Error monitoring under working memory load: an electrocortical investigation

Abstract

Error monitoring is essential for detecting errors, and may facilitate behavioral adjustments that can reduce or prevent future errors. At times, error monitoring must occur while individuals are engaged in other, cognitively demanding tasks that might consume processing resources necessary for error monitoring. Here, we set out to determine whether concurrent working memory load interferes with error monitoring, as measured using event-related potentials, the error-related negativity (Ne/ERN) and error positivity (Pe). Fifty-four participants (n = 33 female) completed an arrowhead flanker task, with trials presented under low (2 letter) or high (6 letter) working memory load. Participants were required to hold letter strings in memory and to recall these letters at the end of a set of flanker trials. Results showed that working memory load reduced the Pe but did not affect the Ne/ERN. Therefore, working memory load appeared to attenuate later, more elaborated stages of error processing, though initial error detection was unaffected. Additionally, high working memory load slowed reaction times overall, but did not lead to a significant increase in errors. As such, slower responses may have helped participants maintain comparable accuracy for low-load versus high-load trials. Overall, results indicate that working memory load interferes with the evaluation of error significance, which could interfere with behavioral adaptations over time.

1. Introduction

Error monitoring is an essential cognitive function that may help people optimize their behavior. Error monitoring is arguably just as important when people are performing demanding cognitive tasks as when they are performing easier tasks. As such, a crucial question is how error monitoring is affected by the consumption of central processing resources. If error monitoring depends on the availability of central processing resources, then error monitoring might be compromised when cognitive demands are high. Alternatively, depletion of central processing resources could result in increased error monitoring to maintain adequate behavioral performance. Understanding how working memory (WM) load affects event-related potentials (ERPs) associated with errors, such as the error-related negativity (Ne/ERN) and error positivity (Pe), could provide insight into how cognitive resource constraints influence error processing.

The Ne/ERN is a negative-going, frontocentral ERP that is maximal 0–100 ms after errors and is larger (more negative) in magnitude than the ERP that occurs after correct responses, the correct-related negativity/CRN (Ford, 1999). In addition to being thought of as measure of error monitoring (Falkenstein et al., 1991; Gehring et al., 1993), it has also been suggested that the Ne/ERN might reflect a relatively automatic process, which might not depend on the availability of domain-general processing resources (Scheffers & Coles, 2000). The Pe is a positive-going, centroparietal ERP component that peaks approximately 200–400 ms following response and is also larger for error versus correct responses (Falkenstein et al., 2000). The Pe has been associated with the processing of error significance, and might therefore be more subject to the availability of central cognitive resources than the Ne/ERN (Overbeek et al., 2005).

WM tasks consume central processing resources and can therefore reduce the availability of these resources for other tasks and stimuli. For example, WM load reduces the electrocortical processing of emotional stimuli (Barley et al., 2021; MacNamara et al., 2011, 2019) and slows extinction learning (Cheng et al., 2022). Past research has also manipulated WM load to explore how consumption of domain general processing resources affects error monitoring. In one study, individuals with obsessive compulsive disorder (OCD) and a group of healthy controls performed an n-back WM task interspersed with a flanker task (Klawohn et al., 2016). Results showed that WM load reduced the Ne/ERN in both groups, though more so in individuals with OCD. Nonetheless, the control block in this study was always presented first, meaning that reductions in the magnitude of the Ne/ERN could have been due to fatigue or other order effects across the task. In another study, WM load was found to reduce the Ne/ERN for errors that were thought to be more motivationally salient, but not for errors that were less motivationally salient (Maier & Steinhauser, 2017). This led the authors to conclude that error detection is an automatic process, because WM load only seemed to reduce the significance of errors and not the overall error monitoring process.

