143-8

COGNITIVE REPAIRS: HOW ORGANIZATIONAL PRACTICESCAN

COMPENSATEFOR INDIVIDUAL SHORTCOMINGS

ChipHeath,RichardP. Larrick, andJoshuaKlayman

ABSTRACT

The literaturein cognitive psychology has described a varietyof shortcomings that preventindividuals from learningeffectively. We review this literature and provide examplesofa numberoforganizational practices that may effectivelyrepairthecog- nitive shortcomingsof individuals. We call these practicescognitiverepairs. We then discusssix tradeoffs that affect the success of cognitive repairs.We close by consideringhow a cognitive perspective might benefitthosewho studyorganiza- tional learning and those who manageit.

Research inOrganizational Behavior, Volume 20, pages 1-37. Copyright © 1998 hy JAI Press Inc. All rights of reproduction in any form reserved. ISBN: 0-7623-0366-2

1

CHIP HEATH, RICHARD P. LARRICK, andJOSHUA KLAYMAN CognitiveRepairs 3

INTRODUCTION

In afamous speech, Hamletdeclares,“What a pieceof work is man.How noble in reason,how infinite in faculties” (Hamlet, II, 2). An observer whosummarizedthe psychologyof the late twentiethcenturywould probably choose very different phrases to describe the humancondition—perhaps,“What fools these mortals be” (Midsummer Night’sDream, III, 2).

Are people“infinite in faculties” and “noble inreason”?HerbertSimon won a Nobel prize for arguingthat social science mustunderstandthe waysthat human facultiesare limited. Instead of being infinite in faculties,Simon’s humans could be only “boundedlyrational” because theircognitive abilities—their ability to perceive,remember,andprocessinformation—were restricted. Well, then, ifpeo- ple arenot infinite in faculties,a.re they “noble inreason”?Cognitive psycholo- gists havespent30 years examining the actualprocessesthat people usewhen they collect information, combineit, anddraw inferences about their world(Nis- bett & Ross, 1980; Kahneman,Slovic, & Tversky, 1982; Holland,Holyoak,Nis- bett, & Thagard,1986).Insteadof depicting people as “noble” (or magnificent) in reason, this researchhas arguedthat peoplereasonin ways that producesystem- atic errors. A pessimistic modern Hamlet might combine the observations of these two research streams and describe humans as equipped with primitive hardwnre andbuggy software.

However, outsidershave not alwaysacceptedthe pessimistic description of human faculties and reason that is found in the research literature.As one skeptic put it, “If we are so stupid,how did we get to the moon?” (Nisbett & Ross,1980).

How should we resolve the apparent discrepancy betweenthe pessimisticliter- atureon human shortcomings and the optimistic evidence of humanaccomplish- ment? One way is todismiss the laboratory research. Someresearchershave argued that the shortcomings that have beendocumentedin the lab areso minor that they do not constitute mistakes ofany real consequence(Funder, 1987; Cohen,1981).Others havearguedthat individuals areless likely to make errors in natural environmentsthan in contrived laboratory experiments (Anderson,1991; Gigerenzer,1996; Cheng& Holyoak, 1985;Hilton, 1995).

We proposeanotherway to resolvethe discrepancy.Unlike someresearchers, we do not dismissthe examples of limitations, errors, and biases reported in thelit- erature; we assume that individuals arelimited, their decision processes are biased, andthat they often make costly mistakes onimportant decisions. We resolve the apparent discrepancy between evidence of individualshortcomings and the empirical factof moonwalksby observing thatindividualsdid not make it to themoon, NASA did.

Organizationslike NASA may have discovered ways toavoid~orrepair theindi- vidual shortcomingsthat have been documentedby cognitive researchers.Orga- nizations may develop such repairs through deliberate analysis, learn them throughtrial and error, or discoverthem through serendipitous accident.In some

cases, repairsmay derive from formal academicdisciplineslike economicsor sta- tistics (e.g., Nisbett,1992;Nisbett, Krantz, Jepson, &Kunda, 1983; Larrick, Mor- gan, & Nisbett, 1990), but in most casesthey will not: They will be ad hoc, intuitive rulesthat emerge from day-to-daypractice. Our thesis,then, isthat indi- viduals indeed face cognitive limitations and shortcomings, but thatorganizations can provide individuals withnormsandproceduresthat mitigate theirlimitations and reduce their shortcomings.

In this paper we describe a variety of potentiallyseriousshortcomingsthat have beendocumentedin researchon humanjudgmentand reasoning. We focus inpar- ticular on learning and hypothesis testing,that is, how people use information to develop and revise theirmental model of the world.Foreach cognitiveshortcom- ing we discuss, we provide examples of organizational practices thatmay repair this shortcoming. We call these practicescognitive repairsto emphasize the fact that they correct somecognitive processthat wasinitially flawed and in need of repair.

We identify potential cognitive repairs to spurresearchersto considerhow such repairs might look and function.Althoughat thispoint, we canmake only ananec- dotal case forinterpretingcertain practicesas “repairs,” we hope that,by pointing out someplausible examples of such repairs, we willpromptresearchersin both psychologyand organizationsto considermore systematicallyhow such repairs might function.

More generally, the concept oforganizationalcognitive repairsillustratesthat researchers may find interesting relationships betweenindividual cognitionand organizational practice.Theserelationships have not received the attention they deserve.On the one side, research in cognitive psychology has largely treated individual learnersas “rugged individualists”who face a difficult environment alone, equippedonly with their own, flawed cognitive strategies. On the other side, organizational research has largely ignored the literature on individualcog- nition, focusing insteadon issues of motivation orincentives.By studyingorgani- zational sources of cognitive repairs, we bringtogethertwo frequently disparate literatures anddemonstratehow knowledge at onelevel of analysiscaninform the other.

By reviewing individual shortcomingsand identifying potential cognitive repairs, wealsohope tocontributeto the academic andpopular literatureonorga- nizational learning (Epple,Argote, & Devadas, 1991; Deming, 1982; Senge, 1990; Cohen, 1991; Miner & Mezias, 1996).One important means tofacilitate learningby organizations is to develop processes that overcome theshortcomings of individuals within the organization.

Below, we start with a briefexampleof the kinds of repairs that weconsiderin this paper. Then we introduce a framework that describes different stages in the learning process, and we use it toreview the literature onindividual shortcomings and to suggest potential repairs.As a preliminary reply toHamlet,we say that even ifruggedindividualsare unlikely to be infinite in faculties and noble inrea-

4 CHIP HEATH, RICHARD P. LARRICK, and JOSHUA KLAYMAN CognitiveRepairs

son, individualswho have access to organizational and cultural repairs maysome- times appearso.

An Exampleof OurApproach

Consider one studythat might be regardedas an ominous indication of ignoble reasoningby individual experts. Hynes andVanmarcke(1976) asked seven“inter- nationally known” civil engineers to predict the height of a structurethat would cause a foundation tofail; they alsoasked the engineers toset a 50 percentconfi- dence interval around their prediction sothat their confidenceinterval was wide enough to have a 50 percent chance of enclosing the true failure height. Theresults were quitesobering: not one engineer correctly predicted the truefailure height within his or herconfidenceinterval.

Evidently, the civil engineers thoughtthey knew more thanthey did—if they had been awareof the limitations of their analysis,they would haveset widercon- fidenceintervals andwould have predicted the truefailure heightmorecorrectly. In the psychologicalliterature this kind of finding has been labeled“overconfi- dence,”and it is not anaberration.Similarresultshave been observed with anum- ber of individual professionals(e.g., Russo & Schoemaker,1992).In summarizing the evidence,Griffin and Tversky (1992) quipped that experts are “often wrong but rarely in doubt.”

To illustrate why this study paintsan ominous picture of individual reasoning, considerthat (unlessyou arereading this paper outside)you are sitting in abuild- ing that was constructedby civil engineers who were substantiallyless accom- plishedthanthe internationally known experts in thestudy. Your civil engineers made numerous decisions to ensure the stability and safety of yourbuilding; they decidedhow strong to makeits roofsupports andhow stableto makeits founda- tton. If evenexpertengineersare overconfident,shouldyou be concerned about your safety?

The answer, we believe, is no. Fortunately, the engineering profession has developeda particular repair, called“safety factors,” that mitigatethe overconfi- dent reasoningof individual engineers.In an actual assignment civil engineers would preciselycalculatethe amount andstrengthof foundationmaterialsneces- sary to hold a structure of a particular height,then they would multiply their pre- cise answerby a safety factor(i.e., a number between three and eight), and use the larger figure to build the foundation.Were the confidenceintervals of the engi- neerstoo narrow?Yes. Werethey toonarrow by a factor of three? No.

Safetyfactors arean example of thekindof cognitive repair weconsiderin this paper.An organization (e.g., an engineeringfirm or the engineering profession at large) provides its members with a repair thathelps combat a systematic and potentially serious bias inindividual judgment. As a result,the organization shields individualswithin the organization from actingon their flawed decisions,

5

and it shields individuals inside andoutsidethe organizationfrom suffering the consequences.

Whatis a CognitiveRepair?

Organizational repairscanroughly bedivided into two classes: (1) motivational repairs increase theenergy and enthusiasm with which individuals pursue a task and (2) cognitive repairs improve the mentalproceduresindividuals use todecide which task to pursue andhow to pursueit. Organizational research onmotivation and incentives can be regarded as the study of motivational repairs(Milgrom & Roberts,1992; Eisenhardt,1989).Organizations may need to repair motivational problems in order to encourage individuals tolearn(e.g., see Heath,Knez, &

Camerer,1993).For example, individualsmay not be willing toexperimentwith newtasks becausethey have becomeendowedwith the benefits associated with the old task.

Although previous work hasrecognizedtheimportanceof motivational repairs, it hasneglected cognitive repairs. Even when individuals have the right incentives and resources, theymay not learnfrom their experience ifthey use the wrong mentalprocess togeneratehypotheses,collectinformation,anddrawconclusions. The civil engineers who misestimated the stability of the clayembankment.were adequatelymotivatedto getthe right answer. However,they did not on their.own invoke the kind of correctives(e.g., safety factors)that might havemadetheir guesses more appropriately cautious.

REVIEW OF INDIVIDUAL LEARNING AND EXAMPLES OF COGNITIVE REPAIRS

In this section we organize the literatureon learning and decision making around three different stages of the learning process. Effective learners must (1) generate hypotheses that explain thecausalstructureof the world, (2)collectinformation to distinguishamong theirhypotheses,(3) draw conclusions that areappropriateand ~autious. Theboundariesbetween these stages arefuzzy—theyare interrelated ~ndinterconnected(Klayman, 1995). However, we distinguish among them

ecause theyinvolve different psychological processes. Our strategythroughoutthe review is to considerfirst the individual then the ‘ganization. For each stageof learning, we describe howan ideal individual irner might reason, andreview psychological research showing how realmdi- luals depart from this ideal. Then, we describe potential cognitive repairsby ich organizations might correct the individual shortcoming in question.

6 CHIP HEATH, RICHARD P. LARRICK, and JOSHUAKLAYMAN CognitiveRepairs 7

Generating Hypotheses

In the first stage of the learning process individuals must generate hypotheses about the relationships amongevents.Subjectto constraints of time andinforma- tion, individuals should generate hypotheses that are deep(i.e., by considering causesthat are more general or systemic) and broad(i.e., by considering a larger numberof potential causes).However,a greatdeal of psychological researchsug- geststhat individuals develop hypothesesthat are shallowand narrow.

Individuals GenerateHypotheses thatareShallowRatherthanDeep

Individuals Searchfor Explanationsthat MakeThemselvesLook Good

Individuals often conductshallow searches when theytry to explain success or failure becausethey search in a self-serving way(i.e., in a way that allows them to feel good about themselves).In a meta-analysisof 91 tests of this self-serving bias, Mullen and Riordan (1988)show that individuals typically conclude that their successesresulted from stable, internal factors (e.g., ability), butthat their Jailures resultedfrom unstable,environmentalfactors (e.g., thedifficulty of the environment, insufficient effort, or bad luck)(see alsoFiske & Taylor, 1991, pp. 78-82).

How might organizations repair self-servingbiases?Some repairsmay be quite simple: Traderson Wall Street are warned, “Don’t confuse brains and a bullmar- ket” (Odean, 1996).This compactphrase prompts individual traders toconsider the base rate of success inthe market, and it makes it moredifficult for them to indulge in self-servingexplanations for their success.

