by James Reason
6 The detection of errors
* * *
So far, we have focused mainly upon the causes of errors, that is, upon the conditions that precede their occurrence and on the cognitive mechanisms that shape their more predictable forms. For the remaining chapters, the emphasis will shift towards their consequences, beginning here with a consideration of the processes involved in the detection and recovery of errors.
To err is human. No matter how well we come to understand the psychological antecedents of error or how sophisticated are the cognitive ‘prostheses’—devices to aid memory or decision making—we eventually provide for those in high-risk occupations, errors will still occur. Errors are, as we have seen, the inevitable and usually acceptable price human beings have to pay for their remarkable ability to cope with very difficult informational tasks quickly and, more often than not, effectively. In conditions where “machines botch up, humans degrade gracefully” (Jordan, 1963). But, as we shall discuss further in the ensuing two chapters, the centralised supervisory control of complex, hazardous, opaque, tightly-coupled and incompletely understood technologies can, on occasions, transform these normally adaptive properties into dangerous liabilities.
If it is impossible to guarantee the elimination of errors, then we must discover more effective ways of mitigating their consequences in unforgiving situations. Many have suggested that this is really the only sensible way of combating the human error problem in high-risk technologies. The first step along this path is to consider what is known about the means by which slips, lapses and mistakes can be detected and recovered.
Despite the obvious importance of the topic, the psychological literature contains very little in the way of empirical studies of error detection or of theories to explain the processes by which people catch and correct errors made either by themselves or by others. What studies there are usually relate to very specific tasks such as reading (Carpenter & Daneman, 1981), writing (Hayes & Flower, 1980) or solving statistical problems (Allwood & Montgomery, 1981, 1982; Allwood, 1984). Only a few investigators have looked at error detection in real-life settings. Among these, the most important findings have been obtained by Woods (1984), who observed how nuclear power plant operators discover their errors in real and simulated emergencies, and by Rizzo, Bagnara and their co-workers (Rizzo, Bagnara & Visciola, 1986; Bagnara, Stablum, Rizzo, Fontana & Ruo, 1987), who investigated the use of a computer database by novices and the production plans made by experienced operators in a steel mill. Other investigators have carried out naturalistic studies of specific error types: slips of action (Norman, 1981) and slips of the tongue (Nooteboom, 1980). Despite the paucity of this error detection research, however, it is possible to give a preliminary account of human error detection and of the factors likely to impede it—though such a discussion will undoubtedly raise more questions than it answers. We begin by considering the various agencies that could be involved in discovering an error.
1. Modes of error detection
Both everyday observation and common sense have shown that there are probably only three ways in which people’s errors are brought to their attention. Most directly, they can find out for themselves through various kinds of self-monitoring. Second, something in the environment makes it very clear that they have gone astray. Or third, the error is discovered by another person who then tells them. Each of these three detection modes is considered further below.
2. Self-monitoring
Just as the control of behaviour resides at several different levels within the nervous system, from spinal nerve reflexes at one extreme to the effortful ‘on-line’ conscious guidance of unpractised actions at the other, so also are error detection mechanisms located at various points across this multilevel control structure. They range in complexity from the automatic correction of postural deviations to the thought-intensive direction of knowledge-based performance.
The essence of feedback control (as opposed to feedforward control) is that it is error driven. Deviations of output from some ideal or desired state are fed back to the controlling agency, which then acts to minimise these discrepancies. This basic loop structure (see Figure 6.1) is present at all levels of action control. What differs between the levels is the means by which discrepancies are detected and the extent to which attentional (as opposed to automatic) processes are involved in their correction. This, in turn, carries implications for the various ways in which the ‘ideal or desired state’ can be defined and represented.
Although we may not know a great deal about the precise operation of these various error detection mechanisms, one thing is clear: the higher the ‘level’ at which they operate, the more they are themselves error prone. Selected examples of error detection processes are considered below, beginning at a relatively low level with posture control.
Figure 6.1. A basic feed-back loop in which the output signal is compared to a reference input signal. The difference between the output and input signals (the error signal) constitutes the input to the controller, which then acts to minimise the discrepancy. The system is thus error driven.
2.1. The automatic correction of postural deviations
Virtually all animals, even very primitive ones, possess some means to keep themselves oriented with respect to gravity. Unstable bipeds such as human beings could not maintain their upright posture without the continuous activity of these wholly automatic mechanisms. For the most part, they do their job extremely well; our dependence upon them only becomes apparent when one or other of their components is damaged. Then the consequences, in the case of sudden onset, are catastrophic. Aside from being unable to stay upright, the sufferer experiences violent dizziness, nausea, vomiting and something very akin to acute depression—like a prolonged case of the worst kind of seasickness (see Reason & Brand, 1975).
