Psychological Review Copyright 1984 by the 1984, Vol. 91, No. 2, 269-276 American Psychological Association, Inc.

Theoretical Note

Automatic and Controlled Processing Revisited

Richard M. Shiffrin Walter Schneider Indiana University University of Illinois

The theory of automatic and controlled processing outlined in Schneider and Shiffrin (1977) and in Shiffrin and Schneider (1977) is defended in the present note. We argue that the criticisms of Ryan (1983) range from irrelevant to incorrect, based on a brief review of data from the 1977 articles and on some more recent publications. The evidence Ryan discusses comes from the prememorized-list paradigm, a par- adigm that undoubtedly involves automatic and controlled processes but probably not automatic detection and controlled search. We argue that a variety of mechanisms consistent with our general theory, some automatic and some controlled, could be operating in the prememorized-list paradigm and can explain the observed results.

A theory of automatic and controlled processing effortless process that is not limited by short-term was outlined and given empirical support in the memory capacity, is not under direct subject control, articles of Schneider and Shiffrin (1977) and Shiffrin and performs well-developed skilled behaviors. It and Schneider (1977). The characteristics of both typically develops when subjects process stimuli in types of processes were established through ex- consistent fashion over many trials; it is difficult to amination of particular examples of each of these suppress, modify, or ignore, once learned. Con- classes of processes. These examples, of critical im- trolled processing is often slow, generally serial, ef- portance in many search and attention tasks, were fortful, capacity limited, subject regulated, and is termed automatic detection and controlled search, used to deal with novel or inconsistent information. and their characteristics were determined empiri- It is needed in situations where the responses re- cally. quired to stimuli vary from one trial or situation In particular, in memory- or visual-search tasks, to the next, and is easily modified, suppressed, or consistent mapping (CM) refers to paradigms in ignored at the desire of the subject. Finally, all tasks which targets and distractors never exchange roles are carried out by complex mixtures of controlled over trials of the study. Varied mapping (VM) refers and automatic processes used in combination. (For to paradigms in which targets on one trial may be a more recent discussion see Schneider, Dumais, distractors on another, and vice versa. We dem- & Shiffrin, in press, and Schneider & Fisk, in onstrated that extended testing in CM paradigms press-a). led to a marked flattening of the set-size functions Ryan (1983) confuses the general concepts of au- and to a number of other prominent effects. We tomatic and controlled processing with the partic- termed the processes used by subjects, automatic ular instances. He argues, probably correctly, that detection. In contrast, the use of a VM procedure, automatic detection and controlled search cannot however extended, leaves intact the form and the explain the results from what we shall term the slope of the set-size function. We termed the pro- prememorized-list paradigm, but this fact is irrel- cesses used in this situation, controlled search. evant if automatic detection and controlled search Based on the findings, we postulated that auto- are not used to produce responses in this paradigm. matic processing is generally a fast, parallel, fairly We argue that other automatic and controlled pro- cesses are at work in the prememorized-list para- digm. Stripped of references to the prememorized-list This research was supported by National Institute of paradigm, Ryan's (1983) note contains little more Mental Health Grant 12717 and a Waterman Research Grant to the first author, and Contract ONR NOOOO-14- than polemics. No criticisms are directed toward 81-K-0034 (NR 150-460) and National Institute of Mental the research reported in our 1977 articles (Schneider Health Grant 31425 to the second author. & Shiffrin, 1977; Shiffrin & Schneider, 1977). In Requests for reprints should be sent to Richard M. those articles we reported roughly 14 new experi- Shiffrin, Psychology Department, Indiana University, ments the results from which were used to generate Bloomington, Indiana 47405. and elaborate the theory. Not one of our studies or

269 270 THEORETICAL NOTE results are mentioned by Ryan. We also used the value may well occur automatically, so that relatively theory to analyze a large number of studies in the few resources might be required to respond; in par- literature, clarifying what had been a somewhat dis- ticular, an explicit search of the memory set is prob- jointed set of findings. He mentions almost none ably not required. of these findings or the explications based on our theory and mentions none of the results in the ar- Automatic Detection and Categorical ticles published by us or by our colleagues since Classification 1977. Despite Ryan's expostulations, we think our 1977 exposition remains cogent and convincing and According to our theory, sufficiently extended based directly on the evidence. consistent testing in the prememorized-list paradigm should lead to the development of two types of au- Prememorized-List-Recognition Paradigm tomatic processes. First, the individual target stimuli should develop an automatic tendency to attract The results discussed in Ryan's (1983) note are attention. This is called automatic detection in the virtually all from the prememorized-list paradigm. 