Human Error

by James Reason

  Barbara Tuchman (1962, p. 38), writing about the German plan for waging the First World War, stated: “Dead battles, like dead generals, hold the military mind in their dead grip, and Germans, no less than other peoples, prepare for the last war.” These words capture a universal truth about human planners. Our cognitive systems make us prisoners of the past, and this offers only a very approximate guide to the future.

  3.9. Pathological underspecificatlon

  Erroneous behaviours associated with various kinds of mental disorders and brain damage can also be viewed as arising from underspecification. Three examples will serve to illustrate this point.

  3.9.1. Frontal lobe damage

  Luria described the disturbances of planning and the difficulties of sustaining intentional activity that stem from massive lesions of the frontal lobes; for example: “One such patient... when asked to light a candle struck a match correctly but instead of putting it to the candle which he held in his hand, he put the candle in his mouth and started to ‘smoke’ it like a cigarette. The new and relatively unstabilized action was thus replaced by a more firmly established inert stereotype. I have observed such disturbances of a complex action program and its replacement by elementary, basic behaviour in many patients with a clearly defined frontal syndrome” (Luria, 1973, pp. 199-200).

  With such extensive damage, it is difficult to pinpoint any one type of underspecification; but the problem seems to lie mainly in the inability of the damaged frontal lobes to formulate or sustain an organized plan of future action. The goal-directed linkages between one action schema and the next seem to be weakened, or perhaps were never properly established, so that the course of action is readily ‘captured’ by the presence of well-trodden actions paths in the ‘vicinity’. Similar observations have been made by Tim Shallice (1986) and Brenda Milner (1986). These pathological behaviours closely resemble the slips of normal individuals in conditions of ‘reduced’ or ‘suspended’ intentionality (see Chapter 3).

  3.9.2. Schizophrenic thought and language

  Also working in a clinical context, but this time with schizophrenics, Chapman and Chapman (1973) have argued that many of these patients’ apparently bizarre utterances reflect an excessive yielding to normal biases, in particular to the tendency to substitute strong associates for appropriate responses. As an illustration, they cite Bleuler’s woman patient who, when asked to list the members of her family, began with ‘Father, son,’ but then concluded with, ‘and the Holy Ghost’. They provided an impressive body of experimental evidence in support of this idea (see Chapman & Chapman, 1973, pp. 119-136).

  The Chapmans have also shown that these tendencies towards strong associate substitutions are present to a lesser degree among normal controls, and show up most clearly when the correct response selection is underspecified, or when the task is too difficult. For example, normal subjects were found to choose associative alternatives on multiple-choice vocabulary items when they did not know the correct answer (Boland & Chapman, 1971; Willner, 1965; see also Section 5, this chapter).

  Many authors have offered attentional deficiencies as the basis for the prevalence of these strong-associate substitutions in schizophrenic thought and language. Thus, Jung (1906) proposed that schizophrenics suffer ‘diminished’ or ‘relaxed’ attention that leads them to produce familiar rather than appropriate assodates. Similarly, Broen and Storms (1966) and Venables (1964) have argued that schizophrenics are characterized by ‘narrowed’ attention, which in turn may be due to heightened drive (Easterbrook, 1959). There is clearly a close correspondence between these notions and the idea of attentional underspedfication at critical points in an action sequence (see Chapter 3). In the case of normal people, these monitoring failures are transitory and relatively infrequent. In certain kinds of schizophrenia, however, the mechanisms directing such attentional checks appear to be chronically impaired.

  3.9.3. Limb apraxias

  Roy (1982) was struck by the similarities between the slips of action observed in normals and those exhibited by patients with limb apraxias, due to cortical lesions in the motor system. He identified three error types common to both normal and pathological states, though occurring far more frequently in the latter. The most common type is the tendency to repeat (perseverate) actions. The second major type is one in which the actions are performed correctly, but in relation to the wrong objects (e.g., using a pencil as a comb). A third type involves the omission of an action altogether. The patient may be aware of what he must do, and can describe how to perform the action, but he is unable to initiate the action on command.

  The precise nature of the underspedfication in these patients is not clear. Roy’s (1982) view is that these errors are multiply determined, that is, many different agencies are likely to be involved in their production. In some cases, the underspedfication arises at the level of intention; in others, it appears to originate in the mediating neural mechanisms. In both, however, there is a tendency for the actions to be overspecified by local, situational factors. One of the consequences of brain damage is that it often leaves the patient excessively context-dependent (Goldstein, 1942; Luria, 1973)

  4. Taking stock

  My purpose so far has been to present sufficient evidence to demonstrate the possible varieties of cognitive underspedfication. An inadequate specification may arise for several reasons and implicate a number of different cognitive mechanisms: incomplete or ambiguous inputs, fragmentary retrieval cues, incomplete or inaccurate knowledge, losses from prospective memory, spillage from the conscious workspace, intentional limitations and either normal or pathologically-induced failures of attentional monitoring. But regardless of this diversity of the possible kinds of underspedfication, the outcomes are remarkably similar. The cognitive system tends to generate responses that (a) are more familiar, more expected and more frequently-encountered than those that might have been intended or judged correct and (b) are context-bound in two senses: they conform both to the current physical situation and to the ‘semantic context’ of the prevailing intention. Another way of expressing this generalisation is to say that when cognitive operations are underspecified, they manifest both frequency and similarity biases to a greater degree than would have occurred had the ‘small print’ of the activity been more precisely stated by either ‘top-down’ or ‘bottom-up’ processes.

