Experimental Slips and Human Error

Abstract

Whereas most humans spend their time trying to get things right, psychologists are perversely dedicated to error. Errors are extensively used to investigate perception, memory, and performance; some clinicians study errors like tea leaves for clues to unconscious motives; and this volume presents the work of researchers who, in an excess of perversity, actually cause people to make predictable errors in speech and action. Some reasons for this oddity are clear. Errors seem to stand at the nexus of many deep- psychological questions. The very concept of error presupposes a goal or criterion by comparison to which an error is an error; and goals bring in the foundation issues of control, motivation, and volition (Baars, 1987, 1988; Wiener, 1961). Errors serve to measure the quality of performance in learning, in expert knowledge, and in brain damage and other dysfunctional states; and by surprising us, they often call attention to phenomena we might otherwise take for granted. Errors also seem to reveal the "natural joints" in perception, language, memory, and problem solving-revealing units that may otherwise be invisible (e.g., MacKay, 1981; Miller, 1956; Newell & Simon, 1972; Treisman & Gelade, 1980). Thus errors show how the nervous system copes with failures of its own goals and constraints; they clearly reveal the limits of our immediate memory and conscious capacity, but they also show the price to be paid for the many unconscious automatisms that come to the fore when people are absentminded or distracted (Reason, 1982, 1983, 1984, 1992). In the world outside the laboratory, human error is thought to be a major contributor to industrial disasters, airplane crashes, medical malpractice, and automobile accidents. And even in the psychology clinic, dysfunctional symptoms appear as persistent, unwanted errors in the actions, thoughts, and perceptions of clinical patients (see Baars, 1988, and Chapter 1, this volume; also Spitzer, 1979). These points are especially applicable to inadvertent errors, or slips, which are generally defined as errors that violate their own governing intentions. Such errors are known to be errors by the person who commits them, but they are made nonetheless. Slips differ from errors made out of ignorance or vii viii Preface forgetfulness of a goal; they are especially revealing of the control problems we encounter in carryng out our voluntary intentions. SLIPS IN MODELS OF LANGUAGE It may seem obvious that an adequate theory of language should explain speech errors as well as fluent, correct speech. But this view is not universally held. For example, Chomsky's widely accepted goal for linguistic theory (1957, 1965) is to generate "all and only the grammatical strings" of a natural language like English. This definition excludes errors. It has taken some time and effort to persuade modem linguists that nongrammatical utterances are worth studying (Fromkin, 1968, 1973). Even today few proposals in the Chomsky tradition venture to account for errors (but see Fay, 1980; also Chen & Baars, Chapter 9, this volume). Likewise, some artificial-intelligence language simulations do quite a good job of analyzing and generating meaningful sentences, but they do not make humanlike errors (e.g., Winograd, 1972). Of course one could always add an error-generating mechanism to these models, but that would be artificial and unparsimonious. To simulate real human speech, a model should produce both correct and erroneous output in a natural way, without extraneous theoretical baggage. A third type of theory has begun to fill this gap. Parallel distributed processors (PDPs) treat the cognitive system as a society of specialized networks, in which activation can spread from node to node in a parallel-interactive fashion. Dell (1985, 1986) has developed a model of sentence production in which several levels of representation exist, including phonemic and lexical. Activation numbers are assigned to lexical and phonemic nodes, and the activation spreads to adjacent nodes both within and between levels. (In "true" PDP models, connections between the nodes are also assigned strength levels, a very important theoretical modification.) When simulations are run with a range of plausible parameters, it appears that this type of network generates correct lexical output as well as spoonerisms and displacements of single phonemes. The model demonstrates the same preference for lexical over nonsense spoonerisms that has been shown to exist in both induced and spontaneous slips (Baars, Motley, & MacKay, 1975; Dell, 1986). Along similar lines, MacKay (1982, 1987) has proposed a spreading-activation model of the serial order issue in speech and action. The general properties of PDPs have been worked out in considerable detail by Grossberg, Rumelhart, McClelland, and others, and the PDP approach has been used quite successfully to model other well-studied phenomena in perception, memory, language