1977 articles and leads to list-length independence. The subject learns lists of items in advance of the Second, the target stimuli should coalesce as a uni- test session, to a criterion assuring that they can be tary category, such that a category code would be reproduced with high accuracy. The test session generated automatically whenever a member is pre- requires that the subject make yes-no decisions dis- sented. This code could then be used to judge list tinguishing list items from distractors not learned membership in a way that would not depend on on the lists. Reaction time is plotted as a function list length (see Shiffrin & Schneider, 1977, Exper- of memorized list length. iment 3). In the prememorized-list studies to date, Atkinson and Juola (e.g., 1973, 1974) suggested sufficiently extended testing has not been used to that subjects utilize a familiarity judgment to re- allow an inference that automatism of either type spond on most trials in this paradigm, with a rapid has developed, although it is conceivable that the search of long-term memory occurring on those training to criterion in advance of the test session(s) trials when the familiarity judgment was not easy. could have led to such development. We agreed that this model was plausible (i.e., Shiffrin The possible use of automatic detection and au- & Schneider, 1977, p. 176). However, we did not tomatic categorical classification is difficult to assess delve into this paradigm in detail because we felt in prememorized-list studies. If the only alternative that numerous mechanisms could be operating, and process were controlled serial search, then such the paradigm was poorly designed to distinguish processes would produce a flatter set-size function. between them. To make our views clear, we now However, processes like familiarity would also pro- consider the various types of processes that subjects duce flatter set-size functions, so more sophisticated could use in the prememorized-list paradigm. methods would have to be used to disambiguate these possibilities. Contextual Familiarity Perhaps the most likely candidate for a process Controlled Serial Search underlying responding in the prememorized-list In prememorized-list studies, the longest lists paradigm is temporal-contextual familiarity. After usually contain more items than can be maintained all, aside from the fact that the list is learned to in short-term store (STS). Thus we would not expect criterion before testing begins, the paradigm is quite controlled serial search to be used in such condi- similar to traditional recognition-memory tasks. In tions. However, there is at least one way in which such tasks most researchers have proposed that sub- subjects are quite likely to use controlled serial jects respond on the basis of familiarity at least search in this paradigm. They may read into STS some of the time. Familiarity judgments presumably some of the memory set (perhaps all of it in the entail an assessment of the recency of the test item, cases where the set size is small enough) and may or the list-context associated with the list item. Fa- carry out a serial search of just this portion of the miliarity can be quite simple as in the models of memory set, perhaps in parallel with some other Mandler (1980) or Atkinson and Juola (1974)—in process. Because such a strategy does not noticeably which cases increases of latency with list length improve performance at large set sizes, controlled would not be predicted—or quite complex, as in serial search is likely to play a significant role and the models of Ratcliff (1978), J. A. Anderson (1973), reduce latencies only when the set size is very small. and Gillund and Shiffrin (1984), in which cases increases of latency with list length would be pre- Controlled Search of Long-Term Memory dicted. (A closely related notion is that of strength as a basis for discrimination that has been proposed Mandler (1980), Atkinson and Juola (1974), and by Corballis, 1975.) The generation of a familiarity others have proposed that recognition involves two THEORETICAL NOTE 271 phases: familiarity and some sort of search of long- prememorized-list paradigm and, as we have argued, term memory. Raaijmakers and Shiffrin (1980, these are either irrelevant or incorrectly applied. 1981) have given a detailed exposition of one model Furthermore, the criticisms fail to take into account of long-term memory search. In their model, search the substantial array of data reported in the 1977 is a complex mixture of controlled and automatic articles that provided the evidentiary support for components and includes a sequential series of the theory. Note well that Ryan makes no attempt samples from long-term memory, so that latencies to provide alternative explanations for the data re- would surely rise with list length, possibly linearly. ported in the original articles. Given these facts, Other search models, such as those of J. R. Anderson there is no real need to carry this note further. (1976) and Ratcliff and Murdock (1976), would Nevertheless, we do not wish Ryan's remarks to also predict latencies