Human Error

by James Reason

  Each schema was originally acquired in relation to a particular context. That contextual information forms an intrinsic part of the schema’s ‘knowledge package’. Any subsequent encounter with that context will raise the schema’s level of activation; thus, another way of looking at the idea of a contextual frame is as a cognate set of currently active schemata.

  While individual schemata are ‘context-dependent’, the cognitive system as a whole is not ‘context-bound’. As indicated earlier, one of the central properties of the attentional control mode is its power to transcend these immediate contextual constraints. But this demands processing resources that are not always available.

  1.4. Contexts are semantic as well as physical locations

  The idea of a semantic context is not novel. The notion was clearly expressed by Hotopf (1980) in his analysis of whole-word slips of the tongue:

  If the word we intend to speak is highly associated with another word that meets the contextual constraints operating within the utterance, then, given a certain time limit, that other word may be produced instead. The error word needs to be a word of high frequency or one whose threshold for production is lowered by other events occurring at the same time for it to have the necessary short latency in response. (Hotopf, 1980, p. 106)

  Some recent studies (Reason & Mackintosh, 1986), using phonological priming to manipulate ‘thresholds of production’, also provide strong evidence for the existence of semantic contextual constraints, at least in one rather specific verbal domain. As is well known to children, same-sound priming can be a very effective way of tricking people into producing what is immediately recognised as the wrong answer to a simple question. The example investigated was that described by Kimble and Perlmuter (1970):

  Q. What do we call the tree that grows from acorns?

  A. Oak

  Q. What do we call a funny story?

  A. Joke.

  Q. What sound does a frog make?

  A. Croak.

  Q. What is another word for cape?

  A. Cloak.

  Q. What do you call the white of an egg?

  A. Yolk (sic!).

  In the first of three very simple studies, we investigated the effects of varying the number of ‘oak-yolk’ primes upon the naming of the white of an egg (Reason & Mackintosh, 1986). A total of 80 subjects were used in the first study, divided into four groups of 20 subjects. Each subject was questioned individually, and no testing session lasted more than about 3 minutes. To establish a baseline, one group was simply asked the key question: “What do you call the white of an egg?” (zero primes). The remaining groups received either one, three or five ‘oak-yolk’ primes (selected randomly from the set of primes listed above) before being asked to name the white of an egg. Subjects were instructed to respond as quickly as possible with the first word that came to mind, and that single word answers were needed throughout.

  The results confirmed the potency of the ‘oak-yolk’ priming effect, and showed that erroneous ‘yolk’ answers increased with the number of prior rhyme primes. The frequency of the primes, as well as the presence of common phonological elements, clearly played an important part in determining the likelihood of a ‘yolk’ response. This is in general agreement with the schema-activating assumptions shown in Figure 4.2.

  Now that we have shown that the effects of phonological priming were incremental, at least within the ‘oak-yolk’ game, the next question concerns the selectivity of this priming. Was it restricted to a particular semantic subdomain, such as the contents of the egg, or could it spread to other parts of the ‘egg’ semantic context? An egg could be regarded as having several semantic subdomains. Only two, however, are of interest here: that relating to the contents (yolk, white, albumen) and that relating to the exterior (shell).

  In the second study, involving three groups of 20 students, subjects received either three, five, or seven ‘shell’ primes, and were asked the same key question (“What do you call the white of an egg?”). These primes consisted of a series of simple questions eliciting monosyllabic answers rhyming with ‘shell’ (e.g., “What rings in a belfry and on a bicycle?”, “From where did people once draw water?”, “The Swiss folk hero that shot an apple off his son’s head was called William ...?”)

  The most important finding was that only 2 of the 60 subjects produced an erroneous, priming-induced, ‘shell’ response to the key question, despite the fact that ‘shell’ is almost as strongly associated with ‘egg’ as ‘yolk’ (Palermo & Jenkins, 1962; Rubin, 1983). These findings are consistent with the notion that the phonological priming effects were confined to the ‘egg contents’ subdomain (as directed by the semantics of the key question).

  An alternative way of testing the relative impermeability of the contextual frame was to prime ‘yolk’ and then direct the semantics of the key question not to the white of an egg, but to the outside of an egg. In the third study, two groups of ten students were used. One group was simply asked the question: “What do you call the outside of an egg?” The other group was asked the same question after first receiving five ‘oak-yolk’ primes. The results were very clear-cut All the subjects in both groups produced the correct ‘shell’ answer.

  Taken as a whole, the results of these three studies provided a clear demonstration of (a) the incremental potency of phonological or ‘shared-element’ priming within the appropriate semantic context and (b) the relative absence of any ‘spillover’ of these priming effects from one contextual frame to another. This suggests that although prior priming may initially activate all words in the mental lexicon sharing the same phonetic and structural characteristics, these effects are only ‘delivered’ to a very specific semantic context by the final key question. This is also borne out out by the absence of any interference from one priming question to the next: each addressed a different semantic context, and none had any high-frequency competitors sharing common elements.