At FloridaPower andLight employeesdevelopeda new way to fightself-serv- ing biases afteran incident thatprominentlyfeatured aJapaneseinspector for the Deming Prizewho laterbecamea folk hero withinthe company(Walton, 1990, p. 61). To impressthe inspector,FP&L managerstook him to visit a new facility that had been constructed faster and moreeconomicallythan any facility in the history of theindustry.However, the Deming inspector did notsimply accept the results at face value and congratulate themon their “quality” projectmanagement; instead, he asked a number ofquestionsto determinewhythey were so successful. Themanagers’answerswere so inadequate that it soon becameclearthatthey did not understandenough about their success torecreateit in the future. The inspec- tor dismissed their “success” in hisJapanese-accentedEnglish—”you were rucky.” Later on his phrase, complete withaccent,became a common repairfor self-serving interpretationsof success.

The Deminginspectordeflated a self-serving biasby consideringalternative hypotheses for success (e.g., luck rather thanskill or knowledge). Traditionally at FP&L, managers were not questioned as long as they achievedgood results. After this incident managerswere much more likely to be asked to explain their sue-

cesses.If they could notdo so, the verdictwould be delivered: “you wererucky” (Walton, 1990,p. 61). The strategic use of the accent was designed to remindman- agers about theearlier incident whereluck produceddramaticresults that were unlikely to be repeated.

Individuals Focuson PeopleRather thanSituations

Individuals alsogeneratea shallow set of hypotheses because social settings tend to highlight people as causes.In Westernculture individuals typically choose to explain events in terms of people’s actions and traits rather than situationalfac- tors (Gilbert & Malone, 1995; Ross, 1977; Ross & Nisbett, 1991). In a recent study observers heardanother studentgive a pro-life or pro-choice speechon abortion. Afterward, observers assumedspeakersheld attitudes consistent with their speeches even though the speeches were derived fromscripts written by the experimentersand even though the observers themselves told the speakers which position to speak for (Gilbert & Jones,1986).Similarly, Deming (1982) describes a company that used a varietyof flammableproducts in their production process. After analyzing the dataon fires, Deming found that the fires were a stable and predictableoutcomeof the production process.However, according to Deming, the companypresidentfocusedhis attentionselsewhere.He “sent a letter toevery one ofthe 10,500employeesof the company to plead with themtoset fewer fires” (p. 325).

People’s actions are frequently more obvious than their situations.Therefore, when individuals generate hypotheses about whyan event occurred, their first hypothesis is likely to bethat someperson caused it(e.g., Ross & Nisbett,1991). This tendency to focuson people rather than situations has beendocumentedby-so manyinvestigatorsin so many situations that it has beencalledthefundamental attributionerror(Ross,1977; for recent reviews see Ross & Nisbett, 1991; Gilbert & Malone, 1995).

Organizations might repair thefundamentalattributionerrorby remindingiridi- viduals toconsidercauses other than people,especiallythe peoplewho are likely to be closest to anyproblem: front-line workers.For example,an old military adage says, “Thereare no such things as bad troops, only badofficers” (Cohen & Gooch, 1990, p. 228). Parallel repairs are found in total quality management (TQM). Ishikawa says, “whenever mistakes occur, two-thirdsto four-fifths of responsibility rests with management” (Ishikawa,1985,p. ix). Such maximsmay partially repair thefundamentalattribution errorbecausethey encourageindivid- uals tolookbeyondthe front line. On the otherhand,they may simply focus the erroron people at ahigherlevel. Thus, abetter repairmay be one from Deming, who tells managers thatof the problems he has seen, “94% belong to the system” (Deming, 1982, p. 315). Ishikawa and Deming both use vivid statistics toover- comethe fundamentalattribution error even though it isunlikely that either has

8 CHIP HEATH, RICHARD P. LARRICK, and JOSHUAKLAYMAN CognitiveRepairs 9

conducteda preciseempirical analysis.Deming’s “94%” is particularlynotewor- thy because ofits apparent precision.

Individuals StopSearchingasSoon as TheyGenerateOne Hypothesis

Self-serving biases andthe fundamentalattribution error are specialcasesof a much broader tendency:Individuals tend to stopsearchingfor a causeas soonas they locate a plausible candidate hypothesis(Gregory, Cialdini, & Carpenter, 1982;Hoch, 1984).

To counterthis generaltendency,organizations have developedsome repairs that arewidely applicableacrossa numberof domains.In onetechniqueknown as the “Five Whys,” workersat Toyota learned to ask “why?”five times beforethey stoppedgenerating hypotheses. Whenthey did so, they were more likely tofind a root cause ratherthan a superficialone.Imai (1986) illustrates thetechniquewith the following example:

Question1: Answer I: Question2: Answer2: Question3: Answer3: Question4: Answer4: Question5: AnswerS:

Why did the machine stop? Because the fuseblew dueto an overload. Why wastherean overload? Becausethe bearinglubrication was inadequate. Why wasthe lubricationinadequate? Becausethe lubrication pumpwas not functioningright. Why wasn’tthe lubricating pumpworking right’? Becausethe pump axle woreout. Why wasit worn out? Because sludgegot in.

Imai arguesthat by asking “why” five times,workers identified “the real cause and therefore the realsolution: attaching astrainerto the lubricating pump. If workers had notgone through such repetitive questions,they might havesettled with an intermediatecountermeasure,such asreplacingthe fuse” (Imai, 1986, p. 50). Another illustrationof the Five Whys deals directly with thefundamental attribution error: “Problem: He doesn’t manage well. (1)Why? He’s not on the floor. (2)Why?He’s in the equipment room. (3) Why? The newest equipment isn’t working. (4) Why? Purchasing gave the supplier a short leadtime. (5) Why?Poor planning system”(Forum, 1992, p. 54). In general, when individuals ask“why” the first time, they are likely to developanswersthat invoke somesalient, recent, or proximal event (e.g.,someperson’sactions). Subsequentwhys are likely to cause individuals to think more broadly and situationally.

Although the Five Whys is an admirable cognitive repair becauseof its power and simplicity,individuals may find it difficult to executeby themselves.When individuals have one good hypothesis in mind,that hypothesisoften blocks their ability to seealternatives(Gregory,Cialdini, & Carpenter,1982; Gnepp & Klay- man,1992; Mynatt, Doherty,& Dragan, 1993).For example, Hoch (1984) found

that subjectswho generatedpro reasons for buying aproducthad more difficulty generating conreasonsimmediately afterward.

If individuals find itdifficult to generate alternate hypotheseson their own,then organizationsmay repairshallow searchby confronting individuals with others who are expertin asking questionsthat reveal deep causes. AtMicrosoft, Bill Gateshas by personal example, encouraged aculturethat relieson relentlessques- tioning. Says oneWindows manager, “yougo into the meetings andyou come out just sweating because, if there is anyflaw, he will land on it immediately and pick itto bits” (Cusumano &Selby, 1995,p. 25). Employees “overuse” terms borrowed from Gates,like “drill down” asa euphemism for “going into more detail” (“What Bill GatesReally Wants,” 1995).

A similar cognitive repair isfound in theorganizationthat administersthe Dem- ing quality prize. Here, official Deming inspectorsexaminemanagers using a techniquecalled“single-caseborequestions.”They begin with broad exploratory queries andthen relentlessly delve down intoweaknessesand omissions inthe answersthey receive. Single-case borequestionssometimes identifycausesthat arequite deep. For example,FloridaPower andLight often had to deal with power outagesthat occurredwhen a treefell on a powerline and severedit. To improve the reliability of its service,FP&L organized a unit to trim all the trees in sites where damage had occurred, and thus prevent future outages. Managers at FP&L congratulated themselves for creating aprocedurethat preventedfutureproblems. However, the Deming inspectors were not satisfied with theproceduresince~itpre- vented problems only in areas that had already experienced acrisis.Theysearched for a solution at adeeperlevel, and askedmanagersa numberof questionsabout what might beconsideredforestry! What kind of trees grow in theregion?Do palms grow faster or slowerthan oaks? Managersat FP&L realized they did not know the answers to these questions, and thatthey had not searcheddeeply enough tosolve their problems. After their experience with single-caseboreques- tions, FP&L managers consulted with foresters anddevelopeda regularmainte- nanceprocedureto trim trees basedon their growth ratesand acrossthe entire region, notjust in areas wheretreeshad previously severed lines Afterparticipat- ing in sessionsof this kind with the Deminginspectors,managersat the firm learned to ask single-case borequestionsin their own internal discussions, thus institutionalizingthis cognitive repair(Walton, 1990,pp. 57-63).

Individuals GenerateHypothesesthat areNarrow RatherThanBroad

In an ideal world individual learnerswould not only generatedeeperhypothe- ses;they would alsoconsidera broad ratherthan narrow set of potentialhypothe- ses. However, even when individuals generate alternativehypotheses,their “alternatives” often differonly slightly from one another,and all lie within the same general frame.For example,participants in oneexperimentwere asked to considerthe serious parking problem facedby their university, and they were

10 CHIP HEATH, RICHARD P. LARRICK, and JOSHUAKLAYMAN Cognitive Repairs 11

given time togenerateas manysolutionsasthey could (Gettys etal., 1987).Com- bined,participants generated about300 solutionsthat researcherswere later able to classify into about seven majorcategories.One category, for example,sug- gested ways to reduce demand for parking (e.g.,by increasing parking fees) and anothersuggested waystouse parking moreefficiently (e.g.,by segregatingpark- ing slots according tosize). The averageparticipantproposed about11 solutions but these11 solutionsrepresented only about three of the seven possiblecatego- ries. The authors askedan independentpanel ofexpertsto compile a completelist of high-qualitysolutions, andthey used thiscompletelist to assesshow many solutionswere missedby eachindividual. The typical participant missed from70 to 80 percent of the high-quality solutions.However, when asked, individuals believedthey had missed only25 percent.

Evenexpertsfail toconsiderahrQadrangeof alternative hypotheses.Forexam- ple, one group ofresearchersshowed professional automechanicsa “fault tree” that listed a number of hypotheses aboutwhy a car might notstart (e.g., battery, starting system,fuel system,ignition). Some mechanics were presented with a “full tree” that containedsevenspecific hypotheses,otherswere givena ‘~pruned tree” that omittedsome important hypotheses (e.g. the ignition system). The resultsindicatedthat when hypotheses were prunedoff the tree, mechanics did not adequatelyconsiderthem (Fischhoff,Slovic, & Lichtenstein,1978).

How might organizationsrepair narrow search by individuals?Individuals might search more broadly ifthey arecuedto think about a problem from different perspectives.At Sharp, employees are told to be “dragonfliesbut not flatfish.” Dragonflieshavecompound eyes and see things from multiple perspectives at once, butflatfish havelargeeyes that onlylookin one direction (Nonaka &Takeu- chi, 1995).

The “dragonfly” repair exhorts individuals toconsiderdifferent perspectives, but thismay be difficult for individuals to do by themselves. Organizations might repairnarrow search more effectivelyby encouraging individualsto- recruit others who havedifferent perspectives. A good example of this is providedby Bridge- stone Tire,which conducts “kokai watches” togeneratealternative hypotheses for improving workpractices.During akokai watch a group ofup to a dozenpeople, from different areas of afactory, gather for a few hours to watch otherswork. In one four-hour watch adozenpeople identified63 potential dangers with a new machine(Walton, 1990,pp. 200-201).

The kokai watchhas a numberof features that ensure that watchersgeneratea broadarray of hypotheses. First, it mandates alargenumber of watchers (up to 12). Second, it selects watchers from a variety of different areas—in one kokai watchthat examineddie and material changes, watchers included a plantvisitor, a memberof the humanresourcesstaff, a chemist, and a project manager. “The idea was that people could observe a process, even thosewho were strangers toit, with fresheyes,seeing thingsthat closelyinvolved workersmight not” (Walton, 1990, p. 200). Third, it ensures that watchers do notdiscardhypothesesprema-

turely. The watchersare instructed to “write down anything,‘Hey, looks like the guy is walking toomuch,’ or ‘Looks like he’snot handlingthe knife right”’ (Wal- ton, 1990,p. 201). Only after watchers generate hypothesesindependentlyarethe results combined andfiltered.