to rise with list length. Note gain credibility through our failure to reply. There- that the availability of alternative processes, such fore we deal with these remarks briefly. as familiarity, might well eliminate the need for controlled search of long-term memory. Load Dependence and Independence Discussion and Summary It is doubtful that anyone could read our 1977 articles without learning that consistent mapping There is strong, independent evidence from a va- of targets and distractors produces automatic de- riety of paradigms that each of the processes dis- tection and small load effects in search tasks, and cussed above exists and can be used in its own right. that varied mapping produces controlled serial In theory, it should be possible for the subject to search and large load effects. For example, in a use any of these processes, either alone, in com- 1 search experiment whose results were given in Fig- bination together, or in mixtures across trials. Note ure 6 of Schneider and Shiffrin (1977), subjects also that all of these processes are consistent with searched for one of four memory-set characters in the general theory of automatic and controlled pro- either VM or CM conditions. As the display size cessing expressed in the 1977 articles. changed from one for four characters, latency rose Given the variety of processes that could be op- 240 ms in the VM condition and just 10 ms in the erating in the prememorized-list paradigm, and the CM condition. Ryan apparently feels, nonetheless, difficulty in choosing among them, it is hard to see that consistency of training does not determine load how the data from this paradigm can be used either effects because Sternberg (1966) found that his fixed- to support or to falsify the concepts of automatic set condition showed the same slope as his varied- and controlled processing. For example, because set condition. We dealt with this finding at length Ryan (1983) thinks that we would argue that con- in the 1977 articles (Schneider & Shiffrin, 1977, p. trolled search is used in the prememorized-list par- 39). In short, Sternberg's results were surely pro- adigm, he concludes that we could not explain the duced by the changes in the fixed set that occurred relatively flat reaction-time functions of set size that after relatively short blocks. Thus his fixed-set pro- can occur on even the first test of a target. Of course cedure was actually close to what we termed VM. we would expect such a result if the subject were Any reader who has doubts concerning the results using any of the processes of automatic detection, of true fixed-set training should feel free to invest familiarity judgment, or category classification. (We a few hours of experimentation: The marked flat- think that familiarity is most likely.) tening of the set-size function is trivial to produce In summary, no crucial conclusions for our theory and must have been replicated hundreds of times can be drawn to date from the results of pre- by the date of this writing. (See Fisk & Schneider, memorized-list studies because the design of the paradigm allows many mechanisms to operate and does not allow these to be disentangled. Because 1 The problems of deciding which processes are used we think that the prememorized-list results are very in the prememorized-list paradigm are compounded by likely the result of a familiarity judgment, or a mix- the typical training and test procedures. For example, the ture of this process with others, the criticisms of lists are learned either in an uncontrolled environment or our theory based on this work are weak at best and in a training session using a learning criterion for stopping. probably irrelevant. It is obvious that a list of 2 items can be learned virtually instantaneously, whereas many trials may be required to learn a list of 30 items. In what sense then can these lists Characteristics of Automatic Detection be said to be equated with respect to any of the processes and Controlled Search that have been discussed, such as attention attraction, cat- egory learning, familiarity, or strength? As another example, Ryan (1983) lists a number of characteristics of the test phase generally involves unequal numbers of tests controlled and automatic processes and argues that for items on lists of different length, and for targets and these are contradicted by existing data. All of the distractors, making it difficult to evaluate learning that substantive criticisms are based on data from the might be occurring during the course of testing itself. 272 THEORETICAL NOTE

1983, for another striking demonstration of this was presented in the 1977 articles. A different sit- phenomenon involving category search.) uation arises when distractors are used that do not Before leaving this topic, we feel it is important match the category of the memory set. In this case, to point out that zero slope is an ideal seldom ex- automatic category-code generation takes place in pected in actual studies using CM training. Even parallel with controlled search and allows truncation automatic processes will probably show a small load of the search and flattening of the set-size function effect. More important, the subject may well carry (see Jones & Anderson, 1982). Automatic detection out controlled search in parallel with automatic and controlled search can also be carried out