That which we mostly take for granted, postural stability, depends upon the complex interaction of the spatial senses—peripheral vision, the semicircular canals and otoliths (together comprising the vestibular system) and the muscle-skin-joint system—with the cerebellum and the spinal reflexes. The subtle orchestration of this multipart stabilising system only becomes apparent when it is abused by disease, by drugs (e.g., alcohol) or, most informatively, by a variety of laboratory manipulations. Among the latter, some potent effects can be achieved by presenting people with large-scale, moving visual scenes in which all the structural elements move in unison—the technique called Vection’.
Figure 6.2. The possible mechanisms governing automatic postural corrections. Inputs from the various spatial senses are checked for their degree of correspondence, one with another, within the comparator system. The output from this comparison process adjusts the position of an internalised body model relative to the primary spatial coordinates. Where discrepancies exist between the inputs from the spatial senses, those from the visual modality override the others. Only the output from the model (the mismatch signal) acts directly upon the corrective musculature.
Imagine the following situation: You are standing upright in front of a striped display that occupies the greater part of your visual field. The stripes now begin to move at a steady velocity in a downward direction while your eyes remain fixed on a target light in the centre of the display. After a few seconds, you will begin to lean forward quite involuntarily, only maintaining your balance (if at all) by fairly strenuous work on the part of your ankle and leg muscles. After about a minute, the visual motion is stopped, and the ‘force’ that caused your body to lean forward is removed. However, instead of drifting back to the upright position, you will now find yourself leaning backwards —to such an extent that you may be forced to hold on to something solid to avoid falling over. But, if instead of the visual motion being stopped, you simply close your eyes, then there is no tilt-back effect. Your body merely returns to the upright position and stays there.
This simple and largely irresistible lean-forward effect, together with its sequelae (see Reason, Wagner & Dewhurst, 1981
), tells us a great deal about the covert processes involved in maintaining balance. A diagrammatic representation of the possible interrelationships between these various postural mechanisms is shown in Figure 6.2.
To explain these effects, it is necessary to assume that each of the spatial senses feeds its inputs to a comparator that checks for the degree of correspondence between their descriptions of the body’s position and motion. The central feature of this theory, however, is that the output from this comparison process is not passed directly to the corrective musculature; instead, it acts to adjust the orientation of some internalised representation of the body’s position in space. Thus, only the position of this ‘body model’ can direct the corrective musculature. In the example given above, the model was ‘fooled’ into thinking that the body was tilting backwards. As a result, it instituted postural adjuments designed to return the body to the upright position. But since the body was upright to begin with, their effect was to tilt it forwards. How could this happen?
Under natural conditions, large-scale movements of the visual field are an invariable accompaniment of self-locomotion. A uniform downward motion of the entire visual scene (relative to some fixed point on the retina) is something that normally goes along with a backward tilt. Vision is the dominant sense. Even when the other senses (as was the case in the tilt effect) are signalling veridical information, peripheral vision, although wrong, has the power to override them and to modify—in this case erroneously—the position of the ‘body model’ so that it adopts a tilt-back orientation. This should not be interpreted as a failure of the postural correction system. Like any biological mechanism, it is precisely attuned to a particular set of environmental conditions. And these do not include the psychological laboratory.
Why does the body overshoot the upright position when the motion is switched off? Having been driven forward by the automatic correction mechanisms, the muscle-skin-joint system has to work very hard to prevent the body from falling flat on its face. These inputs, via the comparator, to the ‘body model’ serve to counteract some of the effects of the visual motion. In effect, the body model reorients itself near to the apparent vertical, even though it is actually tilted forward. When the visual motion is stopped, the lean-forward force is suddenly removed and the effect of this is to ‘flip’ the body model to an apparent upright that lies somewhere behind the true vertical. The body then aligns itself with this apparent upright position rather than with the true vertical, putting its physical reality (as distinct from its cerebellar representation) into a tilt-back position. It does not do this when the eyes are shut during the motion because that signals ‘no vision’ rather than ‘no motion’. Since the eyes are capable of shutting, the comparator is designed to accommodate the simple absence of vision.
The simplest way to understand this process is to imagine an elastic band, knotted in the middle and held at either end. The position of the knot represents the forward or backward tilt of the body model. When the visual motion begins, one end of the band is pulled forward, shifting the knot towards the pulling force. The compensating proprioceptive inputs then start to ‘pull’ at the other end, causing the knot to move back towards the central position. Now the knot is held steady by tension at either end of the band. When the visual motion stops, the force from the forward end of the band is suddenly removed, and the knot now shifts to the backward position. As the body drifts back to the upright, the proprioceptive counterforces are gradually released, but not fast enough to prevent it from adopting a backwards lean in accord with the knot (the body model). Eventually, the knot stabilises in the centre of the rubber band with little or no tension at either end. At that point, the physical body is restored to the true upright. In the case where the eyes are shut, one can think of the forces at either end of the elastic band being reduced at an approximately equal rate.