  2. Similarity and frequency: Cognitive ‘primitives’

  Such fundamental aspects of experience as the degree of likeness between events or objects and their frequency of prior occurrence have been termed intuitive concepts. Similarity and frequency information appear to be processed automatically without conscious effort, or perhaps even without awareness, regardless of age, ability, cultural background, motivation or task instructions (see Wason & Johnson-Laird, 1972; Schweder, 1977; Tulving, 1983; Hasher & Zacks, 1984). There is a strong case for regarding them as being the computational ‘primitives’ of the cognitive system. This argument will be developed further in Chapter 5.

  3. Demonstrations of underspecification

  3.1. Word identification

  The large volume of literature dealing with the recognition of words presented for very brief durations or in noise has been discussed elsewhere (see Howes & Solomon, 1951; Solomon & Postman, 1952; Pollack, Rubinstein & Decker, 1960; Savin, 1963; Spence, 1963; Newbigging, 1961; Neisser, 1967; Broadbent, 1967; Catlin, 1969; Nakatani, 1973). A number of theories have been advanced to explain the word frequency effect, the repeated finding that common words are more readily recognised than infrequent ones when presentation is rapid or attenuated. Newbigging (1961) has provided a useful summary statement: “When a word is presented at a short duration, only a few letters or a fragment of the word is seen by the subject. This fragment may be common to a number of words, and if the subject is instructed to guess the word presented he will respond with the word of the greatest frequency of occurrence (response strength) which incorporates the fragment” (quoted by Neisser, 1967; see also Neisser’s discussion of ‘fragment theory’, pp. 115-118).

  An important corollary to this view is that seen fragments of low-frequency words are prone to being erroneously perceived as high-frequency words sharing the same features (Broadbent, 1967). Both are in keeping with the cognitive underspecification generalization.

  3.2. The recall of verbal list items

  The word frequency effect also appears in the
recall of word lists. Commonly occurring words tend to have a higher probability of recall than less frequent words. Likewise, high-frequency words are more common as intrusions when the subject is attempting to recall a previously presented list. (see Gregg, 1976, for an excellent review of this and the related ‘frequency paradox’ literature).

  3.3. Category generation

  In continuous recall tasks, the only specification offered to subjects is the name of a semantic category, whereupon they are asked to generate as many exemplars as they can without repetition. Beginning with Bousfield and Barclay (1950), several investigators have found substantial correlations between the dominance of a particular item (the number of times a given exemplar is generated by a group of individuals) and its average position in the output order. The most popular exemplars appear earliest. In the case of Bousfield and Barclay’s categories, these correlations ranged between 0.97 and 0.88. In a later study involving 56 semantic categories, Battig and Montague (1969) obtained dominance/output order correlations (for the Maryland sample, N=270) ranging between 0.252 and 0.857. When these correlations are averaged using Fisher’s z-transformations, the mean dominance-order correlation is 0.64. When the same correlations are computed for only the 20 most dominant types in each category, this mean value rises to 0.76. Such findings are in keeping with the predictions of the ‘spew hypothesis’ (Underwood & Schulz, 1960), of Zipf’s Law (Zipf, 1945,1949) and of Hull’s habit strength postulate (see Bousfield & Barclay, 1950).

  The dominance of a particular exemplar within a semantic context (e.g., dog, within the category ‘four-footed animals’) reflects its ‘salience-in-the-world’ for that particular subculture. This, in turn, is likely to correspond to its frequency of encounter by the subculture as a whole. The dominance orders of categories whose types are subject to the influences of fashion or location will vary according to time and place. Thus, when American college students of a generation ago were asked to produce exemplars of ‘a type of dance’, the first seven types in the dominance order were: waltz, frug, twist, fox-trot, cha-cha, monkey and jerk (Battig & Montague, 1969). It is doubtful whether many, or even any, of these types would appear high in a contemporary listing. The dominance-orders of other categories (e.g., precious stones, units of time, relatives, metals, etc.), being universals, are extremely stable across groups within the same broad culture.

  In a recent category generation study carried out in our laboratory, the relationship between output order and frequency of encounter was assessed directly, rather than via dominance assumptions. Different groups of undergraduates were each asked to produce up to 20 exemplars of nine categories, varying in their degree of search specification. The data were collected in individual interviews, lasting about 30 minutes. After they had generated their items, subjects were asked to make metacognitive ratings, on a scale of one to seven, of the following features of each exemplar (fed back to them in random order): recency of encounter, frequency of encounter, feeling of knowing about the item, strength of affect and effort required to elicit the item. The average correlations with output order (computed across all nine categories using Fisher’s z-transforms) were: -0.70 for recency, -0.73 for frequency of encounter, -0.66 for feeling of knowing, -0.55 for strength of affect and + 0.87 for effort. Negative correlations indicated that the most highly rated items were generated soonest.

  This pattern of correlations clearly indicates that items produced early in the output sequence are judged as more recent, more frequent, more known about and more affectively ‘charged’ than those produced later. In general, these data provide strong support for the assertion that underspecification causes the most activated schema to be called to mind first. Although these correlations do not provide unambiguous evidence for the primacy of frequency, it is not unreasonable to assume that factors like recency, knowledge and affective tone depend heavily upon the number of times a particular item has been encountered in the past (see also Matlin, Stang, Gawron, Steedman & Derby, 1979).