to- detection, reducing the latency values at the smallest gether. An extended discussion and some persuasive set sizes (this issue is discussed in the next section). data from Ellis and Chase (1971) is given in Schnei- Thus, it is more important in search studies to note der and Shiffrin (1977, pp. 40-41). We have often the dramatic lowering of the slope as one moves observed in our own work that even lengthy training from VM to CM conditions—and to note the dra- in CM conditions leaves a small but nonzero slope matic practice effects in the CM conditions—than to the set-size function (e.g., see Dumais, 1979). to be overly concerned by set-size functions that This result is presumably due to speeded responding deviate from zero slope. at the smallest set sizes due to controlled search carried out in parallel with automatic detection. Role of Parallel Processes in Search Tasks The best examples of parallelism of automatic and controlled processes occur when two separate Ryan (1983) proposed that subjects in item-iden- tasks with distinct responses are to be carried out tification tasks can use two processes in parallel, simultaneously (rather than one task to be carried with the fastest triggering a yes response. He gives out by two parallel processes). An excellent dem- no details of the component processes, other than onstration of this sort has been reported by Schnei- to suggest that one depends on set size to a greater der and Fisk (1982-a). Subjects were asked to carry degree than does the other, so that one tends to out automatic detection on one diagonal of a four- trigger first at small set sizes and the other triggers character display and controlled search on the other, first at large set sizes. Presumably also, both pro- targets on either diagonal requiring responses. Con- cesses lead to linear functions of set size, because trol conditions required search on only one diagonal. the results of such studies are fit by Ryan with This study utilized a multiple-frame procedure. bilinear functions. He apparently feels that to the That is, a rapid sequence of frames of four characters degree that such a hypothesis is an accurate rep- each was presented, with the subject attempting to resentation of the findings, support for our theory detect the presence of targets on either diagnonal. is eroded. Detection accuracy was the dependent measure. The In fact, just the opposite is the case. As discussed automatic diagonal used consistent mapping of tar- at great length in the 1977 articles, one of the pri- gets and distractors; the controlled diagonal used mary characteristics of automatic processing is the varied mapping. ability of such a process to be carried out in parallel After sufficient practice, subjects instructed to with other processes, without cost. An exception to give maximum effort to the controlled task could this rule occurs if the attention system is itself in- carry out both tasks together at accuracy levels equal voked by the automatic process, in which case in- to those in the respective control conditions. The terference with another control process could take results are shown in Figure 1, in the left-hand panel. place. However, when the task is like those that Note that when both diagonals required controlled Ryan discusses, in which a positive response can search, even great amounts of practice could not be given on the basis of either of two processes— bring the joint-task accuracy levels anywhere close whichever triggers first—then this exception is ir- to the single-task levels. This is shown in the right- relevant. In such a case, there is nothing to hinder hand panel of Figure 1. the subject from using two processes in parallel, as In summary, it is an essential component of our long as at least one is automatic. If one of the pro- theory that automatic and controlled processes may cesses is controlled, then the subject might not occur in parallel, without cost, but that two or more choose to use it, but doing so would only lead to controlled processes cannot exhibit such a property.2 a benefit. Ryan (1983) claims that he (and Forrin & Morin, Whether two parallel processes can be used by the subject, and what the nature of these processes 2 are, depends of course on the task. In VM paradigms If two controlled processes operate together, do not stress short-term capacity, and are at ceiling, costs may of the kind used by Sternberg (1975), in which no not be seen. However, if one adds a third task requiring categorical distinction, automatic detection, or other capacity, so that short-term capacity is stressed, then the response basis is available, only controlled pro- costs caused by the first two control processes used together cessing is used—much evidence along these lines should be observable. THEORETICAL NOTE 273

678 A CM SINGLE VM SINGLE a' (DUAL EMPHASIZED! Figure 1. Detection performance in a multiple-frame search task, measured by A', in single-task and in dual-task conditions. (Single-task performance is given on the axes, whereas dual-task performance is given by the central points, for both tasks separately. Three frame times were used. The left-hand panel shows performance when one task was varied mapping, VM [controlled], and one was consistent mapping, [au- tomatic]. The right-hand panel shows performance when both tasks are VM [controlled]. [Data taken from Schneider & Fisk, 1982a.])