This is not the place to discuss these mechanisms further. The interested reader is directed to the relevant literature: Dichgans and Brandt (1978), and Reason, Wagner and Dewhurst (1981). The point of this discussion has been to reveal the exquisite machinery of the postural correction mechanisms, even though, paradoxically, we have had to demonstrate this by their failures in a nonecological situation. When intact under normal conditions, they function as a near-perfect piece of biological engineering without any recourse to cognitive processes. Indeed, they are only rendered fallible when forced by these high-level control agencies to operate within atypical force environments, such as are found in almost all kinds of human transport.
2.2. The detection and correction of simple motor responses and perceptual discriminations
From the mid-1960s onwards, Patrick Rabbitt and his collaborators at the University of Oxford conducted a series of studies designed, in the first instance, to identify the mechanisms by which errors in simple choice-response tasks were so rapidly and accurately corrected (Rabbitt, 1966, 1967, 1968; Rabbitt & Phillips, 1967; Rabbitt & Vyas, 1970). The investigation began with the observation that error-correcting responses tend to be faster than correct responses (Rabbitt, 1966), or even correct response repetitions (Bums, 1965), in tasks where people were required to press as quickly as possible one of a series of keys at the onset of one of a matching number of signal lamps.
It was found that people could catch and recover nearly all of their errors in these simple keyboard tasks, even though the apparatus itself gave them no indication of whether their responses were correct or not. The efficiency of this detection and correction process was unaffected by the degree to which the signals and the required responses were compatible. Moreover, error corrections occurred, on average, much more quickly than responses to other signals within the task.
The most parsimonious explanation for these results was that people detect their execution errors by comparing what they felt or saw of the wrong response with a record of what they had intended. Expressed in control theory terms, one could say that these execution errors were detected by monitoring the feedback from each response and checking it against some internal model of the correct response. Evidence for the existence of such a response ‘template’ comes from physiological recordings showing the existence of anticipatory muscle tension some 100 milliseconds before any overt limb movement occurs (Megaw, 1972). The fact that movement errors are detected so rapidly after their execution suggests that something like an ‘echo’ of the correct motor program persists after the response has been made, allowing comparison between the actual and intended responses.
Although it is clear from these data that most errors of this kind are caught and corrected by feedback-checking, there are good reasons for doubting that this is the only possible detection mechanism. It is quite possible, for example, that some motor errors in choice-reaction tasks arise not because of the misselection of the response, but because of a misperception of the signal. If a person mistook signal A for signal B, and made a response appropriate to B, there is no way that monitoring the feedback could reveal the error. Discrimination errors of this kind would require a different detection mechanism.
A further set of studies were conducted by this Oxford group to determine whether such perceptual errors can be corrected, and, if so, to identify the means of detection and correction (Rabbitt & Vyas, 1977; Rabbitt, Cumming & Vyas, 1978). The problem with the choice-reaction task, however, is that it is virtually impossible to discover retrospectively whether a particular error was due to a failure in response execution or to a failure in perceptual discrimination. To overcome this difficulty, they employed a visual search task in which subjects inspected one display of letters at a time and were required to respond differently depending upon whether the target set of letters was present or absent in a given display. Two kinds of errors could then be identified, indicating that a target was present when it was not (false identification error) or failing to respond to a target that was actually present (omission error).
It was found that omission errors were much more common than false
identifications, but that many more of them were corrected. In addition, omissions were detected and corrected more rapidly than false identifications.
The general conclusion was that as well as correcting execution errors by feedback-checking, people can also correct some of the perceptual identification errors they make during visual search tasks. The differences in the relative ease of detection of omissions and false identifications could best be explained by assuming that the perceptual analysis of a display persisted for some time after an incorrect response had been made to it.
2.3. The detection and correction of speech errors
Speech, which is made up of readily identifiable linguistic units, allows for a more fine-grained analysis of error detection and correction than do other less easily segmented skill-based activities. Most psycholinguists (see Fromkin, 1980) now accept the existence of an internal monitor or ‘editor’ that checks speech outputs both before and immediately after their utterance. These theorists differ, however, in the degree of detail with which they specify the workings of this error-detection mechanism (see Nooteboom, 1980; Laver, 1980; Garrett, 1980).
Nooteboom (1980) analysed all the errors in Meringer’s 1908 corpus (see Chapter 2) that were discovered and corrected by the speakers. Sixty-four per cent of all the errors in the corpus were corrected by their perpetrators. For both lexical and phonological errors, anticipations were detected more often than perseveration or transposition slips. The detection rate for anticipation slips was between 80 per cent and 90 per cent; for perseverations it was between 55 per cent and 66 per cent; while for transpositions it was between 14 per cent and l8 per cent. This low figure for transposition detections is readily accounted for by the method of classification. If a transposition like ‘heft hemisphere’ is corrected after the first phonemic error and before the spoonerism is complete, it would be classified as an anticipation error. Classifying the first halves of transpositions as anticipation errors could well explain the latter’s relatively high detection rate within the corpus.