A Neural System Eor Error Detection and Compensation

Humans can monitor actions and compensate for errors. Analysis of the human event-related brain potentials (ERPs) accompanying errors provides evidence for a neural process whose activity is specifically associated with monitoring and compensating for erroneous behavior. This error-related activity is enhanced when subjects strive for accurate performance but is diminished when response speed is emphasized at the expense of accuracy. The activity is also related to attempts to compensate for the erroneous behavior. A fundamental characteristic of human cognition is its fallibility. People rarely perform tasks perfectly, even though the costs of imperfection can be devastating (Norman, 1988; Reason, 1990). It is plausible to assume that the prevalence of errors, and their high cost, has led to the evolution of mechanisms that monitor the accuracy of actions and attempt to correct, or compensate for, errors. That such mechanisms exist is, indeed, assumed explicitly or implicitly in many theories of cognition. For example, concepts of error monitoring are included in theories of action (MacKay, 1987), learning (Adams, 1971; Rumelhart, Hinton, & Williams, 1986), speaking (Levelt, 1989), and consciousness (Kosslyn & Koenig, 1992). Monitoring mechanisms are also implied by theories of executive or supervisory cognitive control systems (Logan, 1985; Shallice, 1988; Stuss & Benson, 1986). Given the frequency with which the concept of error monitoring is invoked, it is remarkable that there is little direct neurophysioiogical evidence for the existence of error-detection and -compensation systems (but see Gemba, Sasaki, Address correspondence to William J. Gehring, who is now at the Center for Neuroscience. University of California, Davis, Davis, CA 95616. & Brooks, 1986. for an exception). Much of the work investigating the issue has inferred the existence of a monitoring and compensation apparatus from behavior that appears to be compensatory, as when subjects execute an error and then quickly execute the correct response (Rabbitt, 1966, 1968) or slow down subsequent to errors (Laming, 1968; Rabbitt, 1966). These phenomena, while consistent with the existence of an error-monitoring system, are not conclusive, however, because they could occur without the presence of an errordetection system: The apparent correction could simply be a correct response produced in parallel with, but more slowly than, the error. Furthermore, a response on a trial after an error could be slow because of a persistence of the processing problem that caused the error. More direct evidence for an errormonitoring mechanism comes from descriptions of an event-related brain process that appears to be evoked contemporaneously with the commission of erroneous responses. We (Gehring, Coles, Meyer, & Donchin, 1990) have reported that an error-related negativity (ERN) appears selectively on error trials in choice reaction time experiments. The ERN takes the form of a sharp, negativegoing deflection of up to 10 L̂V in amplitude and is largest at electrodes placed over the front and middle of the scalp. Its onset is shortly after the onset of electromyographic (EMG) activity detected in the limb that is about to make an error, and it peaks about 100 ms following its onset. A similar observation was made, independently, by Falkenstein, Hohnsbein, Hoormann, and Blanke (1990). In this report, we present evidence that the ERN is a manifestation of the activity of a system associated with monitoring the accuracy of the response system and with compensating for errors. Our test is predicated on the assumption that if the ERN manifests the activity of such a system, it will be more active when response accuracy is important to the subject. We predicted that the amplitude of the ERN will vary with the relative weight the subject's task assigns to accuracy and speed. Furthermore, if the ERN is a manifestation of an errorcompensation mechanism, there ought to be a relationship between its amplitude and the dynamics of the erroneous responses. We varied, therefore, the speed and accuracy requirements placed upon the subject, and we measured several performance parameters that may reflect compensatory activity, including the force with which the subject executes a response, the probability of correcting the error, and the speed of responses following the error. We embedded these manipulations and measures in a task known from previous research to produce erroneous response activation (see Coles, Gratton, Bashore, Eriksen, & Donchin, 1985; Gratton, Coles, Sirevaag, Eriksen, & Donchin, 1988).