1969, although those authors made no such claim) data contradict our theory. Exactly the opposite is has obtained data showing independent and parallel the case. controlled processes. The sole basis for this claim As automatic detection begins to develop, it is is the single datum obtained in an experiment re- rather weak and slow and finishes first only when quiring search both through a prememorized list the controlled search is very long in duration. Thus and a short-term varied set: Namely, the first two a flattening of the set-size function is seen first when points on the set-size function for the prememorized set size is very large, and only later does such flat- set exhibits a slope value close to that seen for the tening occur for the smaller set sizes (see the dis- first two points on the varied-set function. cussion in Schneider & Shiffrin, 1977, p. 40). This The logic used here is not very crisp. Although result is predicted for all studies in which the same controlled serial search should give a linear set-size total set of CM targets is utilized in the different function with the typical slope value, the converse set-size conditions (although different instances is not true. Imagine an increasing set-size function randomly chosen from the total set are displayed produced by, say, an automatic familiarity mech- on particular trials). Our own studies use such a anism. This function might increase in negatively procedure, and even Ryan's studies use a nested accelerated fashion, and, somewhere on this func- procedure in which at least some stimuli appear in tion, two successive points might be found to exhibit all set-size conditions.3 a slope value typical of controlled search. It is ob- In summation, parallel processing involving au- vious that one would not want to conclude that tomatic mechanisms is an integral part of our theory. controlled search was being utilized for those two Controlled processing may be carried out in parallel particular set sizes, yet this is in effect what Ryan with automatic processing, at the subject's option, argued in his note. (Also, the data reported by Ryan but two controlled processes cannot be carried out can be accounted for on the basis of an automatic in parallel without the sharing of resources and familiarity process operating in parallel with con- some cost. The data that we know of are consistent trolled search.) with these proposals. Ryan makes another attack on our theoretical position that is rooted in the parallelism of auto- matic detection and controlled search. According 3 The tendency for flattening first to appear for larger to our theory, automatic detection is learned grad- set sizes, due to characteristics of a race between parallel ually in CM conditions; at the start of training the processes, is a factor in all designs. However, when disjoint subject must use controlled search, but as learning character sets are used for different set sizes, more practice may be given to stimuli in smaller sets, causing them to progresses, automatic detection begins to operate exhibit faster automatization and counteracting the first in parallel with controlled search and eventually tendency. In this event, our theory could not make a clear- provides the dominant response basis. Ryan observes cut qualitative prediction. However, in search tasks to date, that the effects of onset of automatic detection are and in Ryan's prememorized-list tasks, disjoint sets have seen at the largest set size first, and claims that such not been used. 274 THEORETICAL NOTE

Factors That Determine Automatization VM condition. Regardless of these ratios, the num- ber of times that any item was presented as a target We are a bit amazed by the suggestions in Ryan's was held constant (except for the last condition, (1983, p. 7) article that we have "no a priori theo- which received 10% fewer target presentations). The retical notion" concerning the factors that determine results showed that the first three conditions were attention demands. Much of our experimental work all superior to the VM control. However, even after has been designed to establish just these factors. 6,000 trials, the fourth condition was close to the Surely no one could read our papers without learn- VM control. It is obvious that inconsistency does ing that CM produces automatic attending to the more than just slow the rate of automatization. consistent targets. When an item appears as a distractor twice as often An experiment by Dumais (1979; reported in as it is used as a target, automatism may not develop Shiffrin & Dumais, 1981) gives additional infor- at all. In addition, the results demonstrated that the mation about the factors that produce attentional lower the consistency, the slower was the rate of demands. We have argued that CM training pro- automatization.4 duces enhanced attending to the consistent targets. The