Otherorganizational procedures also repairnarrow individual searchby ensur- ing that individuals generate hypothesesindependently.For example, when Motorola formscross-functionalteams toevaluate new products,they do not allow employees who have participated in one product team to participate in anotherteam with a similarproduct. This prohibition limitsthe pool of potential team members in acostlyway; evaluation teamsinvolve six to nine people and spendtwo to three months to develop a business plan for the newproduct. How- ever, by consciously disregardingprevious experience,Motorola allows new teams to develop recommendations independently from previous teams. At the same time, Motorolaavoids losing the knowledgeof previous “veterans”—they serve asa “review team” that evaluatesthe recommendationsof the newestteam.2

Other repairsensure that a broad range ofalternativesare consideredsimulta- neously.Somecompaniesdivide aproductdevelopmentteam into competingsub- groups which develop separate projectproposals,and only later recombine to debatethe advantages and disadvantages of theindependent proposals.Again, this strategy is costly because it is redundant.However, it may have advantages becausethe built-in independenceensuresthat different subgroups will approach a problemfrom different perspectives (Nonaka & Takeuchi,1995,p. 14).

CollectingInformation

In the second stage of the learning process ideal learnerscollectinformation to test and revise their hypotheses. Therearetwo main sources of such information: the information that individuals already have in their memory and theinformation that they collect from theenvironment.Both kinds of information have potential flaws, but individuals mightminimizethese flawsif they collectedinformationin a rigorousway. However, learnersdo not always actas thoughthey are awareof the potential flaws in their information—they frequentlycollect only a small, biasedsample.

Individuals OftenCollectSmall Samplesof Information

Individuals often collect only a limited sample of informationbecausethey are constrainedby time or attention.In a classicstudy, Payne (1976) asked hissub- jectsto chooseone apartmentout of a number of alternatives, each of which was describedon severaldifferent dimensions(e.g., rent, cleanliness, landlordquality, noiselevel). Whensubjects chose among only two apartments, theytendedto con- sider all of the information beforethey decided.However, individuals searcheda smaller and smaller percentage of information as more informationbecameavail-

12 CHIP HEATH, RICHARD P. LARRICK, andJOSHUAKLAYMAN Cognitive Repairs 13

able. For example, one subject, whowas decidingamong 12 apartmentscharac- terized on eight different dimensions,looked at only about 25 percent of the information beforemakinga final choice.

It would be reasonable forindividual learners to collect only a smallsampleof

information ifthey performed acost/benefitanalysis and decidedthat collecting a large sample was too costly. However, there is evidence that individuals collect only a smallsampleof information becausethey systematically underestimate the benefits of larger samples. Tversky andKahneman(1971) arguethat individuals typically believe that small sampleswill be quite similar to the population from which theyare drawn. They labeled thisbeliefthe “lawof small numbers” tohigh- light that it contradictsthe statistical “law of large numbers,” whicharguesthat samplescan yield an accurate picture of a population whenthey are sufficiently large. When individuals believe in the law of small numbers,they assumethat any samplewill be sufficient,no matterhow small.

At the extreme,individuals may not collect any information from the external environment becausethey believe that they already have adequate information stored in theirhead.Organizations may overcome thistendencyby encouraging or requiring individuals tocollectlarger samples. This kind of repair ispervasivein writings on TQM. “In promoting statistical qualitycontrol,we have used theslo- gan, ‘Let ustalk with data”’ (Ishikawa,1985,p. 200). At manyTQM companies one of the main principles of the quality effort is “Managementby Fact” (Walton, 1990,p. 37).

And TQM not only talks aboutdata, it provides individuals with toolsthat help themcollectand analyzedata.For example, six of the “SevenTools” of TQM pro- vide ways tocollectdata (e.g., checksheets) or to simplify and display largequan- tities of data(e.g., histograms, scatter plots, Pareto diagrams, control charts) (Deming, 1982;Imai, 1986;Ishikawa,1982, 1985;Juran,1992).

Individuals Collect BiasedSamplesof Information

Individual learners not only collect smallsamplesof information,they alsotend to collect samplesthat are biased (i.e., that are unrepresentativeof the larger world). Consider the commonclaims that “the otherline alwaysmovesfaster” or “it only rains after I wash mycar.” Unless we want to believethat a malevolent

spirit is in charge of such harassment, these examplesdemonstratethat ourmem- oriesdo not store a randomsampleof all waiting times orall rainstorms—weare more likely torememberthe rainstorms thatspoil the finishon-ourfreshlywashed

car. Even whenindividualscollectinformation from theoutsidework1(rath~vtha~i from memory),they do not always attend tothe most relevant andimportantinfor- mation.Below, we discuss anumberof factorsthat might lead individual learners to collect biased samples.

Individuals Only Consider AvailableInformation

As indicatedby the car wash example, individuals often collect biased samples because they collect information that is easilyavailable in memory, for example, because it isespeciallyvivid or recent. The problemis that individuals typically assumethat the information that is availableis also most frequent, probable, and causally important(Tversky & Kahneman, 1973). This assumption is often wrong. Individuals dramaticallyoverestimatethe likelihood of vivid causes of deathlike accidents or homicides, and they underestimate thelikelihood of less vivid causeslike disease or strokes.Individuals estimate that accidents causedas many deaths as diseases andthat homicides wereas commonas strokes. Infact, diseasestake 16 times more livesthan accidents and strokes take11 times more lives than homicides. Individuals also overweight recentinformation. They assume that the most recent flood provides an upper boundon possibleflood dam- age, and the purchase of earthquake insurance “increases sharply after aquakeand then decreases steadilyas memories fade”(Slovic, Fischhoff, & Lichtenstein, 1982,p. 465).

Many organizationsrepair individuals’ tendencyto rely on biased,available informationby instituting a process that collects information more systematically. At a Motorola division that develops equipment forcellular phone systems, one group realizedthat an availability bias was causing it to overlook certaincustom- ers when it evaluated new products. The unit assigned account managers only to large accounts, so when managers evaluated newproducts,they primarily consid- ered the needs and requirements of only largecustomers.However, the unit also serveda numberof smaller customersthat did not have theirown accountman- ager. Together,these small customersaccountedfor a largepercentageof reve- nues. Motorola overcame theavailability bias by developing a Feature PrioritizationProcess; theysurveyedcustomersup to four times a year andthen weightedall of the inputsbasedon customervolume andpriority.3

Hospitals also have a variety ofproceduresto force individuals tocollectinfor- mation more systematically. Trauma physicians are often confrontedby vivid but potentially misleading information. One doctor states that, contrary to what one might expect, stabbings and bullet woundsare “relatively straightforward affairs” becausethey leave“clear trackson the body.” Other injuries are more difficult to treat becausethey leave no visible cues. “It would be all too human to focuson a laceratedscalp—agory but basically insignificantinjury—and miss a fractured thighbone that hadinvisibly severedamajorartery” (Rosenthal,1994,p. 48). The medical profession has developed a series of strict protocolsfor traumasituations that allow doctors toquickly collect all the relevant information, not just that which is salient. For example, when a patient firstentersthe emergencyroom, physiciansfollow the “ABCs”; they establishairway, then breathing, thencircu- lation.4 For situationsthat are more critical, such ascardiacemergencies,proto- cols are even more rigorous and specific.

14 CHIP HEATH, RICHARD P. LARRICK, and JOSHUA KLAYMAN 15 Cognitive Repairs

If individuals tend to focuson information that is highlyavailable,it is notter- ribly surprising that they are frequently unaware of missing information. Even when information is present, learnersdo not pay as much attention to whatdoesn’t happenas what does (Agostinelli, Sherman,Fazio, & Hearst, 1986; Newman, Wolff, & Hearst,1980).

Certain professions andorganizationshave learned to repair the tendency to ignore missing information. Homicide detectives learn to notice the absenceof itemsat murderscenes, sincemany murderers takebacksomething that belongs to them after committingthe crime.“You look atwhat’s beentaken andyou find out who it belongedto originally” (Fletcher,1990,p. 75).

A particularly importantform of missing information is the absence ofexperi- ence with highly unusualevents.Bank examiners rarely see a bank fail,nuclear techniciansrarely see a meltdown, airline personnel rarelywitnessa crash(March, Sproull, & Tamuz, 1991; Perrow, 1984). Certain organizations institutionalize proceduresthat encourage individuals to pay attention to suchinformation:desjtito the factthat such eventsare unlikely to beavailablein their own experience. For example,at the Federal Reserve Bank, which certifies the security of banks, senior bank examinersdeliberately recountstoriesof failed banks to keepjunior exam- iners aware thatthey should bevigilant.5 At one bank’s commerciallending department, senior creditofficers would hold seminars and informal brown-bag lunches todiscusspast lending mistakes, particularly in areas characterizedby unusual or rare events(e.g., “problems with highly leveraged companies, real

6 estate, environmentalliability on contaminatedproperty”). By forcing individu- als to rehearse suchinformation,organizations help individuals learn fromvicar- ious experiencesthat are rare but highlyinformative. Furthermore,organizations remind individuals of potentially painful information that self-serving biases would make thempreferto ignore.

IndividualsCollect BiasedInformation Basedon Their PreexistingTheories

Researchsuggests that individuals tend to think of “facts, experiences, and arguments thatsupporta current hypothesis more readilythan those that refuteit” (Klayman, 1995; see also, Baron,1988; Kunda, 1990; Nisbett & Ross, 1980). Thus, whenindividuals collect information from memory,they may focus on information that supports their preexistingtheories.Individuals may also do this whenthey collectinformation from the externalenvironment.For example, when individuals collect information fromothers,they often askspecific, directive questions thatare likely to elicit the answer they expect (Hodgins& Zuckerman, 1993; Zuckerman,Knee,Hodgins, & Miyake, 1995).

The Chicago Board of Trade has astaffof in-houseinvestigatorswhoscrutinize tradesthat may violate exchange rules. Intheseinvestigations,which areobvi- ously quite sensitive,it is veryimportantthat investigatorsdo not collectinforma- tion that is biasedby their initial theories.To repair thistendency,the investigators

are trainedto avoid questionsthat canbe answered with a yes or no response. “This forcesan investigator to askopen-endedquestionsand allows her to draw out as much information aboutthe situation as possible.” Byasking open-ended questions,investigatorsavoid the possibility of directing the interview in a way thatelicits only information that is consistentwith their preexistingtheories.7

Someorganizationshave developed maxims that seem designed to encourage individuals tocollectdataratherthan relyingon their (potentially biased) theories. At BridgestoneTire employees use two Japaneseterms: genbutsu(actual product) and genba(actual place)(Walton, 1990,p. 194). Theseterms remind employees not to rely on their own theories, but to actuallygo out andinvestigatethe actual product in the actualplace where the problemsarose. Another group (Forum, 1992)uses a similar cognitive repair they call thethreeactualrule: (I) Go to the actual place; (2)Seethe actual problem;(3) Talk to the actual peopleinvolved.

Individuals ConsiderOnly Part of the RelevantInformation

Finally, individual learners maycollectbiased samples because they tendto col- lect information fromonly one small corner of the universe of information. This arises from basic cognitive processes. Memory isassociative—whenindividuals retrieve onepiece of information, theytend to think of other information thatis linked to it by strong associations,commonfeatures,or similar meaning. Even when individualscollect information from the external environment,they are likely to collect informationbasedon the samekind of associative process. Researchin cognitive psychology hasshownthat individuals attend to andprocess information more comprehensively when they have a mentalschemathat tells them what information is needed in a given situation and where to findit (Ander- son, 1995).

Accordingly,organizationscan repair biased informationcollectionby provid- ing individuals with aschemathat reminds them of the full range of relevantinfor- mation.Many schemas of this kindcanbe found in thefinancialservicesindustry, where individuals must assess a wide varietyof information todeterminewhether to buy, sell, or lend.At the FederalReserveBank of NewYork, the BankExami- nationsgroup protects the FDIC insurance fund by ensuring that individual banks are in soundfinancialcondition.Whenreviewing eachbank,examiners use arat- ing system knownas CAMEL: they review Capital adequacy,Assetquality, Man- agement, Earnings, andLiquidity.8 In another bank’s commercial loan department,creditanalysts use the “Five Cs of Credit”:Collateral,Capacity,Cap- ital, Conditions, andCharacter.9

Organizational schemaslike CAMEL and the Five Cs are likely to encourage individuals tocollect a broaderrangeof information than they wouldnormally collect.It wouldbe very easyfor individual learnersto collectinformationon only on themost salientfactors(such as cashflow in a loan decision).Although cash flow is certainlyimportant, it canalsobe misleadingor unreliable,particularly in

16 CHIPHEATH, RICHARD P. LARRICK, andJOSHUAKLAYMAN Cognitive Repairs 17

an environment where conditions are changing. By emphasizing the FiveCs, a bank canrepair the tendencyof individual analysts to neglect informationabout importantvariablesthat are less obvious orare harder to assess. For example, the Five Cs reminds loanofficers to considercharacter—Whatare the management skills of the owners? Dothey have good personalcredit records?Although the answersto such questionsare quite important, individualanalystsmight forget to ask them in a numbers-oriented environmentlike a bank, without acognitive repairlike the FiveCs.