  Whereas the effects of frequency are relatively clear-cut, those of similarity are more subtle. Error forms can resemble the properties of both the current intentional specification and the prevailing environmental cues in varying degrees. The most obvious tendency in mistakes—particularly those in which the problem solver has been limited by an incomplete or incorrect knowledge base—is for the error forms to be shaped by salient features of the problem configuration. In both lapses and slips, however, there can be matching to both intentional and contextual cues. The forms of these execution failures may show close similarities to the intended word or action, as well as, on occasions, being appropriately matched to the situation in which they occur. Exactly what is matched by the error form appears to be related to the extent to which the correct response is, or could be, specified at the outset of the thought or action sequence. Difficulties encountered at the level of formulating the intention or plan tend to create errors that are moulded primarily by immediate contextual considerations; those that occur at the level of storage or execution may reflect the influences of both intentional and environmental ‘calling conditions’ (specifiers). But irrespective of the precise nature of the matching, similarity effects are evident across all error types.

  5. Convergent and divergent memory searches

  In the remainder of this chapter, we shall be concerned with the effects of underspecification upon the retrieval of semantic knowledge items. It is useful to preface this discussion with a brief consideration of two different kinds of memory search: convergent and divergent searches.

  If a person were to be asked: What has four legs, barks, wags
its tail, is usually friendly, has an acute sense of smell, cocks its leg and is called man’s best friend?—there is little doubt that he or she would answer ‘a dog’. If the same person were asked to generate exemplars of the category ‘four-legged animal’, there is also a very strong possibility (see Battig & Montague, 1969) that the response ‘dog’ would occur very early in the output list, probably in the first position.

  Although these two requests yield very much the same initial response, it is clear that they are mediated by two quite different kinds of search. In the first case, sufficient cues (calling conditions) were supplied to identify uniquely a single knowledge structure: dog. In this convergent mode of search, the response was produced entirely by the similarity-matching of supplied cues to stored attributes. In the category generation example, however, only one very imprecise cue was given: the class of animals having four legs. Under these divergent and underspecified search conditions (as seen above), the output is largely determined by frequency-gambling. The first items to be retrieved have been most often encountered in the past. These differences are illustrated diagrammatically in Figure 4.3.

  Convergent and divergent search processes can be thought of as the extremes of a continuum. At the convergent end, the similarity-matching heuristic is dominant, whereas frequency-gambling prevails at the divergent end. In everyday life, however, many searches for stored knowledge items tend to lie somewhere between these two extremes. The calling conditions generated during the course of planning or decision making, for instance, are neither as precise as those in the convergent example nor as vague as in the divergent case. These intermediate searches are likely to produce partial matches between the calling conditions and a limited number of knowledge items. This ‘fuzziness’ is due to either imprecise or fragmentary retrieval cues or to incomplete stored knowledge. These two sources of underspecification will, for theoretical purposes, be treated as functionally equivalent.

  Figure 4.3. Contrasting two extremes of memory search: convergent search, in which the supplied cues uniquely specify a particular knowledge item, and divergent search, in which only the category is given at the outset.

  6. The retrieval of incomplete semantic knowledge

  A cognitive activity in which it is possible to vary both the specificity of the ‘calling conditions’ and the adequacy of the stored schemata is that of answering general knowledge questions. Given some approximate idea of how much (or how little) information a group of individuals has about a particular knowledge domain, and providing the questions are posed in a way that encourages guessing, the answers obtained tend to be highly predictable. For example, ask any British lecture audience to identify the source of the following quotation: “The lamps are going out all over Europe; we shall not see them lit again in our lifetime,” and there is a strong probability that most of them will attribute it to Winston Churchill. Neville Chamberlain and Franklin D. Roosevelt will also receive a few mentions. Hardly any will name the correct source of this far-sighted observation: Sir Edward Grey, British Foreign Secretary, spoken at dawn on 3 August 1914 to the Italian ambassador as the gas lamps were being extinguished on Horse Guards Parade.