findings discussed in the preceding para- Is there any attentional effect for the consistent dis- graphs and the findings in the 1977 articles provide tractors? Subjects were trained extensively in both a great deal of knowledge concerning the conditions CM and VM search conditions until performance that produce automatic detection (and by inference, reached asymptote. Memory set size was always 1, automatism in general). They reinforce the conclu- and display size was both 4 and 16. The set-size sion reached in the 1977 articles: Consistency of slope after training was much larger, as usual, for training is the most important factor determining the VM condition. At this point, a variety of transfer automatization. conditions were carried out. In one case the CM targets were retained intact and new distractors were Control of Processing introduced (either new items or old VM items). In another case the old CM distractors were retained Ryan (1983) implies that subjects do not have intact and new targets were introduced (either new control over controlled processing (and presumably items or old VM items). In all of these conditions do have control over automatic processing). How- transfer was virtually perfect; that is, the set-size ever, his arguments are based solely on our failure slopes were about equal to those in the CM training to equate control processes with consciousness (pp. conditions. 4-5). Other researchers have tried to support such The model proposed by Dumais (1979) to ac- an identification (e.g., Posner & Klein, 1973), but, count for these results emphasizes the role of learn- although we discussed the issue, we were very careful ing to attend. Untrained stimuli are assumed to not to do so. Ryan seems to be arguing that we have a normal, intermediate tendency to attract should have made such an identification, and he attention when presented. Consistent training as a then objects because the identification has problems. target enhances the attention attracting tendency, This is a curious argument indeed. leading to good transfer for such CM targets. Con- Actually, we have quite extensive evidence that sistent training as a distractor decreases the atten- subjects can control their controlled processing. In tion-attracting tendency, leading to good transfer the 1977 articles we showed that subjects can attend when new targets are introduced, because the nor- to a designated diagonal of a display, and Shiffrin mal attention-attracting tendency for these new tar- and Schneider (1974) showed that attention could gets is higher than the reduced tendency for the be given to an emphasized memory-set item. The CM-trained distractors. attention literature is filled with examples such as Perhaps the most important questions concerning those from dichotic listening tasks, in which pro- the development of automatism involve the role of cessing can be directed toward a designated channel consistent training. We have seen that complete (say, a given ear). Also, knowledge of the location consistency (CM training) leads to automatic de- of a threshold visual stimulus improves detection tection, and we have also seen that highly incon- accuracy, even when eye movements are controlled sistent training (VM) does not allow automatic de- (e.g., Posner, 1978). tection to develop. Schneider and Fisk (1982b; also More interesting are demonstrations that auto- reported in Shiffrin, Dumais, & Schneider, 1981) matic processes are resistant to control. In Shiffrin have studied the degree of consistency necessary to and Schneider (1977) several examples of reversal produce automatism. Five conditions were used in which the percentages of time that an item spent 4 We refer the reader to a study by Schneider and Fisk, as target versus distractor were varied. The ratios reported in Shiffrin, Dumais, and Schneider, 1981, for a of target usage to distractor usage were 10:0, 10:5, related result concerning the advantages of finding targets 10:10, 10:20, and 9:61. The first of these conditions and the disadvantages of searching for targets without de- is a standard CM condition, and the last is a standard tecting them. THEORETICAL NOTE 275 training were given. After CM training, the roles of Finally, we want to respond to Ryan's criticism targets and distractors are reversed. Subjects could that our theory applies to a narrow range of stimuli not ignore the previous CM targets, with the result and tasks. The 1977 articles used a fairly narrow that performance dropped even below the VM level range of stimuli and tasks intentionally, to gain the seen at the start of training. Furthermore, relearning stimulus control necessary to reach firm conclu- under CM training conditions took longer than the sions. We of course hope that the principles derived original training (pp. 132, 135). In addition, it was apply quite generally. The interested reader is invited