.18 MethodstoBufferPotentialNegativeEmotionalEffects..........................6.19 USINGERRORSTOIMPROVELEARNING................................... 6.19 CONCLUSION.................................................................. 6.20 INTRODUCTION Nobodywantstomakeerrorsinasituationthatcounts.Theconsequencesofcommittingsuch errors can be devastating. If one is performing a piano solo before an audience, controlling a nuclear reactor, taking SATs, making a medical decision, fighting on a battlefield, or giving a lecture,thelastthingonewantsisanerror.Thisreviewisnotdirectedatthequestionofwhether errors,inasituationthatcounts,aregood—ofcoursetheyarenot.Rather,thequestionishow, 6.2 Metcalfe PS68CH16-Metcalfe ARI 7September2016 12:2 during initial learning and during practice and preparation for a test that counts, one can best gettoastateofperformancethatisoptimalandinwhicherrorswillnotinadvertentlyoccurjust whenoneneedsthemleastandwhentheywilldothemostdamage.Shouldonecommit,explore, examine,analyze,andcorrecterrorsduringlearningandpracticesessions,orshouldoneavoid errorsatallstagesoflearning? It might seem intuitive that if one does not want errors on the test that counts, then one shouldavoiderrorsatallstagesoflearning.Inthisview,committingerrorsshouldmakethose errorsmoresalientand entrenchthembothinto thememoryand theoperating procedures of thepersonwhomakesthem.Exercisingtheerrorsshouldmaketheerrorsthemselvesstronger, thusincreasingtheirprobabilityofrecurrence.Suchaview,whichisconsistentwithanumberof theoldestandmostwellestablishedtheoriesoflearningandmemory(Bandura1986,Barnes& Underwood1959,Skinner1953),suggeststhaterrorsarebadandshouldbeavoidedatallcosts. Inkeepingwiththeseviews,Ausubel(1968)warnedofthedangersoferrorsinthelearning processandsuggestedthatallowingpeopletomakeerrorsencouragesthemtopracticeincorrect andinefficientapproachesthatwillcausetroublebecausetheyaredifficulttooverwritelaterwith correct approaches. He used this reasoning to argue against an exploratory learning strategy, which by its very nature would mean that incorrect paths and faulty approaches and solutions wouldbeencounteredandentertainedbythelearner.Ausubel(andothers)fearedthatwithactive exploratorylearning,thesefalsestartsanderrorswouldbelearnedandwouldmakelearningofthe correctsolutionsandproceduresmoredifficult,ifnotimpossible.Accordingly,activeexploratory learningwastobeavoided. Similarly,Bandura(1986,p.47)urgedthatlearnersshouldbe“sparedthecostsandpainof faultyeffort”andthattheyshouldinsteadreceivetheneededstep-by-stepguidancethatresults inflawlessbehaviorfromtheoutset.Feedbackshouldfocusonlyonthecorrectexecutionoftasks and should take the form of positive social reinforcement, with errors—if any—being ignored. Indeed,errorscanhaveadetrimentaleffectforpeoplewhohaveparticularneurologicaldeficits; thistopicisdiscussedinfurtherdetailbelow.However,afearoferrorstypicallydivertslearning fromhighlyproductivegenerativestrategies,anderroravoidancestrategyisevenmorepernicious becausegeneratingerrors—aslongascorrectivefeedbackisgiven–isactuallybeneficialtolearning. ENCOURAGINGVERSUSDISCOURAGINGERRORS INTHECLASSROOM It is not clear how important the worry about avoiding errors has been in shaping American teaching strategies. It is extremely difficult to obtain accurate data concerning what and how teachers teach, let alone to manipulate that teaching in a manner that would allow confident inferences