  3.4. Recurrent intrusions in blocked memory searches

  Freud (1901) noted that when we are laboriously searching memory for a known name or word whose retrieval is temporarily blocked (the target), other items, “although immediately recognised as false, nevertheless obtrude themseves with great tenacity.” This observation provoked Reason and Lucas (1984a) to test the hypothesis that these ‘recurrent intruders’ in tip-of-the-tongue states (TOTs) were analogous to the strong habit intrusions observed in action slips.

  Sixteen volunteers kept diaries of their TOT states over a period of four weeks. For each resolved TOT, they were required to make a set of standardised ratings. The study yielded data for 40 TOTs. Twenty-eight of these (70 per cent) involved the presence of recurrent intruders: recognisably wrong names or words that continued to block access to the target during deliberate search periods.

  These recurrent intruders were judged by the diarists as being more frequently and recently used than both the related targets and the nonrecurrent intermediate solutions. The diarists also rated recurrent intermediates as being more closely associated with the target than other intermediates generated during that particular TOT state. In 50 per cent of the TOT states with recurrent intruders, the blocking word/name was ranked higher than the target for frequency and recency in the context of the search. In 77 per cent of the blocked TOTs, the recurrent ‘blocker’ was ranked higher than the target on either frequency, recency, or both. Comparisons of these recurrent intruders with the eventually-retrieved target revealed that they almost invariably shared some common phonological, syllabic, structural or semantic features.

  These data are consistent with the view that recurrent intruders emerge in TOT states when the initial fragmentary retrieval cues are sufficient to locate the general context of the sought-for item, but not to provide a unique specification for it. Two conditions need to be satisfied for ‘blockers’ to appear. The first is that the recurrent ‘blocker’ has a high level of activation at the time of the search, where this is probably due to the high frequency of its prior use. The second is that its close structural and/or phonological resemblance to the target make it compatible with some of the incomplete ‘calling conditions’ (or retrieval cues) available at the outset of the search. These, in turn, supply the additional activation necessary to bring it to mind. Subsequently, this ‘blocker’ is likely to recur due to the incrementing effects of the frequency/recency general activators. This unprofitable retrieval cycle will also gain added momentum from the continued sense that its products—the recurrent blockers—are ‘warm’ with respect to the target word. Usually, the only way to break this deadlock is to deliberately discontinue the search. Over 30 per cent of the targets thus blocked were recovered as spontaneous ‘pop-ups’, most often during the execution of some routine activity like washing up or driving.

  3.5. Slips or the tongue

  As with TOT states, slips of the tongue show marked similarities between the actual and target (intended) utterances (see Fromkin, 1973, 1980). According to Dell and Reich (1980, p. 281): “The more similar a given unit is to an intended unit, the more likely the given unit or a part of it will replace the intended unit or a corresponding part of it.”

  Different kinds of similarity effect are likely to operate at the various stages of formulating and executing the articulatory program. Fromkin (1971) has suggested that semantically-related substitutions may occur because of underspecification of the semantic features. At the more detailed level of phoneme specification, however, substitutions are facilitated by a phonological similarity between word segments (Nooteboom, 1969; MacKay, 1970).

  3.6. Slips of action

  Diary studies of naturally-occurring slips of action (see Chapter 3) reveal that they are most likely to be committed during the performance of highly automatised tasks in very familiar surroundings while experiencing some form of attentional ‘capture’ (preoccupation or distraction). A large proportion of these slips (40 per cent in one s
tudy) took the form of well-organised action sequences that were judged by their makers to belong to some other task or activity not then intended. These ‘other activities’ were rated as being recently and frequently executed and as sharing similar locations, movements and objects with the intended actions. The largest single category of action slips were strong habit intrusions, possessing structural or contextual elements in common with the planned actions.

  3.7. Failures of prospective memory

  Except in the case of spontaneous action (see Chapter 1), there is usually an interval between the formulation of an intention and the moment planned for its execution. During this time, the intention must be held in prospective memory. Relatively little is known about the properties of this ‘intention store’ (Harris & Wilkins, 1982; Baddeley & Wilkins, 1984; Harris, 1984). What is certain, however, is that it constitutes one of the more vulnerable parts of the memory system and is thus a major source of cognitive underspecification. Failures of prospective memory—forgetting to remember to carry out intended actions at the appointed time and place—are among the most common forms of human fallibility (Reason & Mycielska, 1982).

  3.8. Planning for uncertain futures

  Plans are, of necessity, based upon ‘best guesses’ about the future state of the world. These emerge from a complex interaction between the perception of its current state and the recall of previous states. Both the perceptual and the knowledge elements of this appraisal can be incomplete or incorrect. Whatever the precise source of the underspecification, however, the outcome tends to favour the selection of either salient (vivid) or familiar (frequently encountered) scenarios for future action (see Nisbett & Ross, 1980; Kahneman, Slovic & Tversky, 1982; Fischhoff, Lichtenstein, Slovic, Derby & Keeney, 1981). We will be looking in some detail at the effects of incomplete and inaccurate knowledge in Sections 5 and 6.