shown that CM targets could not be ignored, even to read results by Schneider and Fisk (1984) and when instructions were given to do so, and when Fisk and Schneider (1983), which generalize the the CM target appeared in a known-to-be-irrelevant stimuli to which these principles apply. display location (pp. 150-151). Conclusion Other Criticism Automatic processes and controlled processes are used in combination in all tasks. Automatic detec- A number of Ryan's (1983) points are unsup- tion and controlled search are particular examples ported, reactive statements, not only irrelevant but of these processes that are used in a surprisingly misleading and incorrect. For example, Ryan makes large number of visual-search and memory-search much of what he says is our claim to a "new and paradigms. Ryan seems to confuse the classes of independent theory". Actually, in the first paragraph processes with the particular examples, which leads of our 1977 article we pointed out similarities to him to use the prememorized-list paradigm to crit- Atkinson and Shiffrin's (1968) article dealing with icize our theory. We certainly think that automatic control processes, and then said: "However, despite and controlled processes are used in that paradigm, the similarities in basic aims, the experiments lie but we rather doubt that controlled search and au- in entirely different areas, and the present treatment tomatic detection play a very prominent role. is to be regarded as a new and independent theory The remaining criticisms raised by Ryan were to be evaluated in its own right" (Schneider & Shif- considered by us in the 1977 articles and were frin, 1977, p. 2). Independence from the 1968 ar- treated there in detail. Our present responses re- ticle, and independence from previous theories in iterate the points made then and allow us to extend general, are hardly equivalent concepts. Indeed, in and to further validate the theory by presenting the 1977 articles we discussed at length a number additional evidence collected since 1977 (but pub- of related proposals by other researchers (e.g., Shif- lished prior to Ryan's submission; we also call the frin & Schneider, 1977, pp. 171-184). reader's attention to newer work that extends these Ryan argued at length in his note that we were concepts: Fisk & Schneider, 1983, and Schneider doing nothing but renaming an old phenomenon & Fisk, 1984, in press-a, in press-b). We feel now and calling it a theory. It is obvious that Ryan failed as we did then, that the theory raises more questions to notice our presentation of 14 studies carried out that it answers. Although we are satisfied that the in our laboratory, not to mention our extensive present theoretical framework provides a coherent review of the literature. way of organizing a large number of empirical re- We want to respond to Ryan's suggestion that sults, the theory is far from immune to criticism, our theory can be criticized because he has obtained and we hope that continuing research leads to a data showing that automatic detection can occur fruitful evolution of the approach. without practice. The argument is based on his ob- servation in the prememorized-list situation that flattening of the set-size function occurs even on 5 Even when attention is restricted to stimuli that do the first tests of the experimental session. This ar- not attract attention automatically when the experiment gument is of course misapplied because learning begins, the amount of practice needed to product autom- potentially capable of producing automatic detec- atization varies considerably with factors such as stimulus- tion can occur during the training phase; more im- ensemble size and similarity of targets to distractors. We portant, a familiarity process would produce a non- have observed significant automatization taking place after linear function from the first test even in the absence just 10 target detections in some studies (although we often of automatic detection. Apart from these obser- give far greater training to subjects to ensure that automa- vations, we are quite aware that some types of au- tization is near asymptote). tomatic detection operate on the basis of innate and/or preexperimental factors; the theory allows for such occurrences. Examples include the atten- References tion demands caused by a sudden, loud noise or by Anderson, J. A. (1973). A theory for the recognition of a stimulus differing from all of the others on simple, items from short memorized lists. Psychological Review, physical characteristics like size, color, or motion.5 80, 417-438. 276 THEORETICAL NOTE