to be made. However, Stevenson & Stigler (1994; see also Stigler & Hiebert 2009) andtheircolleaguesconductedalandmarkstudyinwhichtheywereabletovideotapelessonsin grade8mathematicsclassroomsinavarietyofcountries,includingtheUnitedStates,Taiwan, China, and Japan. Of most interest, given that Japan is by far outstripping the United States in math scores, is the striking difference in the teaching methods used in those two countries. Althoughtheremaybemanyotherreasonsforthedifferencesinmathscores,onehighlysalient differenceiswhetherornotteachersengagewithstudents’errors.Videotapesshowthat,inthe UnitedStates,setproceduresfordoingparticularkindsofproblemsareexplicitlytaught.These correctproceduresarerehearsedandemphasized;errorsareavoidedorignored.Thestudentsare notpassiveinAmericanclassrooms.Ateachermayaskforstudentparticipationinrepeating,for example,aprocedureforborrowingwhensubtracting.Whenaskingaquestionsuchas,“Canyou subtract9from5?”topromptstudentstoanswer,“No,youhavetoborrowtomakethe5a15,” www.annualreviews.org LearningfromErrors 6.3 • PS68CH16-Metcalfe ARI 7September2016 12:2 theteachermayfailtoevenacknowledgethedeviantchildwhosays,“Yes.It’snegative4.”Ifthe responsedoesnotfitwiththeprocedurebeingexercised,itisnotreinforced.Errors(aswellas deviantcorrectanswers)areneitherpunishednordiscussedbutaredisregarded.Praiseisgiven, butonlyforthe“correct”answer. As Stevenson & Stigler (1994) pointed out, praise curtails discussion and serves mainly to reinforcetheteacher’sroleastheauthoritywhobestowsrewards.Itdoesnotempowerstudentsto think,criticize,reconsider,evaluate,andexploretheirownthoughtprocesses.Bywayofcontrast, in Japan praise is rarely given. There, the norm is extended discussion of errors, including the reasons for them and the ways in which they may seem plausible but nevertheless lead to the incorrectanswer,aswellasdiscussionoftherouteandreasonstothecorrectanswer.Suchin- depthdiscussionofthethoughtprocessesunderlyingbothactualandpotentialerrorsencourages exploratoryapproachesbystudents. Insteadofbeginningwithteacher-directedclassworkandexplication,Japanesestudentsfirst try to solve problems on their own, a process that is likely to be filled with false starts. Only after these (usually failed) attempts by students does teacher-directed discussion—interactively involvingstudentsandtargetingstudents’initialeffortsandcoremathematicalprinciples—occur. It is expected that students will struggle and make errors, insofar as they rarely have available a fluent procedure that allows them to solve the problems. Nor are students expected to find theprocessoflearningeasy.Butthetimespentstrugglingontheirowntoworkoutasolution is considered a crucial part of the learning process, as is the discussion with the class when it reconvenestosharethemethods,todescribethedifficultiesandpitfallsaswellastheinsights,and toprovidefeedbackontheprinciplesatstakeaswellasthesolutions. AsStevenson&Stigler(1994,p.193)note,“Perhapsbecauseofthestronginfluenceofbehav- ioristicteaching,whichsaysconditionsshouldbearrangedsothatthelearneravoidserrorsand makesonlyareinforceableresponse,Americanteachersplacelittleemphasisontheconstructive useoferrorsasateachingtechnique.Learningaboutwhatiswrongmayhastenunderstanding ofwhythecorrectproceduresareappropriate,buterrorsmayalsobeinterpretedasfailure.And Americans,reluctanttohavesuchinterpretationsmadeoftheirchildren’sperformance,striveto