A similar conclusion was drawn by LoTemplio and colleagues (2023), who set out to test the compensatory error monitoring hypothesis (Moser et al., 2013). This hypothesis purports that anxiety’s potentiating effect on the Ne/ERN is compensation for worry’s consumption of WM resources. To test this hypothesis, LoTemplio and colleagues (2023) subjected unselected participants to both a state anxiety manipulation and low and high WM load, as they performed a flanker task. Their results revealed that neither anxiety or WM load modulated the Ne/ERN. Nonetheless, because they had asked participants to hold letters in memory over long blocks of 50 flanker trials, they cautioned that their results may not have been attributable to WM load, because the letters could have been transferred to long-term memory at some point during the blocks of trials. Finally, in a study that was only briefly described in a review paper, Moser and colleagues (2013) found that manipulating WM load on a trial-by-trial basis increased the magnitude of the Ne/ERN. Therefore, across the four studies that have examined the effect of WM load on the Ne/ERN, one study found that WM load reduced the Ne/ERN (Klawohn et al., 2016), one study suggested that WM load might reduce the Ne/ERN elicited by some types of errors but not others (Maier & Steinhauser, 2017), one study found no effect of WM load on the Ne/ERN (LoTemplio et al., 2023) and one study found a potentiating effect of WM load on the Ne/ERN (Moser et al., 2013).

Though only two studies have reported on the effect of WM load on the Pe, these results are also contradictory. That is, one study found no significant effect of WM load on the Pe (Maier & Steinhauser, 2017), and another found that WM load reduced the Pe (LoTemplio et al., 2023). There are reasons to believe that elaborative error processing may be more capacity dependent and resource intensive than the early stages of error detection (Rabbitt, 2002), and that it might therefore be more sensitive to the effect of working memory load. Along these lines, inhibitory repetitive transcranial magnetic stimulation (TMS) of the left dorsolateral prefrontal cortex, which is implicated in WM (Barbey et al., 2013; Curtis & D’Esposito, 2003; Rypma et al., 2002), has been found to reduce the amplitude of the Pe, but not the Ne/ERN (Masina et al., 2019). WM load has also been found to reduce the P3, an ERP component that is sensitive to the motivational significance of stimuli (Watter et al., 2001; Wei & Zhou, 2020; Wintink et al., 2001). Similarities between the P3 and the Pe have been noted (Davies et al., 2001; Overbeek et al., 2005; Riesel et al., 2013). For example, like the Pe, the P3 is a centroparietally maximal, positive-going component, that emerges approximately 300 ms post-event/stimulus and is larger for more salient or unexpected stimuli (Donchin, 1981; Duncan-Johnson & Donchin, 1977; Sutton et al., 1965). Moreover, both components reflect the motivational significance of stimuli (Arbel & Donchin, 2009; Leuthold & Sommer, 1999; O’Connell et al., 2007; Ridderinkhof et al., 2009). In addition, the Pe and P3 have been found to covary across participants (Ridderinkhof et al., 2009) and WM load has been found to reduce the P3 elicited by a variety of stimuli, from the position of shapes to sequences of letters (Pratt et al., 2011; Watter et al., 2001; Wei & Zhou, 2020; Wintink et al., 2001). As such, it is plausible that like the P3, the Pe may also be reduced by WM load.

In the current study, we used an experimental design informed by prior work, which we hoped would permit a more conclusive understanding of whether WM load affects the Ne/ERN and the Pe. Sources of variability in prior work have included fixed versus random block order (e.g., Klawohn et al., 2016; LoTemplio et al., 2023), differences in block length (e.g., LoTemplio et al., 2023) and a variety of procedures for modulating WM load (e.g., Maier & Steinhauser, 2017; Moser et al., 2013). Furthermore, some studies included participants with clinical anxiety (Klawohn et al., 2016) or asked participants to worry (or to think positive thoughts) during the task (LoTemplio et al., 2023). In order to interrogate the basic science question as to whether error processing is automatic or depends on WM resources, we decided to intersperse a widely used WM load task (Barley et al., 2021; MacNamara et al., 2011, 2019) with a moderate number of flanker trials in each block (i.e., 15), and to employ an unselected sample that was free from mood manipulation. Low and high working load blocks of a working memory task were presented in a randomized order, interspersed with blocks of 15 flanker trials across the task. In accordance with prior TMS work (Masina et al., 2019) and work characterizing the functional significance of the Ne/ERN and Pe (Riesel et al., 2013), we hypothesized that WM load would reduce the Pe, but not the Ne/ERN. Based on behavioral results from similar studies (Klawohn et al., 2016; Maier & Steinhauser, 2017), we hypothesized that WM load would increase error rate and reaction time.

