psychonomic Bulletin & Review 1997, 4 (1), 63-67

Implicit learning: Indirect, not unconscious

BRUCE W. A. WHITTLESEAand MICHAEL D. DORKEN Simon Fraser University, Burnaby, British Columbia, Canada

Dienes and Berry (1997)argue that there really are two kinds of knowledge, implicit and explicit, that can be characterized through a subjective threshold. Weargue that the subjective threshold does not separate two forms of knowledge, but instead two uses of knowledge, for two different purposes. With Neal and Hesketh (1997),we suggest that an understanding of learning will be better served by inves­ tigating the processing that occurs under various circumstances than by attempting to isolate learning that occurs with awareness from that which occurs without it.

Psychologists love dichotomies. Faced with the com­ authors to conclude that the subjects acquired some knowl­ plexity of behavior, we have traditionally responded by edge implicitly during the training and used that knowl­ subdividing performance into modules governed by dif­ edge implicitly in test. We will question the implicitness ferent principles: procedural versus declarative, seman­ ofthe use ofknowledge in the next section; first, we will tic versus episodic, short- versus long-term remembering, discuss the implicitness ofthe original acquisition. and so on. In attempting to understand how organisms There is an important ambiguity in the meaning ofthe adapt to their environment by changing their behavioral words "learning" and "knowledge." One sense is the ac­ tendencies, we have imposed multiple dichotomies, in­ tual, direct computation ofinformation about a stimulus, cluding controlled/automatic, incidental/intentional, and such as the computation that ROADis a synonym ofSTREET. implicit/explicit. Like the descriptions ofan elephant by In that case, the person actually acquires a knowledge unit, the seven blind men, each of these has merit, but none a representation ofthe experience ofcomputing the rela­ quite captures the subject matter. tionship. The person can know that he has that knowledge In their comprehensive review,Dienes and Berry (1997) by recalling it. However, along with that representation, defend the utility ofthe implicit/explicit distinction in the the person also acquires the potential to perform differ­ implicit learning literature. They argue that many ofthe ently on future occasions. For example, he may experi­ studies they review demonstrate learning occurring above ence savings on a later occasion in which he is again asked an objective threshold but below a subjective threshold, about the relationship ofROAD and STREET. This potential and that this learning has properties different from those to perform is not directly computed and represented in of learning that occurs with awareness. In contrast, we memory, nor is it computable in detail: Any stimulus ex­ argue that every act oflearning carries with it a change in perience provides resources for an infinite number of the potential to perform many other activities. People are possible future occurrences. The person cannot know the unaware ofthese changed potentials, not because the learn­ potential abilities that the new representation entails until ing occurs under a threshold, but because they are only that potential is realized in some future stimulus encounter. incidentally related to what the person actually learns. We On the first occasion that a person compares ROADand argue that implicit learning is just ordinary learning with­ STREET, that person already possesses information about out becoming aware of the implications of that learning each that makes it possible to perform the judgment. But for performing unanticipated activities. that information was .not directly about the relationship between ROAD and STREET, and until it occurs to the per­ The Original Learning: son to perform the comparison, she cannot know that she Implicit Learning or Acquiring a Potential? has the knowledge to do it. In learning separately about In a standard "implicit learning" experiment, subjects the concepts, the person acquired the potential to per­ study a set ofexemplars ofa grammar and are then given form the comparison, but it remains only a potential until a surprise classification test, in which they show sensi­ the person actually computes that relationship. Is this im­ tivity to the underlying rules, without apparent awareness plicit knowledge? Did the person implicitly learn the re­ ofthe basis oftheir success. This observation has led many lationship between ROAD and STREET, and also implicitly learn about all of the other future uses to which that in­ formation could be put, in learning earlier about the sep­ This research was supported by a grant from the Natural Sciences arate concepts? Or did she simply learn about those con­ and Engineering Research Council of Canada. Correspondence con­ cerning this article should be directed to B. W. A. Whittlesea, Depart­ cepts in terms ofwhatever purpose she had at the time, ment of Psychology, Simon Fraser University, Burnaby, BC, Canada and in so doing acquire the potential to perform other V5A 1S6 (e-mail: [email protected]). tasks? Or take a procedural example: a person learning to