learning, and action control (Grossberg, 1982; Hinton & Anderson, 1981; Rumelhart, McClelland, & the PDP Group, 1986; see also Arbib, 1980; Baars, 1980a,b, 1983, 1987, 1988; Gel'fand, Gurfinkel, Fomin, & Tsetlin, 1971; Greene, 1972). Distributed models suggest a new way of thinking about human beings. We normally think of ourselves as guided by an executive "self': ''we'' have Preface ix control over our actions. But distributed systems are strongly decentralized. The intelligence of such systems resides in its independent components and in the interactions between them, so that the whole system behaves much like a market economy, in which thousands of individual transactions take place without centralized intervention, and fundamentals like supply, demand, and price emerge from the system as a whole. Details of processing are generally handled by specialized networks in the processing "society." A decentralized system does not rule out executive control, just as the existence of market forces does not rule out a role for government. But it limits the purview of executive systems and creates possibilities for a reciprocal flow of control between executive and subordinate elements. Such systems have some distinct virtues (e.g., Greene, 1972). Thus, there has been notable progress in understanding and modeling both correct and erroneous language. However, almost no current theory attempts to explain the involuntary character of slips, an issue that is explored in some detail in Chapters 1 and 4. Indeed, it is argued in Chapter 1 that ''local'' theories of speech and action are in principle unable to explain some of the core issues raised by slip phenomena, such as the clear psychological differences between voluntary and involuntary acts (viz. Baars, 1988, and Chapter 4, this volume). PRACTICAL ASPECTS OF SLIPS While the study of slips is contributing to scientific theory, it also has important practical implications. Inadvertent errors can apparently cause serious accidents. Reason (1984, 1992) has analyzed several catastrophic accidents in which fatalities were apparently caused by slips: a London bus driver who crashed a double-decker bus into a low overpass, killing 6 passengers (perhaps because he was in the habit of driving the same route in a singledecker bus); a train collision killing 90 people because the driver made a habitual turn, but this time onto the wrong track; and the notorious runway collision on the island of Tenerife in 1977, in which a KLM Boeing 747 taxied into the path of a landing jumbo jet, killing 577 passengers and crew. The very senior KLM pilot failed to obtain clearance from the control tower before takeoff, perhaps because he had just finished a different assignment involving hundreds of hours of training pilots in a cockpit simulator, in which runway clearance was not required. In all three cases, Reason suggests that the fatal errors may reflect a ''habit intrusion," that is, an inadvertent substitution of a highly habitual and automatic action for the correct one (see also Langer, 1989). A similar pattern of accidents may have occurred in flight training with the B-29 bomber in World War II. The B-29 was a four-engine aircraft designed to fly on only one engine in an emergency. As a routine part of flight instruction, three of the four engines were shut off in midflight, to prove to student pilots that the aircraft would indeed continue to fly on just a single engine. For x Preface this demonstration to work, the fourth engine had to be kept running, however, to generate the electrical power needed to restart the others. With all engines turned off, electrical power would be inadequate to restart any engine, and the aircraft would simply lose altitude and crash. Several training flights came to grief when pilots and instructors inadvertently turned the fourth engine off, rather than turning a second engine on. Norman (1976) points out that, in emergencies like this, panic may serve to degrade rather than improve effective action, by interfering with conscious/limited-capacity processing. SLIPS AND PSYCHOPATHOLOGY Perhaps the most unusual point I will make here is the close connection between slips and the symptoms of clinical psychopathology. Clinical symptoms may be viewed as recurrent, uncontrollable slips in speech, action, emotion, imagery, or thought (Baars, 1988, pp. 246-344). This observation exactly fits the operational definition of slips suggested in Chapter 1, as errors that occur even though they are quickly recognized to be errors when we pay attention to them. Thus, catastrophic mental images in phobias, compulsive behavior, tics, stutters, obsessive thoughts, or excessive and upsetting emotions are quickly recognized by their victims as unwanted and unintended. They are nevertheless difficult or impossible to stop. All of the "Axis I" dysfunctions in the current American Psychiatric Association Diagnostic and Statistical Manual (DSM-III-R; APA, 1987)-the so-called ego-dystonic s