Anderson, J. R. (1976). Language, memory, and thought. Ratcliff, R., & Murdock, B. B., Jr. (1976). Retrieval pro- Hillsdale, NJ: Erlbaum. cesses in recognition memory. Psychological Review, 83, Atkinson, R. C, & Juola, J. F. (1973). Factors influencing 190-214. speed and accuracy of word recognition. In S. Kornblum Ryan, C. (1983). Reassessing the automaticity-control dis- (Ed.), Attention and performance IV (pp. 583-612). New tinction: Item recognition as a paradigm case. Psycho- York: Academic Press. logical Review, 90, 171-178. Atkinson, R. C., & Juola, J. F. (1974). Search and decision Schneider, W., Dumais, S. T, & Shiffrin, R. M. (in press). processes in recognition memory. In D. H. Krantz, Automatic and controlled processing and attention. In R. C. Atkinson, R. D. Luce, & P. Suppes (Eds.), Con- R. Parasuraman, R. Davis, & J. Beatty (Eds.), Varieties temporary developments in mathematical psychology of attention. New \brk: Academic Press. (Vol. 1). San Francisco: Freeman. Schneider, W., & Fisk, A. D. (1982a). Concurrent automatic Atkinson, R. C., & Shiffrin, R. M. (1968). Human memory: and controlled visual search: Can processing occur A proposed system and its control processes. In K. W. without resource cost? Journal of Experimental Psy- Spence & J. T. Spence (Eds.), The psychology of learning chology: Learning, Memory, and Cognition, 8, 261-278. and motivation: Advances in research and theory (Vol. Schneider, W., & Fisk, A. D. (1982b). Degree of consistent 2, pp. 89-195). New York: Academic Press. training: Improvements in search performance and au- Corballis, M. C. (1975). Access to memory: An analysis tomatic process development. Perception & Psycho- of recognition times. In P. M. A. Rabbitt & S. Domic physics, 31, 160-168. (Eds.), Attention and performance V(pp. 591-612). New Schneider, W., & Fisk, A. D. (1984). Automatic category York: Academic Press. search and its transfer. Journal of Experimental Psy- Dumais, S. T. (1979). Perceptual learning in automatic chology: Learning, Memory, and Cognition, 10, 1-15. detection: Processes and mechanisms. Unpublished Schneider, W., & Fisk, A. D. (in press-a) Attention theory doctoral dissertation, Indiana University. and mechanisms for skilled performance. In R. A. Magill Ellis, S. H., & Chase, W. G. (1971). Parallel processing in (Ed.), Memory and control of motor behavior. Amster- item recognition. Perception & Psychophysics, 10, 379- dam: North Holland. 384. Schneider, W., & Fisk, A. D. (in press-b). Processing with Fisk, A. D., & Schneider, W. (1983). Category and word and without long-term memory modification: Attention, search: Generalizing search principles to complex pro- level of processing, and word frequency. Journal of Ex- cessing. Journal of Experimental Psychology: Learning, perimental Psychology: Learning, Memory, and Cog- Memory, and Cognition, 9, 177-195. nition. Forrin, B., & Morin, R. E. (1969). Recognition times for Schneider, W., & Shiffrin, R. M. (1977). Controlled and items in short- and long-term memory. Ada Psychol- automatic human information processing: I. Detection, ogica, 30, 126-141. search, and attention. Psychological Review, 84, 1-66. Gillund, G., & Shiffrin, R. M. (1984). A retrieval model Shiffrin, R. M., & Dumais, S. T. (1981). The development for both recognition and recall. Psychological Review, of automatism. In J. R. Anderson (Ed.), Cognitive skills 91, 1-67. and their acquisition (pp. 111-140). Hillsdale, NJ: Erl- Jones, W. P., & Anderson, J. R. (1982). Semantic cate- baum. gorization and high-speed scanning. Journal of Exper- Shiffrin, R. M., Dumais, S. T., & Schneider, W. (1981). imental Psychology: Learning, Memory, and Cognition, Characteristics of automatism. In J. B. Long & A. D. 8, 237-242. Baddeley (Eds.), Attention and performance IX (pp. 223- Mandler, G. (1980). Recognizing: The judgment of previous 238). Hillsdale, NJ: Erlbaum. occurrence. Psychological Review, 87, 252-271. Shiffrin, R. M., & Schneider, W. (1974). An expectancy Posner, M. I. (1978). Chronometric explorations of mind. model for memory search. Memory and Cognition, 2, Hillsdale, NJ: Erlbaum. 616-628. Posner, M. I., & Klein, R. M. (1973). On the functions Shiffrin, R. M., & Schneider, W. (1977). Controlled and of consciousness. In S. Kornblum (Ed.), Attention and automatic human information processing: II. Perceptual performance IV (pp. 21-36). New "Vbrk: Academic Press. learning, automatic attending, and a general theory. Raaijmakers, J. G. W., & Shiffrin, R. M. (1980). SAM: Psychological Review, 84, 127-190. A theory of probabilistic search of associative memory. Sternberg, S. (1966). High-speed scanning in human In G. H. Bower (Ed.), The psychology of learning and memory. Science, 153, 652-654. motivation (Vol. 14, pp. 207-262). Academic Press. Sternberg, S. (1975). Memory scanning: New findings and Raaijmakers, J. G. W., & Shiffrin, R. M. (1981). Search current controversies. Quarterly Journal of Experimental of associative memory. Psychological Review, 88, 93- Psychology, 27, 1-32. 134. Ratcliff, R. (1978). A theory of memory retrieval. Psy- Received January 3, 1983 chological Review, 85, 59-108. Revision received May 25, 1983 •