avoidsituationswherethismighthappen.” TheJapaneseactivelearningapproachwellreflectsthefundamentalideasofalearning-from- errors approach. Engaging with errors is difficult, but difficulty can be desirable for learning (Bjork2012).Incomparisonwithapproachesthatstresserroravoidance,makingtrainingmore challengingbyallowingfalsestartsanderrorsfollowedbyfeedback,discussion,andcorrection mayultimatelyleadtobetterandmoreflexibletransferofskillstolatercriticalsituations. ERRORGENERATIONANDMEMORYFORCORRECT RESPONSESINTHELAB Considerableresearchnowindicatesthatengagementwitherrorsfostersthesecondarybenefits of deep discussion of thought processes and exploratory active learning and that the view that thecommissionoferrorshurtslearningofthecorrectresponseisincorrect.Indeed,manytightly controlledexperimentalinvestigationshavenowshownthatincomparisonwitherror-freestudy, thegenerationoferrors,aslongasitisfollowedbycorrectivefeedback,resultsinbettermemory for the correct response. Oddly, the researchers who have investigated these processes usually refertoerrorsas“unsuccessfulretrieval,”leavingambiguouswhetheramistakewasproducedor whetherthelearnerhadjustneglectedtogeneratethecorrectresponse. EarlystudiesbyIzawa(1967,1970)showedthatmultipleunsuccessfulretrievalattemptsledto bettermemoryforthecorrectfeedbackthandidaprocedureproducingfewerincorrectresponses. 6.4 Metcalfe PS68CH16-Metcalfe ARI 7September2016 12:2 Kane&Anderson(1978)showedsimilarresults:Attemptingthegenerationofthelastwordofthe sentence,evenifwhatwasgeneratedwaswrong,ledtoenhancedcorrectperformancecompared toreadingthesentencecorrectlyfromtheoutset.Slamecka&Fevreiski(1983)askedpeopleto remembernearantonyms,suchastrivial-vitaloroscillate-settle.Evenfailedattempts(followedby feedbackcontainingthecorrectanswer)improvedlaterrecallofthecorrectanswersoversimply readingthecorrectanswer.Kornelletal.(2015)haveconductedarecentinvestigationofthesame issueandhavereachedsimilarconclusions. Peoplerarelyproducedexplicitmistakesintheearlyexperiments.Themajorityofso-calledre- trievalfailureswereomissionratherthancommissionerrors.However,whenSlamecka&Fevreiski (1983)specificallyanalyzedtheeffectofcommissionascomparedtoomissionerrors,theyfound thatcommissionerrors,aslongastheyweresemanticallyrelatedtothetarget,resultedinbetter laterrecallofthecorrectanswer. Kornelletal.(2009)conductedthefirstdefinitivestudythatdirectlycomparedtheeffectof producingversusnotproducingacommissionerror.Thecueandtheto-be-rememberedtarget wordinmostoftheexperimentsintheirstudywereslightlyrelatedwordpairs.Theycompared aconditioninwhichtheanswer,ortarget,wassimplygiventoparticipants,withnointervening errorgeneration(theno-errorcondition),tooneinwhichtheparticipantswereaskedtoguessthe answerfirstandnearlyalwaysproducedanerrorbeforebeinggiventhecorrectanswertostudy(the error-generationcondition).Theexperimentwascarefullycontrolledtoensurethattheamount oftimespentstudyingthecorrectanswerwasequatedacrossconditions.Kornelletal.(2009)also eliminatedfromconsiderationanyinstancesinwhichthepersondidnotgenerateanerrorinthe error-generation condition. This happened less than 10% of the time. The surprising finding, whichhasnowbeenreplicatedmanytimes,wasthatonthefinaltest,participantsremembered thecorrectanswersconsiderablybetterwhentheyhadgeneratedanerrorthanwhentheyhadnot. Itappears,then,thaterrorgenerationisnotinevitablybadandtobeavoidedatallcosts.Indeed, errorgenerationappearstofosterlearning(seeFigure1). A flurry of replications quickly followed. For example, Huelser & Metcalfe (2012) found a beneficialeffectoferrorgenerationaslongastheerrorspeopleproducedwerenotcompletely unrelatedtothetarget.Theyfoundnobenefitwithunrelatedpairs—afindingthatwasreplicated byGrimaldi&Karpicke(2012)andbyKnightetal.