Individuals WhoCollect BiasedInformation Fail to Correctfor Bias

We have discussed a number of factorsthat might lead individual learners to collect biased information.However, even if learnerscollectbiased information, they might still be able to draw effective conclusion as long as theyrecognizedthe bias and corrected forit. Forexample,supposean individual made the statement, “the otherline alwaysmovesfaster,”but thenremindedherself that suchsituations might be overlyavailablein her memory.This kind of correction improves the conclusions drawn from even a biased sample.On the other hand, even ifindivid- uals are aware thatthey have collected biased information,they may not know how to correctfor biases after thefact. For example, after individuals ask biased questions and therefore receive biased answers,they do not take into accounthow much theanswerswere biasedby their initial questions (Zuckerman, Knee, Hod- gins, & Miyake, 1995).

Because individualsdo not always correcttheir information forbiases,some organizationsattempt to ensure that individualscollectunbiased samplesfrom-the start. Microsoft requires itssoftware developersto use the sameprogramsand machinesthat areusedby their customers.For example, programmers whowere developingthe new Windows NT operating system ran the current day’sversion

of the programasthey programmed the next day’s version. AtMicrosoftthis pro- cess isknown as “eating yourown dog food.” It ensuresthat developers collect a large, unbiasedsample of information about the current state of the program.If Windows NT crashed while a developer wasdesigninga new printerdriver, he had tofix the problem with NT before he could return to his driver (Cusumano & Selby, 1995,p. 331). Microsoft also requiresdevelopersto use the same machine

usedby customers,a requirement thathas been “controversial at times” because developers like to have thefastest, coolest machineson their desks. However, when developershavebettertechnology than theaveragecustomerthey collect biased information abouthow well their software programsperform.One man- ager said,“every time I’ve had a project where thedevelopershad hardware that was a generation beyond what customers had, the [software] always hadperfor- mance problems” (Cusumano &Selby, 1995,p. 347). Byrequiringdevelopersto use the same machines as their customers,Microsoft forces them to collectan

unbiased sample of information about theoperatingspeed and memory demands of the softwarethey are developing.

DrawingConclusions

After generating hypotheses and collectinginformation,ideal learners should evaluatethe information they have collected and drawconclusionsthat areappro- priate andcautious.Researchers have suggestedthreemain classesof problems that real individuals facewhenthey interpretevidence. First, they often weigh information in a way that is not statisticallyappropriate—forexample, they emphasizethe importanceof extreme evidence butthey do not emphasize therel- ative amount of extreme versus non-extreme evidence. A second, even more insidious problem is that individuals use their initial theoriesto interpretthe evi- dence. While individuals may readily accept information that is consistentwith their initial hypothesis,they cast a critical eyeon information thatcontradictsit. Third, as a resultof the two previous processes andothers,individuals frequently drawconclusionsthat are overconfident and overly optimistic.

Individuals WeighVivid and Extreme EvidenceMore Heavily

Once individuals have collected information,how shouldthey combine it and weigh it? An ideal learnerwould weigh informationbasedon the quality of the information.However,actual learnersdo not alwaysassignappropriateweights to all aspectsof the decision.For example, they tendto weigh more vivid, easily imagined information more heavily(Keller & McGill, 1994).They also focuson the extremity or strength of theavailableinformation (e.g., the warmth of arec- ommendation letter) without adequately attending tothe amount or weightof the evidence(e.g., the writer’s amountof contactwith the recommendee)(Griffin & Tversky, 1992).

If individuals tend tooveremphasizevivid or extremeinformation, organiza- tions might prevent thisby requiring individuals toconsciouslyclassifyinforma- tion according to its appropriate weight. Manycompanieshave internal audit groupsthat examine the records of company divisions to ensure that they areusing proper accounting proceduresand spending moneyon legitimate expenses.An audit usually uncoversa variety of major andminor “exceptions”(i.e., situations where correctprocedureswere not followed). One auditor saysthat auditors must be careful not to “place too much emphasison memorableerrors, e.g., an error in the president’s expensereportor the misuseof the companycar.” One auditing group repaired this temptationby first classifying eachexceptionas major or

in minor then consciously ignoring theminor issues.

Consistent with the tendency tooverweightthe extremityof information and ignore the amount, individuals frequentlyplace higher weighton one vivid case than on a much larger sample of information. Joseph Stalin is reported to have

18 CHIP HEATH, RICHARD P. LARRICK, andJOSHUA KLAYMAN CognitiveRepairs 19

said, “The death of a single Russian soldier isa tragedy.A million deathsis a sta- tistic” (Nisbett & Ross,1980,p. 43). In a studythat supports this observation,Bor- gida andNisbett (1977) showedsome students a statistical summary ofhow dozensof students hadratedvariouscoursesin the previous term. Other students attended a panel discussion, during whichtwo or three upper-division students rated eachcourseon a numerical scale and providedsomegeneric, uninformative comments. Despite the factthat the statistical summary provided students with a larger amountof information, individuals who heard the smallsample of vivid information were more likely to change thecoursesthey selected.

Microsoft also discoveredthat individualsdiscountlargesamplesof statistical information.At one point, Microsoft started surveying users to seehow many of them found it easy tousea particularfeature. Softwaredevelopers often refused to believe the statistics.“The usability groupwould tell the development group‘Six out often couldn’t dothis.’ And thedeveloper’sreactionwould be, ‘Where’d you find six dumbpeople?”’(Cusumano &Selby, 1995,p. 379).In order to repair this tendencyto ignore base rateinformation,Microsoft made the information more vivid. It built a “usability test lab” where developers can watch real usersstruggle with new productsfrom behind a one-waymirror. Instead of presentingdevelop- ers with pallid statistics, the test lab presents them with real people(albeita mxtch smallersample).The lab managersaysthat when developers see auser, “twenty ideasjust immediately come tomind. First of all, you immediately empathize with the person. Theusual nonsense answer‘Well, they canjust look in the manual if they don’t know how to useit,’ or ‘My idea is brilliant;you just found ten stupid people’...that kind of stuffjust goes out thedoor...” (Cusumano &Selby, 1995,p. 379). This cognitive repair is interesting because it uses onekind of bias (over- weighting of extreme, orvivid information) to fight another (underweighting of statisticalinformation).

IndividualsUse Their PreexistingTheoriesto Interpret the Evidence

Individuals not only weigh information inappropriately,they also havediffi- culty interpreting information independently of their preexistingtheories.Instead of using the information to test theirtheories,they use their theories to test their information. This often leads them todiscountinformation that disagrees with their preexistingbeliefs.

In a classic demonstration of such discounting, Lord, Ross, and Lepper (1979) selectedundergraduateswho strongly supported or opposed capital punishment and presented them withtwo purportedacademic studiesthat evaluated capital punishment’seffectivenessusing very different methods. A study using one methodfound that capital punishmentwas effective and a study using the other methodfound it was ineffective (the researcherscounterbalancedwhich method was associated with whichresult). Participantsapplaudedthe positive aspects of whichevermethodsupported theirown preexistingtheory, andthey critiqued the

“design flaws” in the other. In fact, after receivingmixed results fromthe two studies, subjectsbecamemoreconvincedof the validity of their original position. Seemingly, they regardedthe evidence as “one good studythat supports my beliefs, and one lousy studythat draws the wrong conclusions.”Individual sub- jects thus failed to evaluate the incoming information separately from theirpreex- isting theories. Unfortunately, similar results have been notedwith professional scientists(Mahoney,1976;Koehler, 1993).

One bank helped itsloan officers repair the waythey interpret evidenceby encouragingthem toconsidera nonstandardtheory of lending.In mortgagelend- ing, loan officers oftenlookfor reasonsto deny loans because loans are difficult to make (they are subject to a mountain ofregulations)and potentially quite costly (e.g.,foreclosureon a badloan maycostup to 20% of the property value). Thus, the initial hypothesis in manyloan decisions isthat an applicant should be denied a loan unless proven otherwise. One mortgageloan departmentgrew at an annual rate of30 percentby forcing loan officers toconsideran alternative to the standard hypothesis.Instead of asking whetheran applicant should bedenieda mortgage loan, it asked whether theapplicantshould beapproved.This reversalled the departmentto develop specialprogramsfor qualified applicantswho had low incomes or other specialcircumstances.11

Individuals use their theoriestodevelopexpectationsabout what is normal, and they frequently label unexpected events as “problems” or “failures.” These labels may be misleading,however, particularly in research anddevelopmentwhere unexpected events may point the way to importantbreakthroughs.One research organization has developed a repair that discourages individuals from thinking that unexpected events are failures(Sapolsky,1997).JacksonLaboratoriesbreeds mice that exhibit physiologicalor behavioraltraits that are of interest to medical researchers.For example,it sells mice that lack growthhormoneto researchers who are interested inunderstandingthe biology of mammalian growth. It found that the animal technicians (e.g.,the peoplewho cleanedthe cages) often noticed unusual behaviorthat was scientifically important. The mice that lacked growth hormone were discoveredby a technician who noticed aparticularmouse that didn’t grow at a normal rate. Another technician noticed a mousethat didn’t respond normally to the loudnoisesthatoccurredwhen the cages werecleaned-- its offspring were found to be susceptible tohereditarydeafness.After several experienceslike this where unexpectedbehaviorproducedimportantdiscoveries, the company started holding regular meetings with animaltechniciansto inquire whetherthey have spottedanything unusual.Theseforums for highlighting the importanceof unexpected eventsare called“deviant searches.”

CRSS,an architecturalfirm, developeda special position to repair the problem of theory-based interpretationof evidence.“Most designerslove to draw,to make ‘thumbnail sketches’,”says one manager, but this rush todraw conclusionsis often premature.CRSScreateda uniquejob description,the “programmer,” to ensurethat somemembers of its design teams were not allowing theirown theo-

20 CHIP HEATH, RICHARD P. LARRICK, andJOSHUAKLAYMAN Cognitive Repairs 21

ries to dominate the waythey evaluated information from clients.Programmers arenot in charge ofdesigningor problem solving,they are in chargeof “problem seeking.” They are trained touse techniques that help them to resist premature conclusions,and thus listen more carefully to clients. “The experienced, creative I programmerj withholdsjudgment,resists pre-conceived solutions andthe pres- sure to synthesize...he refuses to make sketches until heknows the client’sprob- lem” (Peters,1992,p. 402).

Often, organizationsensure that individuals weigh information effectivelyby forcing them to interact with others who might weighthe informationdifferently. One researcher hasexploredwhether training as a scientist cures the problemsthat other individuals have in evaluatingevidence(Dunbar,1995).The answer isno. For example,scientists,especiallyyoung ones, often believethat a singleexperi- mental result hasjust resolvedan important problem. However, when Dunbar studied aset of microbiologylabsthat had been particularly successful, he found that they placed more emphasison group lab meetings. At thesemeetingsanindi- vidual scientist presented his or herresultsto a varietyof skeptical,uninvolved peers.Whenthe individual scientist presented a striking new piece of evidence (e.g., I have detectedEnzymeZ in a biological process where ithas never been observedbefore), theindividual’s peers were typically quite willing to propose alternate ways of interpreting the evidence (e.g.,the samplewas contaminated with residual Enzyme Z from a prior procedure). In successful labs, even when individual scientists failed to weigh a particular piece of evidence appropriately, their peersdid so for them. Moreover, the most successful labs werethose that includedmemberswith different training and backgrounds.Such “lab meetings” are not limited to successfulmolecularbiology labs; similar meetings takeplaceat venture capitalfirms wherefirms decide whether to allocate money tonew ven- tures (Kaplan,1995).