63 Copyright 1997 Psychonomic Society, Inc. 64 WHITTLESEA AND DaRKEN play the piano is acquiring finger dexterity skills that to accuracy while performance is above chance. Both are may later facilitate learning to type or to be a mechanic. common-sense ideas, corresponding to the intuitive no­ This person will probably not realize the implication of tion of what it means to be conscious ofone's ability or the early training at that time, and later may not connect knowledge. However, these ideas turn out to be more com­ that practice with ease in learning the new vocation. plex than they initially appear. Clearly, there is implicit transfer of skill, the potential for We believe that the idea of"being aware or not aware which was created in the early training: but piano practice ofthe knowledge" miscasts the way in which people come is not implicit learning to be a mechanic. to a belief about what they know. We argue that what Similarly, subjects memorizing instances ofan artifi­ people say about their knowledge and the confidence cial grammar compute some properties of those items, that they express in their performance are the product of encoding bigrams, substrings, or representations ofwhole an inferential process. People possess representations of items. This learning enables them to classify novel ex­ the events they have experienced, and those representa­ emplars as legal or illegal, because legal items are neces­ tions drive future performance. However, people do not sarily more similar to encoded representations than are have direct, conscious access to those representations. illegal items. That is, they directly encode information Instead, when a stimulus is encountered for some purpose, about items in one task that indirectly gives them the abil­ they can only observe their performance and attempt to ity to deal with related stimuli in an unanticipated task. judge the likely source of their behavior, given their un­ However, the learning ofitems in the memorization task derstanding of the requirements of the task (cf. Nisbett is not direct learning ofthe ability to classify other items: & Ross, 1980). We argue that failure to develop the atti­ It is simply learning ofthe items for the purpose given by tude that one has knowledge ofhow to perform the trans­ that task, which incidentally grants the potential to per­ fer task in an "implicit-learning" experiment is not the form other activities on related stimuli. The ability to clas­ product of learning below a threshold of awareness. In­ sify later items is implicit in the original learning, in the stead, it occurs because the subjects do not interpret their sense ofexisting as a potential, but not in the sense ofa behavior as evidence of such knowledge, under their the­ direct computation ofinformation relevant to that activ­ ory about what is necessary to perform the task. That is, ity. In consequence, we argue that the original acquisi­ we argue, consciousness is an interpretive activity and tion ofknowledge in an "implicit learning" experiment not a passive knowing. is not qualitatively different from learning performed This characterization of learning as acquiring the po­ under any other circumstances. tential to perform unanticipated future tasks and ofmeta­ The original motive for considering the possibility of knowledge as interpreting one's performance goes to the implicit learning as a separate mode oflearning was the heart of the controversy over implicit learning. If the idea that people memorizing instances of a grammar also episodic-processing account of learning (reviewed by learn general, abstract information about the deep struc­ Neal & Hesketh, 1997) is correct, then people learn dur­ ture of the domain without being aware ofthe latter (e.g., ing the training session in both actual and potential senses. Reber & Allen, 1978). That is, it was suggested that they Inprocessing training items to meet the demand to mem­ acquire the ability to perform the transfer task through orize, people compute some organization of each train­ direct computation and representation of information ing item (whether at the bigram, substring, or item level). that is irrelevant to their overt task. Ifpeople did directly They thus know the individual items to some extent, in compute such extra information without intending to, the sense of having representations of encoding those there would be good reason to describe that learning as items. That knowledge grants the potential to discrimi­ a special mode that was different from explicit learning. nate between novel legal and illegal items in the classi­ However, as discussed in the review by Neal and Hes­ fication test. They can also be aware of possessing this keth (1997), there is no positive evidence of such extra knowledge, in the sense that in a recognition test they computation. We conclude that the ability to perform a can discriminate those items from others, and do so with transfer task is simply the realization of a potential that a feeling of familiarity for the original items. J However, was