(2012)andthatisconsistentwithSlamecka& 1.0 0.9 Error free Error generating 0.8 ct 0.7 e r or 0.6 c n o 0.5 ti or 0.4 p Pro 0.3 0.2 0.1 0.0 Three Four Five Six Experiment Figure1 Proportionofcorrectresponsesfollowingerror-generatinganderror-freestudy.Figureadaptedwith permissionfromKornelletal.(2009). www.annualreviews.org LearningfromErrors 6.5 • PS68CH16-Metcalfe ARI 7September2016 12:2 Fevreiski’s(1983)earlieranalysis.Similarly,Kangetal.(2011)foundnoeffectwhenparticipants had absolutely no idea what the answers to factual questions might be but were nevertheless forcedtoguess.Itappearsthattobebeneficial,theguessneedstobesomewhatinformedrather than a shot in the dark. Interestingly, in the related-pair case in which a large beneficial effect of committing errors was found, the participants were metacognitively unaware of the benefit. Evenimmediatelyaftertheyhadexperiencedthetaskandhadevidencedabenefitof20%(i.e., roughlythedifferencebetweenaC andanA,ifithadbeenacoursegrade),participantsthought − thattheerror-freeconditionhadresultedinbetterrecall(Huelser&Metcalfe2012).Thislack ofawarenessofthebenefitsoferrorgenerationmaycontributetotheaversiontoerrorsinthe AmericanteachingstyleevincedinStigler’swork. Inamoreeducationallyrealisticstudy,Richlandetal.(2009)foundthaterrorgenerationwas relatedtoenhancedmemoryformaterialfromreadingpassages.Intheirstudy,materialfromthe passagewastestedbeforeparticipantshadachancetoreadthepassage,resultinginparticipants generatingmanyerrors.However,latercorrectmemoryforthematerialthathadbeenpretested in this way was greatly enhanced. More recently, instead of using simple word pairs, Kornell etal.(2015)usedgeneralinformationastheirstimuli,askingpeopleeithertotrytogeneratethe responses(resultinginmanymistakes)ortoreadthequestionandcorrectanswer.Inthissituation, too,generationoftheerrorhelpedlatermemoryforthecorrectresponse.Thus,itappearsthat generatinganerror,ifitisrelatedinanywaytothecorrectresponse,enhancesratherthanimpairs memoryforthecorrectanswer. FeedbacktoErrors Whenpeoplehavemadeanerror,correctivefeedbackiscrucial(Andersonetal.1971;Butler& Roediger2008;Hancocketal.1992;Kornell&Metcalfe2013;Lhyle&Kulhavy1987;Metcalfe &Kornell2007;Metcalfeetal.2007,2009;Pashleretal.2005).Itisnotenoughtosimplytell learnerswhethertheywererightorwrong.Peoplegetvirtuallynobenefitunlessthefeedbackthey receiveprovidesthecorrectanswer,asPashleretal.(2005)showed(andseeBangert-Drownsetal. 1991,Moreno2004).Furthermore,peopleneedtounderstandandpayattentiontothefeedback. Manystudiesshowingalackofbenefitfromfeedback(andsomeshowingalackofbenefitfrom committingerrors)failtoensurethatthefeedbackisprocessed.Whenthecorrectanswerismade available, though, and people appreciate that the answer is correct as well as why that answer is correct, they are able to integrate that information into memory and improve performance (Andersonetal.1971). Furthermore,Metcalfeetal.