IndividualsDraw Conclusions thatare Overconfidentand Overly Optimistic

Imagine that individuals havegenerateda set of hypotheses, collectedsome new information, and interpreted the relevanceof the information for the initial hypotheses. How much confidence should they place in the conclusionsthey have drawn?If individual learners were adequately cautious, their conclusionswould reflect the degree ofuncertaintyin the dataon which they are based.Over the years, researchhas documentedthat individuals often express morecertaintyin their conclusionsthan is warrantedby the factsavailable to them (or by their actual performance). This kind of problem has beendocumentedextensively in laboratorystudies,but also infield studies of individualjudgmentin a variety of professions,like the civil engineers in theintroduction(Griffin & Tversky, 1992; Lichtenstein,Fischhoff, &Phillips, 1982; Russo & Schoemaker,1992).

Individualsoften exhibit aparticularkind of overconfidencethat we mightlabel a planningfallacy (Buehler, Griffin,& Ross,1994)or anoptimismbias. Thisopti-

mism bias ispervasivein work environments. Software developers at Microsoft often experience burnout becausethey “grossly underestimate”how long it will takethem to accomplish certaintasks (Cusumano &Selby, 1995, p. 94). Organi- zationsdo not always successfully overcome this individual bias. A study ofpio- neerprocessplantsrevealed that thetypical plant experienced actual construction costs that were almostdouble the originalprojections;similarly, a study ofstart- ups showed that morethan 80 percentfell short of their projected marketshare (Davis, 1985). Theseexamples suggest that individualsdraw conclusionsthat underestimate the amount of uncertainty and error in their predictions, butthey tend todo it asymmetrically—theyrarely overestimate a project’s cost or time to completion.

Of course, individuals maydisplayan optimismbiasbecausethey confrontmis- alignedincentives.Perhaps if engineerscorrectlyestimated the true cost of a new energy plant, decision makers might choose not to buildit. However, the real causesof theoptimismbias seem to be cognitive, since individuals areoverconfi- dent by the same magnitude even in labexperimentsthat rewardaccuracy.For example,individuals typically assumethat their predictions are more precisethan they are.Whenthey are askedtoset confidenceintervals around a quantity, so that theirconfidenceinterval has a98 percentchanceof including the true answer,they are typically surprisedby the true answer not2 percent of the time, but20 to 50 percent (Lichtenstein, Fischhoff, &Phillips, 1982; Russo & Schoemaker,1992).

How might organizationsrepairindividual tendencies towardoptimism bias and overconfidence? One strategy is to allow individuals to makeoverconfident predictions, thenadjustthem overtly. This was the strategypursuedby the engi- neering profession withits safety factors. Microsoft uses asimilar strategy tocor- rect the overly optimistic projections of individual softwaredevelopers:It has rules about the amountof buffertime that should beaddedto projects.For reason- ably well-understoodprogrammingchallenges,suchas applicationsprogramslike Excel andWord, Microsoft typically adds buffer timethat constitutes30 percent of the schedule.However,for operatingsystemslike Windows, where developers mustcreatea system thathas to mesh effectively with numerouspiecesof hard- ware and software, Microsoft may add buffertime that reaches50 percent (Cusumano &Selby, 1995).Similar repairs have evolved inotherindustries. At one Big Six accounting firm, where teams mustprepareformal plans for acon- sulting engagement,projectleadersdevelop their best estimatesof time, expense,

t2 andcontingencycosts, thenincreasethe finalnumberby 15 percent. This repair has evolved despite the fact that this environmentprovidessome incentives to underestimatecosts—bidsthat are too highmay not be accepted.

WhenMicrosoft adds buffer time to a schedule, it corrects the predictions of overconfident individualsby overriding them.However, it has also developed proceduresthat help individual developers decrease their initiallevel of overcon- fidence. For example, the company has improved itsschedulesby requiringdevel- opers tocreatea detailed work planthat specifies which tasksthey will perform

22 CHIP HEATH, RICHARD P. LARRICK, andJOSHUAKLAYMAN Cognitive Repairs

during specific windowsof time. Says one manager, “The classicexampleis you aska developerhow long it will take him to do something andhe’ll say a month, becausea month equalsan infinite amount of time.And you say, ‘Okay,a month has22 working days init. What arethe 22 thingsyou’re going to do during those 22 days?’ And the guy will say, 0h, well, maybe it will taketwo months.’ Even by breakingit down into22 taskshe realizes,‘Oh, it’s a lot harder than I thought”’

(Cusumano &Selby, 1995, p. 254). Some organizations repair overconfidenceby forcing individuals to interact

with others whoare trained to question their conclusions. For example,the Penta- gon for manyyears had what they calledthe “murder board,” a group ofexperi- enced officersthat reviewed theplans for importantmissions, withthe goal of killing the mission. According to Pentagonlore, the failed Iranian hostage rescue during theCarteryears wasnotvettedby this boardbecausehigh governmentoffi- cials were too concerned about securityleaks.13

Other organizationshave developednorms of frank feedback to ensure that individuals question others’ conclusionshonestlyand openly.In its featureanima- tion unit, Disney regularly holds “Gong Shows” where personnel(including departmentsecretaries)canpitch ideas to a group ofseniorexecutives. Gong Shows mayattract40 people whopresenttheir idea to theexecutivesand other presentersfor three to five minutes. The seniorexecutivesare careful to give exceptionallyfrank feedback atthe end of the session,highlightingbothgood and bad aspectsof the presentations. “Somebody may have agreatconcept, but the story may not be very good. [Wecan’tsay] ‘Oh, that’s fabulous. Great pitchguys!’ and whenthey leave, mumble,‘That was awful!’....We don’t pull our punches. [Eventually] people begin tounderstandthat no matterhow good,bad,or indiffer- ent theidea, it can be expressed,accepted, and thought about”(McGowan, 1996).

GENERAL DISCUSSION

In this paper we have reviewed the literatureon individual learningusing a simple frameworkthat considersthreebroad stagesof the learning process.We argued that ideal learnerswould generate a broad and deepset of hypotheses,test them with a large, unbiasedset of information,and draw conclusions in a cautious and balancedway. Thepsychologicalliteratureindicates, however,that realindividu- als are not ideal learners; they think and act in ways that reduce their ability to learneffectively.

Fortunately, individual learners do not have togo it alone.We have argued that organizationsfrequently repair theshortcomingsof individual learners through the use of sayings, informal routines, and formal procedures.We believe the examples we have offered illustratethe tremendouspromise of organizational sources of cognitive repairs.

23

Nevertheless,we do not thinkthat cognitive repairs will overcomeevery indi- vidual problem. Cognitive repairs are heuristics—like themental processesthey repair,they arepragmaticand often efficient, but alsoapproximateand inexact. For example,they maysolve 75 percent of individualshortcomingswhile incur- ring only one-thirdof the costs of optimalprocedures(e.g., fromeconomicsor sta- tistics). However,they are unlikely to beperfect.

Consider thefive whys. It undoubtedlyprompts individuals to think more deeply about causes, but it is only a roughheuristic.Why five questionsand not three orseven?And which questions? “Problem: He doesn’t manage well.” (1) Why? He doesn’t manage conflict well. (2)Why? He grewup in a dysfunctional family. (3) Why?His parentswere alcoholics...” In this example, the answers took an unhelpful detouraway from potential solutionssometimearound answer 2.

Even when repairs are reasonablyeffective, they may still leave roomfor fur- therrepair.Consider,for example, the military’s partial repairfor thefundamental attribution error: “There are no bad troops,only bad officers.”This adage may repair tendencies to attribute blame tothe people whoareclosest to a problem (the troops who areon the battlefield);however,it merely focuses attentionon another group of people. Thus, itmay prevent individuals from fixing systems orproce- duresthat have basic flaws (Cohen &Gooch, 1990).A more effectiverepairmight say, “There are no bad people, only bad systems.”

Otherrepairsmay be imperfect becausethey fix one problem well, butexacer- bate others.For example, the Five Cs may help individual loanofficers collect more kinds of information but theymay createsecondaryproblems.First, by emphasizing character, the Five Cs may provoke thefundamentalattribution error. Second, although theyexpandthe set of factors loan officers willconsiderin a loan decision,they may also institutionalize any tendencythat they may have to ignoreotherpotentially relevant factors. Third, they may help loan officerscollect information, butthey do not necessarily help theminterpretit. They seem toindi- cate that each C should beweightedequally, whereasan ideal statisticalmodel would weigh someCs more heavilythan others.

As these caveats illustrate, cognitive repairs areunlikely to completelyrepair the shortcomingsof individual learners.Thus, when we assesswhethera given cognitive repair is successful, we mustconsiderthe costs and benefitsof the repair. Below,we considersix dimensions that may affect the costs and benefitsof repairs, and therefore their success.

Tradeoffs Associatedwith Successful Repairs

In order to be successful, a cognitive repair must be effective—it must mend

some individual shortcoming and improve learning relative to the status quo.To be truly successful,however,a cognitive repair must also beacceptedin the orga- nization and actively used. A repair that is not implemented is not arepair.

24 CHIP HEATH, RICHARD P. LARRICK, andJOSHUA KLAYMAN CognitiveRepairs

Cognitive repairsare a kind of innovation,and as such, theiruse will undoubt- edly be affectedby many of the characteristics that havepreviously beenmen- tioned in literatureson diffusion andadoption (Rogers,1995; Scott, 1995).We will focus on innovation characteristicsthat are particularlyrelevantfor cognitive repairs. Cognitiveshortcomingsnot only createthe need for a repair,they also limit what repairsmay succeed.

Below, we consider six dimensions that affect whether a repair will be success- ful: simple versuscomplex,domain-specific versus domain-general,familiar ver- sus novel, corrective versuspreventative,social versusindividual, and top-down versusbottom-up.(We will typically focus on the endpointsof these dimensions, but they shouldbe regardedas continuousrather thandichotomous.)Most dimen- sionsinvolve tradeoffs.For example, qualitiesthat make a repair more effective in solving an individual shortcoming sometimes reduce the chancesthat it will be acceptedand usedby individuals.In the absenceof perfectly effective andaccept- able repairs, we mustrecognizeand understandthe tradeoffsthat make repairs more or less successful.

Simple versus Complex

One obviousdimensionalong which cognitiverepairs vary is whetherthey are relatively simple or complex.Many of the repairs we have discussed in this paper are strikingly simple—they require an individual torememberand apply aproce- durethat is only a fewstepslong (e.g., the five whys or the physiciansABCs). In contrast,many of the proceduresthat are taught as formal repairs in academic environmentsare quite complex, andinvolve many stagesof sorting, arranging, and calculating(e.g., formal financial or statisticalanalysis).

Simple repairshave profound advantages over complex repairs. First,they are more likely to be used because thecostsare small; individuals will find iteasierto learnand implementshorterprocedures. By contrast, complex repairs typically require extensive backgroundknowledgeand tax basic cognitive resourceslike attention and memory (Bell,Raiffa, & Tversky, 1988; Nisbett & Ross, 1980). Thus, when individuals encounter a complex repair, they are likely to perceive the costs of learning it as large and immediate, and the benefits of using itas small, uncertain, and delayed.

Second, simple repairs are easier torememberand reconstructthan complex repairs, and this increases the probability that individuals will accurately apply them and accuratelytransmitthemto others. Because complex repairs require individuals to remember a number of stages,they are more likely to be distorted whenthey are transmitted from individual toindividual. This problem will bepar- ticularly pronounced in situationsthat requirelearningby observation andimita- tion (DiMaggio & Powell, 1983). Individuals who learn a repair through observationmay find it difficult to infer the completerules of behaviorfor com-

25

plex repairsbecauseinformation about the rules is incomplete, unavailable, ordis- tributed across time in a waythat makeslearningdifficult.

Although simple repairshave profound advantages overcomplexrepairs, they also havesome disadvantages. Fundamentally, thetradeoff between simple and complex repairs is atradeoffbetween ease ofuse and accuracy.Complexproce- dures are oftencomplicatedbecause they attempt to be precise. Simple repairs gain easeof useby sacrificing accuracy.For example, a simple aphorism suchas “don’t confuse brains and a bullmarket” suggests the correct direction toadjust one’sjudgment,but providesno guidance aboutexactlyhow much one shoulddis- credit the success of an individualtrader. To precisely estimate theamount of credit due to brains versus themarket,an individualwould have to perform a more complex procedure,suchascalculatingthe overall market performance andmea- suring an individual’s performance relative to the dispersion and centraltendency of the market.

Domain-Speczflcversus Domain-General

Cognitive repairs also vary in the rangeof contextsto which they canbe applied (Nisbett, 1992), with some repairs being relatively more domain-specific and somebeing more domain-general.Domain-specificrepairs are tailorednarrowly for a specific context(e.g., the FeaturePrioritization Process at Motorola or the Five Cs of Credit). Domain-generalrepairs aredescribedso generally, and abstractlythat they apply across mostjudgmenttasks(e.g.,the Five Whys or most economicor statistical principles).