created in the act ofencoding items for the purpose they do not become aware that it is that knowledge that ofmemorization. permits them to succeed in the classification test. Being informed by the experimenter that test items differ in Metaknowledge: Awareness or Inference? conformity to a set ofabstract rules, the subjects assume Wenow turn to the question of people's ability to iden­ that they could only succeed in discriminating the items tify the source oftheir ability to perform the transfer task. if they possessed abstract knowledge. Being unable to Dienes and Berry (1997) argued for a threshold account, recall any such knowledge (for the good reason that it in which knowledge that was inaccessible to conscious­ was never computed), the subjects lack confidence in ness could still control behavior. They offered two criteria their classification judgments and claim to know nothing for establishing that knowledge was beneath a subjective that could assist them in the test. In effect, the subjects threshold: the guessing criterion, that subjects believe lack meta-metaknowledge: Operating under the wrong they are guessing while performance is above chance, and theory about what knowledge is necessary in order to the zero-correlation criterion, that confidence is unrelated perform the surprise classification, they are unable to in- IMPLICIT LEARNING: INDIRECT, NOT UNCONSCIOUS 65

terpret their ability to recall individual training items as between zero (no confidence) and unity (full confidence), metaknowledge about their ability to classify the stimuli. and was also greater for typical than for similar stimuli. This understanding of metaknowledge can be illus­ That is, (1) the subjects had some knowledge enabling trated using Chan's (1992) data, which Dienes and Berry them to classify test stimuli above chance, (2) the subjects (1997) quote as a demonstration ofperformance beneath claimed some confidence in having such knowledge, and a subjective threshold. Chan's memorization and rule­ (3) the patterns ofconfidence and classification were cor­ searching subjects both apparently used knowledge of related. Thus, by both the guessing and zero-correlation bigrams to classify test items as legal or illegal, but the criteria, the subjects had metaknowledge about their clas­ memorizationsubjects' confidence was less related to their sification performance. When pronunciation came first, accuracy than was that ofrule-searching subjects (zero­ the pattern was reversed. Subjects were now more accu­ correlation criterion). Dienes and Berry concluded that rate on similar than on typical stimuli; confidence was memorization subjects lacked metaknowledge that rule­ around.7, but was now greater for similar than for typi­ searching subjectspossessed. Agreed-but was the knowl­ cal stimuli. We interpreted this reversal to mean that sub­ edge they used to classify items inaccessible to the mem­ jects had shifted the basis ofboth judgments from typi­ orization subjects (below a threshold) or was it instead cality to similarity. However, in terms of the criteria of that memorization subjects failed to interpret their item implicitness, the subjects again appeared to have meta­ knowledge as being relevant to the legality decision? knowledge. Chan's further finding that confidence was found to be In the third case, when pronouncing the item was in­ related to accuracy when memorization subjects were serted between the other tasks, accuracy was greater for later asked to rate bigrams for grammaticality suggests typical than for similar stimuli, indicating use of typi­ to us that the subjects were able to recall and interpret their cality information. However, the subjects were more knowledge of bigrams as relevant when an appropriate confident on similar than on typical items, indicating the context was provided at test. use of similarity rather than typicality. Thus, by the zero­ A study by Whittlesea, Brooks, and Westcott (1994) correlation logic, the subjects did not know what infor­ offers another insight into the relationship between clas­ mation they had based their classification decisions on: sification accuracy and confidence. They exposed sub­ in fact, they used a completely different basis to perform jects to exemplarsof two categories,each based on a proto­ the two decisions, even though the second was a decision type. The subjects were trained twice on the same stimuli about the accuracy ofthe first. In a grammar experiment, before being tested. In one training phase, they were re­ this is parallel to basing one's classification on abstracted quired to copy the stimuli, thus encoding each item as a knowledge about the general deep structure of the set, unique entity; in the other, they were asked to judge the but basing one's confidence about the accuracy of that frequency with which each feature of the item had oc­ classification on the ease with which a similar training curred in previous items ofthat category, thereby learn­ item comes to mind. However, the subjects certainly ing about the typicality ofthe features. thought they knew how well they had classified: confi­ There were two types of