(2009)showedthatthefeedbackdidnothavetobegivenwithin momentsoferrorcommissiontobeeffective.Theycontrastedaconditioninwhichpeoplewere giventhefeedbackimmediatelyuponerrorcommissionwithoneinwhichitwasdelayedbyuptoa week(andtheycontrolledforlaguntiltest,whichwasoneweekfromfeedbackinbothconditions, sotheamountoftimeuntilthetestdidnotdifferinthetwoconditions).Inboththeimmediate anddelayedfeedbackconditions,participantswererequiredtoattendtothefeedback,insofaras theywereinstructedtotypetheanswerintothecomputer.Thestudyfoundthatcollegestudents performedequallywellintheimmediateanddelayedfeedbackconditions,whereaschildrenin grades3to5didbetterwhenthefeedbackwasdelayed.Interestingly,Kulik&Kulik(1988)noted that whether delayed or immediate feedback produced better results differed between studies conductedintheclassroomandthoseinthelaboratory.Labstudiestendedtoshowthatdelayed feedback was better, whereas classroom studies favored immediate feedback. They concluded, however,thattherealdifferencebetweenthesestudieswaswhetherthelearnerspaidattentionto thefeedback.Studentsintheclassroomarehighlyengagedinknowingtheanswerstoquestions 6.6 Metcalfe PS68CH16-Metcalfe ARI 7September2016 12:2 rightaftertakingatest.Theypayattentiontothefeedbackwhenitisgivenimmediately.However, theirinterestflagswithalongdelay,especiallyifthefeedbackiscursory.Iftheteacherneedsor wantstodelayfeedback(toallowforconscientiousscoring,forexample),itseemsclearthatthe students’interestinthequestionsneeds tobereignited,andthepresentationofthecorrective feedbackneedstobeengaging. When feedback is elaborative or scaffolded, its beneficial effects are also increased (Finn & Metcalfe2010).Theprovisionoffeedbackcanbe,asitisintheJapaneseclassrooms,extensiveand henceveryhelpfulinallowingstudentstounderstandtheunderlyingconstructsandtogeneralizeto newsituations.Feedbackprovidesanopportunityforexploringanddeeplyanalyzingtheprinciples underlyingtheproblems,theanswers,andthereasoningleadingtotheanswers—betheyright orwrong.Siegler(1995)conductedaparticularlyinterestingstudyillustratingtheimportanceof focusingstudents’attentiononthereasoningunderlyingthefeedback.Hehadthreeconditions: (a) a feedback-only condition; (b) an explain-own-reasoning condition, in which the children wereasked,“Whydoyouthinkthat?”;and,mosteffectively,(c)anexplain-the-correctreasoning condition, in which the students, after having been given the corrective feedback, were asked, “HowdoyouthinkIknewthat?”AsSiegler(2002,p.40)noted,“Havingthechildrenexplain anotherperson’scorrectreasoninghastheadvantageofbothdiscoveryanddidacticapproaches to instruction. It is like discovery-oriented approaches in that it requires the child to generate a relatively deep analysis of a phenomenon without being told how to do so. It is like didactic approaches in that it focuses the child’s attention on the correct reasoning. Thus, it combines someoftheefficiencyofdidacticinstructionwithsomeofthemotivatingpropertiesofdiscovery.” SourcesofErrors Another fundamental question concerns whether the source of an error modulates its efficacy forlearning.Theliteratureindicatesthatalthoughself-generatinganerrorandbeingexposedto externallypresentedirrelevantinformationmayseemsimilaronthesurface,importantdifferences mayexist.Forexample,Grimaldi&Karpicke(2012)showedthaterrorgeneration,asopposedto merelybeingpresentedthecorrectresponse(inaparadigmlikethatofKornelletal.2009),had abeneficialeffectonlatermemoryforthecorrectanswer.However,wheninsteadofhavingthe participantgenerateanerror,theexperimenterpresentedanincorrectwordduringwhatwould havebeenthegenerationperiod,memoryforthecorrectanswerwasharmed.Furthermore,itwas