  Aside from the fact that the person who actually made this eerily familiar remark is now a fairly obscure figure, it is not difficult to see why such a predictable misattribution occurs. The topic and style of the quote point to a statesman, the language is English and the sense locates it on the eve or during the early days of a world war. For most people questioned, this combination of cues inexorably signifies just one person and one period: Churchill, the great speechmaker, at the outset of the Second World War. With the exception of those with specialist knowledge, the category ‘English statesman on the eve of a world war’ contains, for most British people, perhaps three or four names, of whom Churchill’s is the most frequently encountered (even though he did not hold office in September 1939).

  This section is concerned with the cognitive processes that yield such common but wrong answers to general knowledge questions, particularly in domains where the information in semantic memory is both relatively sparse and unevenly distributed. We begin by sketching out the processes involved in knowledge retrieval, and then discuss the findings of three studies that test certain hypotheses derived from this view.

  6.1. Retrieval mechanisms

  It is argued that knowledge retrieval involves at least four kinds of cognitive activity: a metacognitive assessment of whether or not the sought-for item is likely to be available in semantic memory; the similarity-matching of retrieval cues to the attributes of stored knowledge structures; resolving conflicts created by the partial matching of several ‘candidates’ by gambling in favour of high-frequency alternatives and the inferential work performed by the conscious workspace, the product of which is a revised set of ‘calling conditions’.

  6.1.1. Frequency information as the basis of epistemic awareness

  A major determinant of the effort that will be invested in searching for the answer to a general knowledge question is what Klatzky (1984) has termed epistemic awareness: the feelings of knowing (FOK) about what one knows. There is now a considerable body of evidence to show that adults are able to make reasonably accurate judgements of their subsequent success in an effortful memory search (Hart, 1965; Arbuckle & Cuddy, 1969; Blake, 1973; Gruneberg & Sykes, 1978; King, Zechmeister & Shaughnessy, 1980; Zechmeister & Nyberg, 1982; Reason & Lucas, 1984a).

  A substantial proportion of everyday memory work can be likened to directing a very small ‘access window’ through an enormous knowledge base in an effort to locate stored items of information that satisfy current demands. Without the ready availability of something akin to an index or knowledge map, this would be an exceedingly difficult task. If it were necessary to search in an undirected and serial fashion for every item of sought-for knowledge, an unacceptably large part of mental life would be consumed in retrieval activities. Metamemoric FOK assessments clearly play a critical role in the management of cognitive affairs. But upon what basis are these assessments made?

  Of special relevance to this question is how a person decides that he or she does not know something. Glucksberg and McCloskey (1981) proposed a two-stage model of question-answering to account for different types of ‘don’t know’ decisions: the one rapid and confident, the other slow and uncertain. In the first stage, a preliminary search of memory is conducted to locate any stored information that may be relevant to the question. If nothing is found, a rapid ‘don’t know’ decision is made. But if relevant facts are found, these are examined in detail to determine whether they indicate an answer to the question. If these retrieved facts fail to provide an answer, a slow ‘don’t know’ decision results. In either case, however, the discovery of relevant facts is expected to delay the outcome.

  Unfortunately, this account begs the question of how the preliminary memory search is made. Consistent with this chapter’s general theme, it is proposed that initial appraisals of the knowledge base depend critically upon the immediate availability of crude frequency data regarding the number of times particular stored items have been encountered in the past or have co-occurred with other items. This view rests on two well-founded assumptions regarding the status of frequency information within the memory system. First, frequency-of-encounter information appears to be stored in memory by “an implicit or automatic encoding process” (Hasher & Zacks, 1984). Second, most of the experimental evidence favours a multiple trace hypothesis in which each encounter with a given item is recorded as an additional trace on a ‘pile’ of like traces (Hintzman, 1976; Hintzman, Nozawa & Irmscher, 1982). According to this notion, frequency information is represented in a privileged analogue form (i.e., by the ‘height’ of the ‘pile’ of like traces). It is also assumed that each individual trace preserves episodic information regarding both the context of the encounter and connections with co-occurrent items.

  What is being proposed here, then, is that people are able to make more o
r less instant ‘frequency maps’ of particular semantic regions in a way that is analogous to brain heat scans. The metamemorial information so obtained, though crude and relative, forms the basis of a memory management system by which the limited ‘access window’ is steered (or drawn) towards promising areas of the knowledge base. By the same token, if a ‘scan’ throws up little or no frequency information—either in regard to the previous occurrence of single knowledge items or concerning the co-occurrence of multiple items— then there is little to be gained by searching further. The crucial point is that frequency-derived epistemic awareness of the ‘fatness’ of a particular knowledge ‘file’ appears to be available to the memory searcher before he or she has had an opportunity to examine its contents in any detail. This is in keeping with Oldfield and Wingfield’s (1964) notion that crude assessments of frequency are as immediate as the act of recognition itself. When we recognise something, we not only know that we have it in memory, we also know approximately how much is known.