2. Method

2.1. Participants

Eighty undergraduate students provided written informed consent and participated in the study for course credit. Twenty-three participants committed fewer than six errors in at least one of the conditions (Olvet & Hajcak, 2009) and were therefore excluded from further analysis. Three additional participants were excluded due to technical issues during data collection, resulting in a final sample of 54 participants (33 female; age M = 19.70 years, SD = 1.45). Study procedures were in compliance with the Helsinki Declaration of 1975 (as revised in 1983) and were approved by the Texas A&M institutional review board.

2.2. Experimental Procedure and Task

Participants completed a modified version of an arrowhead flanker task (Eriksen & Eriksen, 1974), performed under high-load and low-load WM blocks (Figure 1). Trials were arranged into 22 blocks of 15 trials per block (11 low-load; 11 high-load), presented in random order. This led to a total of 330 trials, consisting of four trial types across the task: correct responses made under low-load (low-load correct); errors committed under low-load (low-load error); correct responses made under high-load (high-load correct) and errors committed under high-load (high-load error).

Figure 1.

The working memory flanker task. In each low-load and high-load block, participants viewed a string of 2 or 6 letters, prior to performing 15 flanker task trials, with congruent (not depicted here) and incongruent (depicted here) flankers. At the end of each block, participants were asked to enter the letters presented at the beginning of the block.

At the start of each block, a white fixation cross was presented on a black background for 500 to 1000 ms. Next, either a two-letter (low-load) or six-letter (high-load) string was displayed for 5000 ms. Letter strings were created using a random number generator (Reed, 2022). Vowels were not included in the strings; there were 60 two-consonant strings and 60 six-consonant strings (Ashcraft & Kirk, 2001). This was followed by a series of flanker trials where participants viewed a fixation cross for 200 ms, then five arrows presented for 200 ms against a black background; participants were asked to respond as quickly and as accurately as possible to indicate the direction of the middle arrow by pressing the left or right mouse button. Participants had up to 1800 ms from the onset of the flankers to respond. Half of the trials in each of the high-load and low-load blocks were congruent (< < < < < or > > > > >), whereas the other half in each block type were incongruent (< < > < < or > > < > >). Trial order was random for each participant and the intertrial interval varied from 200 to 1200 ms, during which time a white fixation cross was presented on a black background. After completing 15 trials of the flanker task, the text “What were the letters? (then press enter):” was presented in the center of the screen. Participants were instructed to use the keyboard to enter the letters they had viewed at the beginning of the block and to enter them in the same order that they had originally been presented. At the end of each block of trials, participants received feedback designed to counteract highly conservative responding. If their flanker accuracy was below 75%, participants received the feedback, “Please try to be more accurate when responding to the arrows”; if flanker accuracy was greater than 90%, they received the feedback, “Please try to respond faster to the arrows”; all other flanker accuracy scores received the feedback, “You’re doing a great job responding to the arrows.”

Before beginning the task, participants completed a practice block containing 10 trials with only the flanker task and two trials with the WM and flanker task together. Participants received self-paced breaks in between each block. Stimuli were presented using Presentation software (Neurobehavioral Systems Inc., Albany, CA).

2.3. EEG Recording and Data Reduction

Continuous EEG was recorded using an ActiCap and the ActiChamp amplifier system (Brain Products, Gilching, Germany). Thirty-two electrode sites were used based on the 10/20 system. The electrooculogram (EOG) was recorded from four facial electrodes: two electrodes were placed approximately 1 cm above and below the right eye, forming a bipolar channel to measure vertical eye movement and blinks and two electrodes were placed approximately 1 cm beyond the outer edges of each eye, forming a bipolar channel to measure horizontal eye movements. EEG data was digitized at a 24-bit resolution with a sampling rate of 1000 Hz.