Domain-specific rules haveat leasttwo advantages overdomain-generalrules. First, individualsfind it easier torecognizethat a domainspecific rule is relevant becausethe situation itself remindsthem to usethe rule (e.g., acreditanalyst who has learned to think aboutthe Five Csof Credit will find it difficult to think about lending decisionswithout consideringall five categories of information). Second, individualsmay find it easierto apply domain-specific thandomain-generalrules. Consider, forexamplea loan officer who istryingto apply a general rulelike “cal- culate the netpresentvalue (NPV)” of makingthe loan.This domain-generalrule appliesto many more financial decisions thanthe Five Cs; but it contains no hints about how it should be applied to aloan decision. In contrast, the Five Cs pointout specificaspects of theloan decisionthat might affect the loan’squality. Similarly, securities traders mightfind it hardto benefitfrom a domain-generalwarning against self-serving biases(e.g., “pay attention tosituational determinantsof suc- cess, and don’t over-attribute achievement topersonalcharacteristics”).In con- trast,they are unlikely to miss the point of a moredomain-specificwarning not to confuse brains and a bullmarket.

Although domain-specificrules have advantages, they also have limits.Their specificcontentwill make them more likely tospreadwithin their domain, but it may also preventthem from spreadingacross domains.For example,engineers

26 CHIP HEATH, RICHARD P. LARRICK, andJOSHUA KLAYMAN Cognitive Repairs 27

have safety factors and softwaredevelopershave buffer time, butknowing about safetyfactors does notautomaticallysuggest the need forbuffertime. And even a single individualmay use a rule effectively in one domain butfail to-seeits impli- cations foranother.Auditors areoften quitegood at ignoring theirpreexistingThe- ones about aclient’s financial health whenthey investigateauditing problems. However, they arelesslikely to do so whenthey confront similar problemsoutside

the auditing domain, even if the problem relates to their other professionalactivi- ties (Smith & Kida, 1991).

A second limitationof domain-specific repairs is that theyaretightly tailored to fit a particular task andenvironment.Because of this tightfit, they maybe less successful thandomain-generalrepairs when the task environment is in flux. A buffer factor designed during a specific period oftime—”multiply all time-to- delivery estimatesby 1.5—maylose its effectiveness whentechnologicalor eco- nomic conditions change. Considerthat Microsoft had to develop separate buffer factors to repair overconfidencein applications andoperating-systents?.In general, domain-specific rules will be helpful incompaniesor divisions where the tasks and environmentsare stable over time, whiledomain-generalapproacheswill be helpful in situations where tasks and environments change frequently(e.g., at higher levels in an organizationwhere tasks and decisions areless routine).

A potential method of combining the advantages of domain-specific and domain-generalrules maybe to give individuals a domain-specific repair and then train them to generalize that repair to otherdomains(Fong & Nisbett, 1990;Lar- rick, Morgan,& Nisbett, 1990).Individualstypically find it easier to generalizeby analogy from onespecificinstance toanotherthan to map fromgeneraLprinciples down to specifics (Bassok,1990).For example, peoplewho learn to ignore sunk costs in financial investments correctlyrecognizethat this rule applies toinvest- ments of time as well(Larrick, Morgan, & Nisbett,1990).Similarly, amanagerin industry may find it easier to applythe specific military adage about their being no such thingas bad troops” than to apply a more generallessonabout thefun-

damental attribution error.

CorrectiveversusPreventative

Cognitive repairs also differ in whethertheypreventor correcttheshortcomings of individuals. Corrective repairs interveneduring or after aparticularcognitive process (e.g., the accounting team that corrects their tendency tooverweightvivid exceptions by forcing themselves to consciously classify each exceptionas major or minor). At the extreme, a corrective repair might only intervene at thevery end of a processto correctthe overall outcome(e.g., safety factors). Preventative repairs intervene early in a cognitive process before shortcomings have had a chance toact. Microsoft prevents developers from acquiringa biased sample about the speed of their programsby forcing them to develop programson the same machines usedby customers.

Someshortcomingsareeasierto correctthan others.Forexample, when ashort- coming arises because individuals have the wrongrule, they maynot find it diffi- cult to substitute a different rule(Wilson & Brekke, 1994).Trauma physicians may learnto check airway before breathing, and accountantsmay learn to ignore vivid but minorexceptions. In general, corrective repairs will be appropriate when individuals accept the need for a repair andthey understandhow to execute the correction.

However,when a shortcoming arises becauseof some basic cognitive process, organizationsmay need to intervene more forcefullyby bypassingor eliminating the faulty process (Arkes,1991;Wilson & Brekke, 1994).Forexample,individu- als may find it difficult to generate abroadand independentset of hypotheses because associative memory leads them toconsiderthe samealternativestheypre- viously considered.Theoretically, Motorola could instruct individuals whoare developing anew consumerproduct to “ignore what you’ve done in the past and approach this problemcreatively.”However, individuals mightfind it difficult to ignore their previoussolutions.Thus, Motorola prevents the problemby prohibit- ing them from serving on morethan one productdevelopmentteam. Similarly, the Chicago Board of Trade could warn itsinvestigatorsto discountthe answersthey receive whenthey ask leading questions. Instead it prevents individualinvestiga- tors from asking yesor no questions, and thusensuresthat they receivelessbiased information in thefirst place.

Familiar versusNovel

Repairsmay also vary in the extent to which theyarenovel ratherthan familiar. Novel repairs require individuals to change their assumptions or tolearna proce- dure from scratch. For example the “programmers” at the CRSSarchitecturalfirm had to learn to resist theirtendencyto sketch solutions before evaluatingall the informationfrom a client. On the otherhand,familiar repairs buildon preexisting knowledge (e.g., the CAMELschemafor bank examiners or the ABCs fortrauma physicians).Theserepairs have familiarcontent-traumaphysicians knowthat they should attend tobreathingand circulation, and bank examiners knowthey should pay attentionto capital and earnings. They also have afamiliar form—they are organizedby a simple acronymthat individuals already haveas a part of their mental lexicon.

Familiarrepairsmay be at anadvantageovernovel repairsbecausethey areless costlyto use and their benefitsmay be moreapparent.CAMEL and the ABCs reduce costsby using a familiaracronymto remind individuals tocollecttypes of information that they know they should be collecting. In contrast,the CRSSpro- grammershad to work hard to overcome thebehaviorsthey had learnedas archi- tects, andthey mayhave questioned the benefits of theelaboratenew procedures they were being taught.Familiarrepairsare alsoless likely to provoke resistance

28 CHIPHEATH, RICHARD P. LARRICK, and JOSHUAKLAYMAN Cognitive Repairs 29

thannovel repairs. Anything that requires individualstothrow out old practices or adopt new beliefsmay be technically and psychologically difficult.

However, familiar repairsmay sometimes be too familiar for theirown good. First, they may be lesslikely to createenthusiasm.If individuals thinkthat a new repairdiffers only trivially from current practice,they mayseeno advantage toit. Because individuals often ignore thefamiliar, would-be change agents often strive to createthe perception that their programs arenovel and unique(Abrahamson, 1996).Second, familiar repairsmay be subject todistortion.If a repair seemspar- tially familiar, individuals may neglect itsunfamiliar aspects or force them to mimic the more familiar aspects (aprocessthat psychologists call assimilation). For example, the propertechniquefor brainstorming requires aspecificsequence of steps:first, a creative, idea-generation stage which does notallow criticism, then a stage where ideasare evaluatedand selected. Although organizationsfre- quently report thatthey use brainstorming, careful examinationrevealsthat the organizationsare actually engaged in a more familiar activity: a basic business meeting (Zbaracki, in press). Thenovel aspectsof the brainstorming procedure, suchas the separation of stages and the “no criticism”rule, are often lostas brain- storming isassimilated to the more familiar practice of the standard business meeting.In the end, only theattractivelabel remains. In situations whereassimt- lation is a problem, repairsthat are novel may be less likely to suffer distortion than repairsthat are more familiar becausenovel repairsdo not evokethe preex- isting knowledgethat leadsto assimilation.

SocialversusIndividual

Many of the cognitiverepairs we have considered are social;they work because individuals interact with others(e.g., single-caseborequestions in the Deming Prize organization,or the murderhoard atthe Pentagon).Otherrepairsare indi- vidual; individualsapply them to theirown learning processes without theinter- vention of others(e.g., individuals learn to use the Five Whys, and individual investigatorsat the Board ofTrade learn to avoid “yes or no” questions).

In general, we suspect that manysuccessfulrepairs will be social becauseindi- vidualsmay not recognizethe need to repair themselves. The very cognitiveshort- comings that organizations might hope to repairwill make it difficult for individuals to see theirown flaws. As we havediscussed,individuals tendtoretain current assumptionsin the face of conflicting data (Klayman,1995). Also, they interpretevents in waysthat protect their self-image; theyavoid potentiallythreat- ening feedback(Frey, 1986; Larrick, 1993)and attribute their poor outcomes to luck or forces outside theircontrol. Although individualsmay find it hard torec- ognizetheir own biases,they may findit easier torecognizethe biases ofothers. Many of the repairs we document have the feelof (friendly) socialgamesmanship. Forexample, learners at FP&L did notconsiderthe hypothesisthatthey had been “rucky”—their colleagues considered it for them. Similarly,during weekly micro-

biology lab meetings,researchersdid not have to suggest alternative ways ofinter- preting theirevidence,their peers didso. Social competitionamong individuals aids the spreadof repairs even when individualsare overconfident and believe they would have donejust aswell without therepair.

Social repairsdo have to overcomesomedisadvantages.For example,individ- uals may not appreciate otherswho attempt to repair their biases, andthey may dismissthe repairattemptsas theproductof picky or abrasive personalities. Thus, social repairs may be moresuccessfulwhen an individual understandsthat his or hertormentorsare playing an important, formal role. Individualsmay find it easier to entertainan antagonist’scritiques when he or she islabeledas a“devil’s advo- cate,” or whenthe individual is appearing before the “murder board.” Disneyclar- ified the role of the evaluators and theoccasionby establishingits norm of frank feedback andby labelingits tryouts as “The Gong Show.”

Eventually,social repairsmay be transformed into individual repairs asindivid- ualslearnto imitate the patterns of analysis forcedon them by others. In order for individuals to learn, they needvivid, immediate feedback.Social encountersare likely to provide a keysource of such feedback.For example, when Deming examiners asksingle-case-borequestions, or when lab colleaguestry to shoot holes in a lab presentation, individual learnersmay eventually learn to engage in preemptiveself-criticismin order tolookbetterduring social encounters(Tetlock, 1992).(Many academicpapers arebetterbecause authorslearnto mentallysimu- late potentialreviewer’scomments.)Such repairsinvoke social forces at twodif- ferentlevels: individuals who anticipatesocial interactionmay be more aware of someof their own shortcomings, andthenactualsocial interactionmay overcome additionalshortcomingsthat individualscannotrepair on their own.

Top-DownversusBottom-Up

Some cognitive repairs originate from“top-down” within an organization.Typ- ically these repairsare deliberately designed and implemented by managers or outsideexperts.Others arise frombottom-upthrough informal observationorser- endipitousdiscovery fromthe people who aredoingthe work.

The source of the repair isimportantbecause it is likely to affect itsform. In general, bottomup repairs, suchas organizational adages, willbe simpler and more domain-specific than top-down repairsdesignedby technically sophisti- catedengineers,statisticians, or management gurus.Thelocal originof bottom-up repairs may also make them feel more familiar and acceptable than top-down repairs. Thus, theorigin of a repair will be highlycorrelatedwith many of the tradeoffs we have already discussed.

More importantly, the origin of the repair is also likely to affecthow potential adopters perceiveit. Top-down repairsmay be perceivedwith suspicion orhostil- ity precisely becausethey originateoutsidethe organizationor are imposedfrom above. Front-line workers may doubt that outsiderslike consultantsunderstand

30 CHIP HEATH, RICHARD P. LARRICK, andJOSHUAKLAYMAN Cognitive Repairs 31

their situationwell enough to make wise recommendations. When managerssug- gest a repair,they seem toimply that employees havebeenperformingpoorly or cannot be trusted to perform theirjob correctly. If so, then individualsmay resist adopting a repairbecauseof the same kind of self-serving biases we discussedear- lier.