alsoharmfultorestrictwhatthepersonhadtogenerateasanerror(aconstrainedalternative). A similar phenomenon has been observed with self-generated as compared to experimenter- presented mistaken items when people were in tip-of-the-tongues states. Kornell & Metcalfe (2006)notedthatself-generatederrors(which,typically,theparticipantsknewwerenotthecorrect answers)didnotblockretrievalofthecorrectanswer,asopposedtowhennoincorrectanswerwas produced.Incontrast,whenexperimenter-presentedincorrectanswersweregivenwhilepeople were in tip-of-the-tongue states, the incorrect answers harmed retrieval of the correct answer in comparison with having nothing presented (Smith & Blankenship 1991). Determining how self-generatedwronganswersdifferfromexperimenter-presentedwrongitemsmaybeimportant for understanding why error generation helps. It is possible that self-generated errors tap into theperson’sownsemanticmemorystructureandthuscanserveasmediatorstohelptheperson gettothecorrectanswer.Externallyimposeddistractingitemsmayfailtotakeadvantageofan individuals’internalmentalstructureandthusfailtoservethemediationalrole.Suchitemsmay simplybedistracting. Whathappensifapersonobservesanotherperson—say,afellowstudent—makeanerrorthat isthenfollowedbyfeedback?Doesthisexposurehavethesamebeneficialeffectasself-generating www.annualreviews.org LearningfromErrors 6.7 • PS68CH16-Metcalfe ARI 7September2016 12:2 one’sownerror,orisitlikebeingexposedtoirrelevantexperimenter-presenteddistractions?This isaninterestingquestionwithpracticalimplications,butitisanissuethatresearchhasnotyet addressed. CONFIDENCEINERRORS ErrorsVersusGuesses Thecriticmayarguethatperhapsthepeopleintheexperimentsdescribedabovedidnotreally believeintheirresponses:Theywerejustgeneratingguesses,notgenuineerrors.Perhapsthere isnobenefittogeneratingerrorswhenpeoplereallyendorsetheirerrorsandhavestrongbeliefs thattheyarecorrect.Thedatashowgreaterbenefitwhentheerrorswererelatedtothetarget thanwhentheywereunrelated,whichsomewhatoffsetsthiscontention.Evenso,fewclassroom teacherswouldaskstudentstomemorizelistsofwordpairssuchasOscillate-SettleorHillside- Bankerorsimilaritemsusedintheexperimentsonerrorgeneration.Kornellandcolleagues’(2015) experimentwithrealgeneral-informationquestions—whichshowederror-generationbenefitsin areal-worldfactualtask—helpsallaysomeconcernsthatbenefitsmayobtainonlywithguesses andnotwithrealerrors. Nevertheless,wemayaskhowmuchbeliefisenoughtomakeaparticularresponseagenuine errorratherthanjustaguess.Inthenextsection,wereviewtheliteratureontherelationbetween individuals’degreeofbelief,ortheirconfidenceinthetruthoftheiranswerstogeneral-knowledge questions,andtheirpropensitytocorrectthoseanswersandrememberthecorrectionswhenthey turnouttobeerroneous. CorrectingHigh-ConfidenceVersusLow-ConfidenceErrors A number of studies have investigated the correction of errors as a function of the individual’s confidence in the error. Typically, a factual question is asked of the participant; the individual generatesananswerandthenrateshisorherconfidencethattheerroriscorrect.Next,feedback isgiven,inwhichthecorrectanswerisprovided.Forexample,participantsmightbegiventhe question, “What kind of music is associated with the Cajuns in Louisiana?” Participants then giveananswer(perhaps“jazz”)andratetheirconfidenceinthecorrectnessoftheanswerthey produced.Finally,participantsaregiventhecorrectanswer(inthiscase,“zydeco”).Incontrastto