critical test stimuli, both of dence in their accuracy was again high (about .7 on the which were novel. One type (called typical) shared three zero-to-one scale). features with its prototype and also three with the most Clearly, the subjects did not have explicit access to the similar training item; the other (called similar) shared knowledge they used to perform the classification, or only two features with its prototype but four features with they would have used that same knowledge to evaluate the most similar training item. In making decisions about their confidence: they did not have metaknowledge about test stimuli, the subjects could rely on information about the basis oftheir performance. However,their confidence particular training items, in which case they would per­ ratings were not random: they were based on whichever form better on similar than on typical stimuli or, alter­ form ofknowledge the subjects had used in the immedi­ natively, rely on knowledge about the typicality of the ately preceding task. In that sense, the subjects were highly component features ofthe item, resulting intypical items' sensitive to the specific forms ofknowledge they used to being judged better than similar items. perform other judgments. Did they know what they knew? At test, the subjects were asked to classify each item Wesuggest that this is the wrong question. Just as peo­ and then to rate their confidence that the classification ple's "implicit, unconscious knowledge" is often not was correct. In addition, they were asked to pronounce the knowledge, in the sense of lying in a precomputed rep­ item. We knew from previous studies that this task is resentation, but instead consists of the potential to per­ usually performed using similarity information. The ques­ form toward some unanticipated stimulus-so people's tion was whether performing that task would influence "explicit, conscious knowledge" about their abilities is the basis of performance ofthe other two tasks. Pronun­ often not knowledge, in the sense oftheir having direct ciation was required after the other two, or before both, access to the actual basis of performance. Instead, it is an or between the other two. attribution about what one must know, an inference based When pronunciation came last, the subjects classified on one's judgment about how one is performing now and the stimuli above chance, and were more accurate on typ­ what is required to perform the task (cf. Nisbett & Ross, ical than on similar stimuli, indicating use of typicality 1980). We believe that subjects in this study based their information. Overall confidence was around .5 on a scale confidence judgments on a feeling ofgoodness, a feeling 66 WHITTLE SEA AND DORKEN resulting from the relative fluency with which various change in surface features is limited following the mem­ tasks can be performed on different stimuli (e.g., Jacoby, orization task, not because that task produces implicit Kelley, & Dywan, 1989; Whittlesea, 1993). More typical knowledge, but because it encourages learning about the stimuli and stimuli that are more similar to training items unique surface characteristics of items. Similarly, the are (for different reasons) easier to process than are less broad transfer that accompanies rule eduction occurs not typical or less similar stimuli (Whittlesea, 1987). The because the knowledge is explicit, but because it is ab­ feeling of confidence that these subjects experienced came stract, independent of the features to which it is applied. not from privileged access to the basis on which they made Whittlesea and Dorken (1993) asked some subjects to their classification decisions, but from the ease they ex­ memorize grammatical items; others were shown the perienced in processing the stimuli. same stimuli, each with a marker under one ofits letters, Thus, we argue that the fundamental problem with the and asked to determine whether that letter was repeated search for a way to distinguish implicit versus explicit elsewhere in the item. Neither group was informed about knowledge is that knowledge has been characterized as the existence ofa grammar. At test, the subjects were in­ the possession of some precomputed fact to which the formed ofthe grammar and asked to classify novel items. subject has explicit or only implicit access. Instead, knowl­ Both groups were tested on items presented in the same edge is the potential to perform some activity or judg­ letter set as that of the training items and on the same ment on an object, and metaknowledge is an inference items presented in a novel set ofletters (the order of tests about the source ofthat performance in prior experience. being counterbalanced). The memorization group showed Performance is supported by representations ofprior ex­ above-chance classification on the same-letters test but perience, but the ability to perform some activity toward less ability in the novel-letters condition, demonstrating a stimulus without developing an attitude ofknowing the the narrow transfer described by Dienes and Berry (1997). basis for that performance does not demonstrate the op­ However, the other group