Top-downrepairsmay also be resisted becausethey will be perceivedasdriven by politics or fashion and notby the demands of the task. Some top-downrepairs may be resisted becausethey seem too political. Particularly when top-down repairs rely on fixed procedures,they may provoke resistance because individuals may think thatthey aredesigned tocentralizecontrol or removeindividual-discre- tion. Other top-downrepairsmay be resisted becausethey seem tobe driven by fashion. Institutional theorists contend that organizations adopt new practices for reasonsother than greaterefficacy. Organizational membersmay share similar, cynical intuitions (Adams,1996),and will resist repairsthat they seeasmerewin- dow dressing oras this year’s fad. Whenindividuals have trouble recognizing their shortcomings, thenthey may be particularly likely to attribute top-down repairs to politics or fashion becausethey will not recognizethe repair’s true value.

Bottom-up repairs will often benefit from their local, homegrown origins. Local repairs have a meaningful history that makesthem memorableand appealing. Even a repairthat is potentially threatening, suchas “you were rucky,” may be more acceptable if organizational members see itas their own invention. Just as lawyers are entitled to tell lawyerjokes,organizational members are entitled to develop self-critical repairs and toconveytheir insider statusby using iliem. And homegrown repairs evoke a stronger sense of ownership; at the same time that they call attentionto a potential shortcoming,they also givethe n’~er creditfor fix- ing it.

Conclusions: SuccessfulRepairs

We have considered six differentdimensionsalong which cognitive repairscan be classified. For example, the physician’s ABCsare simple, domain-specific, corrective,familiar, individual, and top-down. Although we have suggestedsome advantages and disadvantagesofeach endpoint of each dimension, we believe-that our discussion suggests at leastsomepreliminary conclusions about successful repairs. Forexample,because individuals have limited faculties,organizations who wish individuals to learncomplex,domain-generalrepairs will find them- selvesdevoting a greatdeal of scarce time, money, and effortto ensurethat such repairs are learned and used. Similarly, because individuals are overconfident about theirown conclusions, they may not spontaneously execute individual repairs tocorrecttheir own biases. Basedon the advantages and disadvantages we have considered, we suspect that the most successful repairs will be simple, domain-specific,socially administered, and evolved frombottom-uprather than

developed from top-down.We find this conclusion intriguingbecauseit describes repairs that differ sharply from those that arerecommendedin academiclitera- tureson decision analysis, statistics, and economics.

Implications for Research

Cognitive psychologistsoften think of people asruggedcognitive individual- ists, constrainedby their own cognitive limitations and a poor environment in which to learn. Cognitiveresearcherscontinueto argueoverhow well individuals actually perform the cognitive tasksthey encounterin their lives (see Anderson, 1991, and accompanyingcommentaries,and the debate between Gigerenzer, 1996; and Kahneman &Tversky, 1996). However, it is importantto remember that much of what people do, including muchof their cognition, takes place in the contextof organizationsand other social structures.

Some recent approaches in psychology doexplorecultural andsocial contribu- tions to individual learning. For example, work on “transactivememory” (Weg- ner, Erber, & Raymond, 1991; Liang, Moreland, & Argote, 1995) showshow individuals reduce their memory limitationsby distributing memory-heavytasks across multiple people. Thus,therecan becollectivememory that does notreside in any individual. Our concept of organizational repairs is in the samespirit, but it deals with“higher order” cognition: Reasoning and decision making can alsobe improved through socialstructureand culturalbootstrapping.

We also believethat organizationalpsychologists couldbetterunderstandorga- nizational processes ifthey thoughtmore about the cognitiveprocessesof the indi- viduals who makeup the organization.Researchtying individual psychology to organizationalbehavior certainlyhas along and venerable pedigree (March & Simon, 1958),but recently,some researchershave expressed concern that that approachis still underutilized. For example, in the contextof institutional diffu- sion processes,Zucker(1991) has warnedthat “without a solid cognitive, micro- level foundation, we risk treatinginstitutionalizationas ablackbox at theorgani- zational level, focusing oncontentat the exclusion of developing a systematic explanatorytheory of process” (p.105). And Miner (1994) warns,“most evolu- tionary discussions of organizational changediscussroutines asthoughthey exist independentof individual human beings” and“evoke images ofdisembodiedenti- ties removed from day-to-day human interaction” (p. 87). Knowledge ofindivid- ual cognitioncan be crucial tounderstandingwhy thingshappenastheydoiri-an organization.Forexample,why do engineeringfirms use a systemby which the best engineers make their bestestimates,only to have the firm second-guess them by adding a huge safety factor?Any explanationwould beincompletewithout an understandingof individual overconfidence.Understandingthe abilities andcon- straints of individualspermits a kind of cognitive archaeologyof organizational practices thatmay allow organizationalresearchersto better understandwhy cer- tain rules, norms, andproceduresdevelop in organizations, andwhy others fail.

32 CHIP HEATH, RICHARD P.LARRICK, andJOSHUAKLAYMAN Cognitive Repairs 33

Implicationsfor Practice

Managers who think explicitly about cognitive repairswill, we think, be in a betterposition tofoster improvements in their organizations. Managers already think about factors such as incentive systems andinformation-technoLQgyas tools to fosterlearning andinnovation.We believe that cognitive repairswill be a useful additionto the toolbox. Managerswho consciouslyconsiderindividual cognition may be able torecognizea larger number of repairopportunitiesandmay target top-down repairs moreeffectively. Furthermore,they may designmore effective repairs ifthey take a cognitive approach andconsiderrepair dimensionslike the six we discussedearlier.

Even whenrepairsare developedbottom-uprather than top-down, a manager who is informedabout individualcognitionmight have a positive influence.As has beenobservedby researchers who think about evolutionary approaches to organizations, oneof the critical components of organizational learning is tostart with a rich and varied pool of alternative practices(Levitt & March, 1988;Miner, 1994). Savvymanagers canenhancethe pool of alternatives that areavailableby teachingorganizationmembers about the concept of repairs andby encouraging them toidentify existing repairs and to seek new repair opportunities. Managers canrecognizeand reward individuals whodiscovercognitiverepairs,and-they-can disseminate effectiverepairs via demonstration, training,communication,and rotation of personnel.

Managersmay also find it veryimportantto think about cognitive repairs when they evaluate existing organizationalpractices. Consider, for example, a new executivewho discovers that her development groupis split up into six separate teams, eachtrying to solve the same problem without anycommunicationwith each other untilwell into the process.Thismight seemlike a paradigmof bureau- cratic inefficiency. Yet, this is the kind of systemthat Motorola has found to be effective ingeneratinga broaderset of options. Without understanding thepoten- tial value of this repair, the newexecutivemight besadly surprisedby the results of streamlining theprocess.

Managers might also benefit from a cognitive approachbecausecognitive repairs, like other innovations,may suffer from the law of unintendedconse- quences. A repairthat is intended tofix one problem can well end up creating another.Recall, for example,Microsoft’s laudableattempt to make thecustomer moresalientto program developersby having them watch live-customers attempt to use their products in the usabilitytestlab. Thetestlab repairedthe developer’s tendency to beunmovedby cold statistics, but it probablyexacerbatedtheir ten- dency to believethat small samples were reliably representative.In responseto seeing onecustomerin the test lab,developersmight wastetimealtering a feature that would have been okay for mostcustomers.Managers who take a cognitive approach would,we hope, be more likely to avoid unwantedside effects, or at least bein a betterposition to recognize and cope with them.

Final Words

In contrast to Hamlet’s enthusiasm, we argue thatthereis good reason to be aware ofthe limitations of individual learners. People are not “infinite infacul- ties” or “noble in reason.” As individuals, we make systematicmistakesthat com- promiseour ability tolearnfrom our experiences and tounderstandthe world.On the otherhand, wemortals are notall fools. We are able toform socialstructures that havethe potential to magnifysomeof our abilities and to minimizesomeof our failings. In thispaperwe have concentrated ondemonstratingthat effective organizational repairs happen. Wedo not mean to imply thatorganizationscure any and all individual shortcomings, nor eventhat organizationsalways make thingsbetterrather thanworse.Nevertheless, we do believe that theorganizations in which we work can provide us with norms, reminders, protocols, andproce- duresthat help usmovebeyond our individual shortcomings.

ACKNOWLEDGMENTS

We thankthe Graduate Schoolof Business,University of Chicagoandthe FuquaSchoolof

Business,Duke University, for researchsupport. For helpful commentson this project, we

thank participantsat Rod Kramer’s 1996 conferenceon OrganizationalLearning; we also thank Jim March, Cade Massey, Elizabeth McSweeny, Sim Sitkin, Dick Thaler, Elke

Weber, andJohnWright.

NOTES

1 We primarily want to distinguish problemsof incentives fromproblemsof mental process. In this reviewwe will not distinguish between mental errors that arise from “motivatedreasoning”and thosethat arise from “colder” processes.

2. Personalcommunication,Abhay Joshi. 3. Personalcommunication,RobertAlan Fisch. 4. Personalcommunication,Jan Elsbach. 5. Personalcommunication,Francisco Bayron. 6. Personalcommunication,DeanSiewert. 7. Personalcommunication,JustinScott Bradley. 8. Personalcommunication,Francisco Bayron. 9. Personalcommunication,DeanSiewert.

10. Personalcommunication,Ken Myszkowski.

II. Personalcommunication,Leo Tucker. 12. Personalcommunication,Raymond Stukel. 13. Personalcommunication,John Payne

REFERENCES

Abrahamson, E. (1996). Management fashion.Academy ofManagementReview, 21, 254-285. Adams,S. (1996). The Dilbert principle. New York: Harper Business.

34 CHIP HEATH, RICHARD P. LARRICK, andJOSHUAKLAYMAN Cognitive Repairs 35

Agostinelli,0, Sherman,S. J., Fazio, R. H.,& Hearst, E.5. (1986). Detecting and identifying change: Addition versus deletion.Journal ofExperimental Psychology. Human Perception and Perfor- mance, 12, 445-454.

Anderson,J.R. (1991). Is human cognitionadaptive? Behavioral and Brain Sciences,14,471-517. Anderson,JR. (1995).Cognitivepsychologyand its implications.New York: W.H. Freeman& Co. Arkes, H. R. (199 I). Costs and benefitsofjudgmenterrors: Implications for debiasing.Psychological

Bulletin, 110, 486-498. Baron,J. (1988). Thinking and deciding. Cambridge, UK: Cambridge University Press. Bassok,M. (1990). Transferof domain-specificproblem-solvingprocedures.Journal ofExperimental

Psychology: Learning, Memory, and Cognition, 16, 522-533. Bell, D. E., Raiffa, H.,& Tversky, A. (1988).Decision making: Descriptive, normative, and prescrip-

tive interactions.Cambridge: Cambridge University Press. Borgida, E.,& Nisbett, R. E. (1977). The differential impactof abstractvs. concrete information on

decisions.JournalofAppliedSocialPsychology7, 258-271. Buehler,R., Griffin, 0., & Ross,M. (1994). Exploring the “planning fallacy”: Why peopleunderesti-

mate their task completiontimes. Journalof Personality and SocialPsychology67, 366-381. Cheng,P. VV., & Holyoak, K. J.(1985). Pragmatic reasoningschemas.CognitivePsychology,17, 391-

416. Cohen, L.1. (1981). Can human irrationality beexperimentally demonstrated?Behavioral and Brain

Sciences, 4, 317-331. Cohen,M. D. (1991). Individual learning and oganizational routine: Emerging connections.Organi-

zation Science, 2, 135-139. Cohen,E. A., & Gooch,J. (1990).Military misfortunes:The anatomyoffailurein war. New York: Free

Press. Cusumano,M. A., & Selby, R. W. (1995).Microsoft secrets.New York: Free Press. Davis, 0. (1985,March-April). New projects: Bewareof false economies.Har’.’ard BusinessReview,

95-11)1. Deming, W. E. (1982).Out of the crisis. Boston: Massachusetts Instituteof Technology Center for

Advanced Engineering Study.

DiMaggio, P. J., & Powell Xv. W. (1983). The iron cage revisited: Institutionalisomorphismandcol- lective rationality in organizational fields.American SociologicalReview,48, 1,47-160.

Dunbar,K. (1995). How scientists reallyreason:Scientific reasoning in real-world laboratories. In R. J. Steinberg& J. Davidson (Eds.),The natureof insight.Cambridge,MA: MIT Press.