thepredictionsofavarietyoftheoriesthatsuggestthatresponsesinwhichoneishighlyconfident should be particularly difficult to overwrite, the high-confidence errors are more likely to be correctedontheretestthanareerrorsendorsedwithlowerconfidence(e.g.,Butleretal.2008, 2011;Butterfield&Mangels2003;Butterfield&Metcalfe2001,2006;Cyr&Anderson2013;Eich etal.2013;Fazio&Marsh2009,2010;Iwakietal.2013;Kulhavyetal.1976;Metcalfeetal.2012; Metcalfe&Finn2011,2012;Sitzman&Rhodes2010;Sitzmanetal.2015).Thishypercorrection effectoccursbothwithimmediateretestandwhentheretestisgivenataconsiderabledelay(Butler etal.2011,Butterfield&Mangels2003,Metcalfe&Miele2014).Astrongdegreeofbeliefinthe truthofone’serrorsmakesthemmore,ratherthanless,susceptibletobeingcorrectable. ExplanationsoftheHypercorrectionEffect Anumberofassociativetheoriesofmemoryandoftherelationofmemorytoconfidenceseem toindicatethatthehypercorrectioneffectshouldnotbefound;indeed,perhapseventhereverse shouldbeobserved.Responsesthataremadewithhighconfidencearethoseinwhichtheperson believes most and are thought to be the strongest in memory (e.g., Ebbesen & Rienick 1998). 6.8 Metcalfe PS68CH16-Metcalfe ARI 7September2016 12:2 As such, they should be most easily accessible and most resistant to interference as well as the mostdifficulttooverwriteorreplacewithanewresponse.Certainly,inalldatapresentedtodate on the hypercorrection effect, the correlation between confidence in one’s first responses and thecorrectnessofthoseresponsesishigh:Theresponsesinwhichpeoplearehighlyconfident arenearlyalwayscorrect,indicatingthatingeneralpeopleknowwhattheydoanddon’tknow andthattheirhigh-confidenceresponsesarestronglyandreadilyretrievedfrommemory.Most errorsthatpeoplemakeareassignedlowconfidence(andbecausetheyareweak,shouldbeeasy tooverwritewithanewresponse).Peoplemakehigh-confidenceerrorsonlyrarely.Butifsucha high-confidenceresponsewereinerror,ittoo—likecorrecthigh-confidenceresponses—should bestrong,entrenched,anddifficultratherthaneasytochange.Theconsistentfindingofahyper- correctioneffect,wherebyhigh-confidenceerrorsarecorrectedmorereadilythanlow-confidence errors,fliesinthefaceoftraditionalinterference,responsehierarchy,andassociativetheories. Twononmutuallyexclusivefactorsarecentraltothehypercorrectionphenomenoninyoung adults,thoughthesearenottheonlyfactorsatplay.Thefirstrelatestothesurpriseindividuals experienceatbeingwrongwhentheyweresuretheywereright.Thesecondrelatestothestructure ofthesemanticnetworksurroundinghigh-ascomparedtolow-confidenceerrors. Becausetheyaresurprised(andperhapsembarrassed)athavingmadeamistakeonaresponse theystronglythoughtwascorrect,individualsmayrallytheirattentionalresourcestobetterre- memberthecorrectanswer.Severallinesofevidencesupportthissurprise/attentionalfactor.For example, Butterfield & Metcalfe (2006) conducted an experiment designed to evaluate the ex- tenttowhichonetaskdrewattentionfromanother.Subjectswererequiredtodetectwhenvery softtonesoccurredwhiletheyweredoingthegeneral-informationerror-correctiontask.Some ofthetoneswereintentionallypresentedsimultaneouslywiththecorrectivefeedbackfollowing high- and low-confidence errors. Participants were more likely to fail to detect the tones that werepresentedatthetimeofhigh-confidenceerrorfeedbackthanthosepresentedatthetime oflow-confidenceerrorfeedback. Thisresultindicates thatbecause participants’attentionwas