performed at identical levels in eration of a special implicit-learning mechanism. the two tests. The analysis of repetition patterns that the second group was asked to perform provided them with Qualitative Differences: Two Modes? abstract knowledge about the stimuli, information that was Dienes and Berry (1997) pointed out that finding of equal applicability when the surface structure was knowledge to be below a threshold of awareness is not of changed. We thus argue that the narrow transfer that is value unless knowledge that is above and below that often observed in implicit-learning experiments is not a threshold is qualitatively different. They argue that there distinguishing feature ofa special implicit-learning mode. really are qualitatively different modes of learning and It is simply a result of the fact that, in order to distract knowledge, distinguished not only by metaknowledge, subjects from the general rules of the set, researchers have but also by breadth oftransfer, probability of success in used induction tasks that cause subjects to focus on sur­ test, and robustness. We will comment only on the first face properties of individual items. two grounds; Neal and Hesketh (1997) have anticipated Success in transfer. Dienes and Berry (1997) re­ (however implicitly) our rebuttal to the third. ported evidence that people who approach complex tasks Transfer to related tasks. Dienes and Berry (1997) in a relatively passive manner, simply memorizing the in­ pointed out that transfer is often smaller when test stim­ stances, perform as well as, if not better than, people who uli are presented in novel surface features than when they attempt to educe the deep structure ofthe task. They con­ are presented in the same surface features as training cluded that implicit learning leads to as much, if not more, items. They concluded that implicit knowledge is to some success than explicit learning in dealing with complex extent perceptually bound and that this limitation on trans­ domains. fer reveals a form oflearning that is qualitatively differ­ Unquestionably, people learn different things about a ent from an explicit form. domain ofstimuli when they are told that the domain has Presumably, the contrast is with learning in the form a deep structure than when they are not. However, the su­ ofhypothesis testing, when the subject is told that train­ periority of memorizing instances over attempted rule ing items possess a deep structure and is asked to dis­ eduction can be explained without resorting to an im­ cover it. We are not aware of studies on the breadth of plicit learning mechanism. First, at any moment during transfer following such training, but we anticipate, with training, a subject is faced with a particular instance but Dienes and Berry (1997), that it would differ from that never encounters the deep structure as a unitary whole, seen after memorization: A subject who learned a useful unless mentally. That makes it easier to learn about indi­ rule would have equal success in transfer on stimuli pre­ vidual members ofthe set than to learn about the abstract sented either in the same or different features, because relations among features across items. To learn the sep­ the rule is applicable regardless ofthe features in which arate stimuli, one has only to study each as it is presented. items are presented. But to learn about the general abstract structure, one must Clearly, memorizing instances and educing rules are hold in mind a large variety ofpotentially important struc­ different activities, resulting in qualitatively different con­ tural units withoutconfusing them. sequences for future performance. However, that is not a Worse, the subjects have no chance of hitting on the sufficient basis on which to argue that they reflect ex­ right hypothesis: The only correct hypothesis is the com­ plicit and implicit modes of learning. Transfer across a plex set ofrules used to generate instances. However, the IMPLICIT LEARNING: INDIRECT, NOT UNCONSCIOUS 67 subjects might try one or two simple hypotheses that are real case for accepting the idea ofa subjective threshold correlated with the actual grammar. Initially, they are ofmetaknowledge, or for qualitative differences in per­ likely to select a few salient parts ofearly items, testing formance that result directly from the implicit acquisi­ the frequency with which those parts occur in succeed­ tion ofknowledge. Instead, we agree with Neal and Hes­ ing stimuli. This puts them at a disadvantage in two keth's (1997) conclusion that abandoning the concept of ways. First, they have not attended (and so not encoded) awareness as the crucial dimension, and instead investi­ the remainder ofeach string, and so are missing a lot of gating the consequences of variation in processing re­ the important structure. Second, if the parts they initially sulting from differences in tasks, contexts, and prior his­ select are oflow frequency across the set, they are likely tory will prove most fruitful in our understanding of to abandon those hypotheses as false (although any part learning. ofa training item represents an aspect ofthe deep struc­ ture), and begin testing some other parts. Because ofthe REFERENCES variability ofstimuli, they are likely to conclude that no CHAN,C. (1992). Implicit cognitive processes: Theoretical issues and hypothesis has much merit. In the end, they have no use­ applications in computer design systems. Unpublished doctoral dis­ ful rules, and also have not encoded much about the sertation, University ofOxford. structure of particular items. DIENES, Z., & BERRY, D. (1997). Implicit learning: Below the subjec­ In contrast, although memorizing the items is not easy, tive threshold. Psychonomic Bulletin & Review, 4, 3-23. JACOBY, L. L., KELLEY, C. M., & DYWAN, 1. (1989). Memory attribu­ it nevertheless causes the subjects to attend to all ofthe tions. In H. L. Roediger III & F.I. M. Craik (Eds.), Varietiesofmem­ components that are relevant in defining the deep struc­ ory and consciousness: Essays in honour ofEndel Tulving (pp. 39 1­ ture. Even poor encoding ofwhole items (e.g., the encod­ 422). Hillsdale, NJ: Erlbaum. ing ofsome bigrams) is thus likely to grant the subjects NEAL, A., & HESKETH, B. (1997). Episodic knowledge and implicit learning. Psychonomic Bulletin & Review, 4, 24-37. some basis for future well-formedness judgments. Thus, NISBETT, R. E., & Ross, L. (1980). Human inference: Strategies and in dealing with complex systems, it may easily be more shortcomings ofhuman judgement. Englewood Cliffs, NJ: Prentice­ useful to encode some clear examples ofthe effect ofthe Hall. system than to attempt to educe its principles. Ifdeliber­ REBER, A. S. (1976). Implicit learning ofsynthetic languages: The role ate rule-induction succeeds in identifying the underlying of instructional set. Journal ofExperimental Psychology: Human Learning & Memory, 2, 88-94. principles, then it grants perfect control; but ifit fails, it REBER, A. S., & ALLEN,R. (1978). Analogic and abstraction strategies can leave the learner with no basis for future judgments. in synthetic grammar learning: A functionalist interpretation. Cog­ In contrast, learning about instances grants an imperfect nition, 6, 193-221. but robust basis for future performance. REBER, A. S., KASSIN, S. M., LEWIS, S., & CANTOR, G. (1980). On the relationship between implicit and explicit modes in the learning ofa Moreover, the superiority of memorization over rule­ complex rule structure. Journal ofExperimental Psychology: Human eduction instructions is not an absolute fact; it depends on Learning & Memory, 6, 492-502. the way in which training items are shown. It occurs when WHITTLESEA, B. W. A. (1987). Preservation ofspecific experiences in items are presented one at a time, in random order (e.g., the representation of general knowledge. Journal ofExperimental Reber, 1976).However,when the items are presented in an Psychology: Learning, Memory, & Cognition, 13, 3-17. WHITTLESEA, B. W.A. (1993). Illusions offamiliarity. Journal ofExper­ organized fashion, all at once, performance in the rule­ imental Psychology: Learning, Memory, & Cognition, 19,1235-1253. learningconditioncan surpass thatinthe memorization con­ WHITTLESEA, B. W. A., BROOKS, L. R., & WESTCOTT, C. (\994). After dition (Reber, Kassin, Lewis, & Cantor, 1980). We con­ the learning is over: Factors controlling the selective application of clude that any style oflearning succeeds, to the extent that general and particular knowledge. Journal ofExperimental Psychol­ ogy: Learning, Memory, & Cognition, 20, 259-274. it is supported by the circumstances in which it occurs. WHITTLESEA, B. W. A., & DORKEN, M. D. (1993). Incidentally, things Fundamentally, better transfer following memoriza­ in general are particularly determined: An episodic-processing ac­ tion than following hypothesis testing merely indicates count of implicit learning. Journal ofExperimental Psychology: that the former learning does not consist of hypothesis General, 122, 227-248. testing. It does not indicate that that learning consists of NOTE abstraction, or that it is passive, or implicit, or ofa differ­ ent mode oflearning, controlled by different principles. 1. That this awareness is not true access to representations ofprior ex­ Memorizing instances is an activity that is different from perience is demonstrated by the fact that people regularly mistake novel hypothesis testing, but it is not a mysterious process re­ but grammatical items for items they have actually seen (cf Whittlesea & Dorken, 1993). In a recognition test, in contrast to their claims ofnot quiring a special implicit-learning mechanism. knowing what enables them to classify novel items, people falsely claim to be able to know which items they have seen previously. Our claim is Conclusion that metaknowledge is an inference based on the success ofcurrent pro­ Dienes and Berry (1997) have provided the most com­ cessing, not direct access to the representations ofpast events that con­ prehensive summary to date of evidence for a distinction trol current performance. between implicit and explicit learning and knowledge. (Manuscript received August 18, 1996; However, we cannot agree that this evidence makes a revision accepted for publication October 24, 1996.)