Eisenhardt,K. M. (1989). Agency theory: An assessment andreview. Academyof Management Review, 14(1), 57-74.

Eppel, D.,Argote, L., & Devadas, R. (1991). Organizational learningcurves:A methodfor investigat- ing intra-plant transferof knowledgeacquiredthroughReamingby doing. OrganizationalSci- ence, 2, 58-70.

Fischhoff,B., Slavic,P.. & Lichtenstein, 5.(1978). Fault trees: Sensitivityof estimatedfailure proba- bilities to problem representation.Journal of Experimental Psychology: HumanPerception andPerformance,4, 330-344.

Fiske, S. T., & Taylor,S. E. (1991).Social cognition(2nd ed.). NewYork: McGraw-Hill. Fletcher,C. (1990).What cops know.New York: Pocket Books. Fong, G. T., & Nisbett, R. E. (1990). Immediate anddelayedtransferof training effectsin statistical

reasoning.Journalof Experimental Psychology:General, 120, 34-45. Forum Corporation(1992).P.R.O.G.R.E.S.S.tool kit. Forum Corporation. Frey, D. (1986). Recent research on selective exposure to information.Advancesin Experitnental

SocialPsychology19, 41-80. Funder,D.C. (1987). Errors andmistakes:Evaluating the accuracyof social judgment.Psychological

Bulletin, 101, 75-90.

Gettys, C.F. et al. (1987). An evaluationofhuman act generation performance.OrganizationalBehav- ior and Human Decision Processes,39, 23-51.

Gigerenzer, G. (1996). On narrownorms and vague heuristics: A replyto Kahneman and Tversky (1996).PsychologicalReview, 103, 592-596.

Gilbert, D. T., & Jones, E. E.(1986). Perceiver-induced constraint: Interpretationsof self-generated reality. Journalof Personality and SocialPsychology50, 269-280.

Gilbert, D. T.,& Malone,P.S.(1995). The correspondence bias.Psychological Bulletin,117, 21-38. Gnepp,J., & Klayman,J. (1992). Recognitionof uncertaintyin emotional inferences: Reasoningabout

emotionallyequivocalsituations.DevelopmentalPsychology28, 145-158. Gregory, Xv. L., Cialdini, R. B., & Carpenter, K.M. (1982). Self-relevant scenarios as mediatorsof

likelihoodestimatesand compliance: Does imagining makeit so?Journalof Personality and Social Psychology43, 89-99.

Griffin, 0., & Tversky,A. (1992). The weighingof evidence andthedeterminantsof confidence.Cog- nitive Psychology24,411-435.

Heath,C., Knez,M., & Camerer, C. (1993). The strategicmanagementofthe entitlement processin the employmentrelationship.StrategicManagementJournal, 14, 75-93.

Hilton, D.J. (1995). The social contextof reasoning: Conversational inference and rational judgment. Psychological Bulletin, 118, 248-271.

Hoch,S. J. (1984).Availability and interference in predictive judgment.JournalofExperimentalPsy- chology:Learning, Memory, andCognition, 10, 649-662.

Hodgins,H. S.,& Zuckerman,M. (1993). Beyond selecting information: Biasesin spontaneousques- tions and resultantconclusions.Journalof Experimental SocialPsychology29, 387-407.

Holland, J. H., Holyoak, K.J., Nisbett, R. E., & Thagard,P. T. (1986).Induction: Processesof infer- ence,learning, and discovery.Cambridge,MA: Bradford Books/MIT Press.

Hynes,M., & Vanmareke, E. (1976).Reliability of embankmentperformancepredictions.Proceedings of the ASCE Engineering Mechanics Division SpecialtyConference.Waterloo, Ontario,Can- ada: University of Waterloo Press.

lmai, M. (1986).Kaizen: The key to Japan’scompetitivesuccess.New York: McGraw-Hill. Ishikawa, K. (1982).Guide to quality control(2nd revised ed.). Hong Kong: Asia Productivity Orga-

nization. Ishikawa, K. (1985).What is totalqualiry control? TheJapanese way.Englewood Cliffs,NJ: Prentice-

Hall. Juran, J.M. (1992).Juranon quality by design:The new stepsfor planningquality into goods andser-

vices.New York: Free Press. Kahneman, D., Slovic,P., & Tversky, A. (1982). Judgmentunderuncertainty: Heuristicsand biases.

Cambridge, England: Cambridge University Press. Kahneman, D.,& Tversky, A. (1996). On the realityof cognitive illusions.PsychologicalReview,103,

582-591. Kaplan, J. (1995).Start-up: A silicon valiry adventure.Boston: Houghton, Muffin Co. Keller, P. A., & McGill, A. L. (1994). Differencesin the relative influenceof product attributes under

altemative processing conditions: Attributeimportanceversus attribute easeof imagability. Journalof Consumer Psychology3, 29-49.

Koehler,I. J. (1993). The influenceof prior beliefson scientificjudgmentsofevidencequality. Orga- nizational Behavior and Human Decision Processes,56, 28-55.

Klayman, J. (1995). Varietiesof confirmation bias.The Psychologyof Learning and Motivation,32, 385-418.

Kunda, Z. (1991)).The case formotivatedreasoning.Psychological Bulleti,t,108, 480-498. Larrick, R. P. (1993). Motivational factorsin decision theories: The roleof self-protection.Psycholog-

ical Bulletin, 113, 440-450. Larrick, R. P., Morgan, J. N., & Nisbett, R. E. (1990). Teaching the useof cost-benefitreasoningin

everyday life.PyschologicalScience,1, 362-370.

36 CHIPHEATH, RICHARD P. LARRICK, andJOSHUA KLAYMAN Cognitive Repairs 37

Levitt, B., & March, J. G. (1988). Organizational leaming.AnnualReviewofSociology14, 319-40. Liang, D. Xv., Moreland,R., & Argote, L. (1995). Group versusindividual training and groupperfor-

mance:The mediatingfactorof transactivememory.Personality& SocialPsychologyBulletin, 21, 384-393.

Lichtenstein, S., Fischhoff,B., & Phillips, L. D. (1982). Calibrationofprobabilities:The stateofthe art to 1980. In A. Tversky, D. Kahneman, &P. Slovic (Eds.),Judgment underuncertainty:Heu- ristics and biases(pp. 306-334).New York: Cambridge.

Lord, C., Ross, L., & Lepper,M. R. (1979). Biasedassimilationand attitude polarization: The effects ofprior theorieson subsequentlyconsidered evidence.Journalof PersonaliryandSocialPsy- chology,37, 2098-2109.

Mahoney,M. J. (1976).Scientist as subject:Thepsychologicalimperative.Cambridge,MA: Ballinger. Making quality more than afad. (1992, May 18). Fortune,pp. 12-13. March,J. G., Sproull, L. S., &Tamuz,M. (1991).Learningfrom samplesofone orfewer. Organization

Science,2, 1-13. March, J. G.,& Simon, H. A. (1958).Organizations.New York: John Wiley. McGowan, J. (1996, April1). How Disney keeps ideascoming. Fortune,pp. 131-133. Milgrom, P., & Roberts,1. (1992).Economics,organization, and management.EnglewoodCliffs, NJ:

Prentice Hall. Miner, A. 5. (1994). Seekingadaptiveadvantage: Evolutionary theory and managerial action. In J. A.

C. Baum & J. V. Siogh (Eds.),Evolutionarydynamicsof organizations (pp. 76-89).New York: Oxford University Press.

Miner, A. S. & Mezias, S. J. (1996).Ugly duckling no more: Pasts and futuresof organizationalleam- ing research.OrganizationScience,7, 88-99.

Mullen, B., & Riordan, C. A. (1988). Self-serving attributions for performancein naturalisticsettings~ A meta-analyticreview. JournalofAppliedSocialPsychology18, 3-22.

Mynatt, C. R,Doherty, M. E., & Dragan,Xv. (1993). Information relevance, working memory, and the considerationof altematives.QuaterlyJournalof ExperimentalPsychology46A,759-778.

Newman, J.,Wolff, Xv. T., & Hearst, E. (1980). The featurepositive effect in adult human subjects. Journal ofExperimental Psychology: Human Learning andMemory, 6, 630-650.

Nisbett, R., & Ross, L.(1980). Human inference: Strategiesand shortcomingsof social judgment. Englewood Cliffs,NJ: Prentice-Hall.

Nisbett, R. E. (1992). Reasoning, abstraction, and theprejudicesof 20th centurypsychology. In R. E. Nisbett (Ed),Rulesfor reasoning(pp. 1-12).Hillsdale, NJ: Erlbaum.

Nisbett, R. E.,.Krantz,D. H.,Jepson,D., & Kunda, Z. (1983). The useof statisticalheuristicsin every- day reasoning.PsychologicalReview,90, 3 39-363.

Nonaka, I.,& Takeuchi, H. (1995).The knowledge-creating company: How Japanese companiescre- ate the dynamicsof innovation.New York: Oxford University Press.

Odean,T. (1996). Volume,volatility, price and profitwhen all traders areabove average.Working paper.University of California, Berkeley.

Payne, J.Xv. (1976). Task complexity andcontingentprocessingin decisionmaking: An information search and protocol analysis.Organizational Behavior and HumanPerformance,16, 366-387.

Perrow,C. (1984). Normal accidents:Living with high-risktechnologies.New York: Basic. Peters,T. (1992). Liberation management:Necessarydisorganizationfrr the nanosecond nineties.

New York: Alfred A. Knopf. Rogers, E.(1995). Thediffusion of innovations(4th ed.). NewYork: Free Press. Rosenthal,E. (1994,May). Invisible injuries. Discover,pp. 48-51. Ross, L., & Nisbett,R. E. (1991).Thepersonand the situation: Perspectivesofsocial psychology.New

York: McGraw-Hill. Ross, L.(1977). The intuitive psychologist and his shortcomings. In L. Berkowitz (Ed.),Advancesin

experimentalsocialpsychology(Vol. 10). New York: Academic Press.

Russo,J. E.,& Schoemaker,P.J. H. (1992). Managing overconfidence.SloanManagementReview,33, 7-17.

Sapolsky,R. (1997, November). Ontheroleof upholsteryin cardiovascularphysiology.Discover,pp. 58-66.

Scott,Xv. R. (1995). Institutions andorganizations.ThousandOaks, CA: Sage. Senge,P. M. (1990). Thefifth discipline. New York: Doubleday Currency. Slovic,P., Fischhoff,B., & Lichtenstein,5. (1982). Facts versus fears: Understandingperceivedrisk.

In A. Tversky, D. Kahneman,& P. Slovic (Eds.),Judgmentunderuncertainty: Heuristicsand biases(pp. 463-492).New York: Cambridge.

Smith, J. F., & Kida, T. (1991). Heuristics and biases: Expertise and taskrealismin auditing. Psycho- logical Bulletin, 109,472-489.

Tetlock, P. E. (1992). Theimpactof accountabilityon judgmentandchoice: Toward a socialcontin- gency model.Advancesin Experimental SocialPsychology25, 331-376.

Tversky, A., & Kahneman, D. (1971). Belief in the lawof small numbers.PsychologicalBulletin, 76, 105-110.

Tversky, A., & Kahneman, D. (1973).Availability: A heuristicforjudging frequency and probability. CognitivePsychology5, 207-232.

Walton, M. (1990).Demingmanagementat work. New York: Putnam. Wegner, D. M., Erber,R., & Raymond,P. (1991). Transactive memoryin close relationships.Journal

of Personality and SocialPsychology61, 923-929. What Bill Gates really wants. (1995,January16). Fortune,p. 34. Wilson, T. D., & Brekie, N. (1994). Mental contaminationand mentalcorrection:Unwantedinflu-

enceson judgmentsand evaluations.Psychological Bulletin,116, 117-142. Zbaracki,M. (in press). The rhetoric and realityof total quality management.AdministrativeScience

Quarterly. Zucker, L. G. (1991). Postscript:Microfoundationsof institutional thought. InXv. Xv. Powell & P. J.

DiMaggio(Eds.),The newinstitutionalismin organizationalanalysis(pp. 103-107).Chicago: Universityof Chicago.

Zuckerman,M., Knee, C.R., Hodgins, H. S.,& Miyake, K. (1995). Hypothesis confirmation: Thejoint effectof positivetest strategy and acquiescenceresponseset.JournalofPersonalityand Social Psychology68, 52-60.