capturedbythehigh-confidenceerrorfeedback,theyhadlesscapacitytodetectthetones. Fazio&Marsh(2009)testedformemoryforthesurroundingcontextthataccompaniedfeed- back. They showed that the surrounding context of high-confidence error feedback was better rememberedthanwasthesurroundingcontextoflow-confidenceerrorfeedback,andtheyinter- pretedthisresultasfavoringtheattentionalexplanation. Butterfield&Mangels(2003)usedyoungadultparticipantsinthefirstevent-relatedbrainpo- tential(ERP)studyonthisparadigm.Themostsalientresultfromtheirstudywasthatparticipants showedavoltagedeflection—theP3a—thatpastliteratureonERPshaslinkedtosurprisereac- tionsinwhichpeoplerallytheirattention(Friedmanetal.2001).Theliteraturehasalsoassociated thisdeflectionwithenhancedmemory(Palleretal.1985).Butterfield&Mangel’s(2003)study revealedthattherewasaconfidence-gradedP3awhencorrectivefeedbackwasgiven,aresultthat hasbeenreplicatedandextendedbyMetcalfeetal.(2015)(seeFigure2).TheP3awaslargestfor thefeedbacktohigh-confidenceerrorsandsmallestforthefeedbacktolow-confidenceerrors. Theseresultssupportthecontentionthathigh-confidenceerrorsaresurprisingandthatincreased attentionispaidtocorrectivefeedbacktosucherrors. Finally,Metcalfeetal.(2012)conductedastudyonhypercorrectionusingevent-relatedfunc- tional magnetic resonance imaging. Participants answered questions for several hours, giving theirconfidenceintheiransweroutsidethescanner.Participantsthenenteredthescannerand were presented with questions, their original answers, their original confidence ratings, and fi- nallythecorrectanswer.Whenthebrainactivationstothefeedbacktohigh-andlow-confidence errorswerecompared,itwasfoundthatmedialfrontalareasthatprominentlyincludedthean- teriorcingulate—anarearelatedtosurprise,errordetection,andattention—weredifferentially www.annualreviews.org LearningfromErrors 6.9 • PS68CH16-Metcalfe ARI 7September2016 12:2 P3a 375–425 0.9 Young 300 900 25 Young Young d robability correcteon final test 0000....5687 Older 10O µldV+–er 300 900 425–475 P3a amplitude toerror feedback 211050 Older P 0.4 5 High Low High confidence High Low confidence confidence Low confidence confidence confidence Δ = 1.0 µV Figure2 Event-relatedpotentials(ERPs)synchronizedtotheonsetofcorrectivefeedbackgiventohigh-andlow-confidenceerrorsmadeby youngadultsandolderadults.TheERPtracingsshowaprominentP3atofeedbacktohigh-confidenceerrors.Thebehavioralresults showingtheproportioncorrectonthefinaltest,forbothgroupsasafunctionofinitialerrorconfidence,areshownontheleft.Figure adaptedwithpermissionfromMetcalfeetal.(2015). activated.Otherareas,suchasthedorsolateralprefrontalcortexandthetemporalparietaljunction (seeFigure 3),werealsodifferentiallyactivated.Theseresultsalsoimplicateasurprise-related explanationoftheenhancedencodingassociatedwiththefeedbacktohigh-confidenceerrors. Thesecondfactorthathasbeenimplicatedinthehypercorrectioneffectisagreatersemantic knowledge in the domain of the high-confidence errors than in the low-confidence error do- main. Butterfield & Mangels (2003) noted that participant-ascribed familiarity to high- and to Figure3 Contrastmapofhigh-confidenceerrorsversuslow-confidenceerrors.Redareasweremoreactivefor high-confidenceerrors,andgreenareasweremoreactiveforlow-confidenceerrors;bothareashadan uncorrectedthresholdofp<0.001.FigureadaptedwithpermissionfromMetcalfeetal.(2012), c 2012by ⃝ theMassachusettsInstituteofTechnology,http://www.mitpressjournals.org/doi/abs/10.1162/jocn_a_ 00228. 6.10 Metcalfe