Offline, data were processed using BrainVision Analyzer 2 software (Brain Products GmbH, Gilching, Germany). The signal from each electrode was referenced offline to the average of the left and right mastoids (TP9/10) and band-pass filtered with high-pass and low-pass filters of 0.1 and 30 Hz, respectively. For the ERN/CRN and Pe/Pc, data were segmented for each trial beginning 500 ms prior to response and lasting for 1500 ms (1000 ms beyond response). Eye blink and ocular artifacts were corrected using the method developed by Miller, Gratton and Yee (1988). Artifact analysis was used to identify a voltage step of more than 50.0 μV between sample points, a voltage difference of 300.0 μV within a trial, and a maximum voltage difference of less than 0.50 μV within 100 ms intervals. Trials were also inspected visually for any remaining artifacts, and data from individual channels containing artifacts were rejected on a trial-to-trial basis.

For response-locked ERPs, correct and error trials were averaged separately for low-load and high-load trials, and baseline correction was performed using a 200 ms window from 500 to 300 ms before response. We used a functional localizer approach to score both the ERN/CRN and Pe/Pc, in which waveforms were inspected for the largest error versus correct response across conditions to determine time windows and scalp distributions for each component (Luck, 2014). The Ne/ERN and its counterpart on correct trials, the correct-related negativity/CRN, were scored at a pooling of Fz and Cz as the average activity on error and correct trials, respectively, from 0 to 100 ms after response (Klawohn et al., 2020). The Pe and its counterpart on correct trials, the Pc, were scored at Pz as the average activity on error and correct trials, respectively, from 200 to 400 ms after response (Riesel et al., 2012).

2.4. Flanker Task Behavior

Flanker error rate was calculated separately for each participant as the percentage of incorrect responses in each condition (i.e., low-load correct, low-load error, high-load correct, high-load error). Reaction time was recorded as time in ms between flanker stimuli onset and participant response and was calculated separately for each condition.

2.5. Statistical Analyses

The ERN/CRN, Pe/Pc, and reaction time were analyzed using a 2 (load: low, high) X 2 (response: correct, error) within-subjects repeated measures analysis of variance (ANOVA). Significant interactions were followed up using t-tests as appropriate. Error rate was analyzed using separate paired-samples t-tests, comparing low-load versus high-load trials. WM performance, calculated separately for each participant as the percentage of correct responses in low-load and high-load trials, was assessed to ensure the task was difficult enough/as a manipulation check. All analyses were performed using SPSS statistical software version 28.0 (IBM, Armonk, NY). Additional analyses examining single-trial ERPs with multilevel models (MLMs) and analyses with incongruent trials only are detailed in the supplementary information section and described briefly, below.

3. Results

All data are open and available on the Open Science Framework (https://osf.io/2pgzq/). Table 1 presents task behavior (flanker task and WM), ERN/CRN amplitudes, and Pe/Pc amplitudes, shown separately for low-load and high-load conditions. As expected, participants accuracy was lower on high-load (M = 72.22%, SD = 18.56) than low-load (M = 96.30%, SD = 6.96) WM trials, t(53) = 9.38, p < .001, d = 1.28.

Table 1.

Means [95% CIs] for ERPs, flanker error rate, flanker reaction time (RT), and working memory accuracy.

	Low-Load	High-Load
ERN (μV)	1.42 [−0.42, 3.26]	1.24 [−0.86, 3.33]
CRN (μV)	6.09 [4.61, 7.56]	5.20 [3.79, 6.61]
Pe (μV)	10.65 [8.95, 12.35]	8.39 [6.41, 10.38]
Pc (μV)	1.93 [0.83, 3.04]	1.30 [0.27, 2.33]
Flanker error rate (%)	12.31 [10.38, 14.24]	13.17 [11.00, 15.33]
Flanker error RT (ms)	316.33 [297.19, 335.47]	337.04 [317.39, 356.69]
Flanker correct RT (ms)	379.55 [366.58, 392.53]	389.37 [375.73, 403.01]
WM accuracy (%)	96.30 [94.40, 98.20]	72.22 [67.16, 77.29]

3.1. ERN/CRN

Figure 2A depicts grand-averaged waveforms for error and correct responses at Fz/Cz where the Ne/ERN and CRN were scored and Figure 2B shows headmaps depicting voltage difference distributions for error minus correct responses, shown separately for low-load and high-load trials. As expected, the Ne/ERN was more negative than the CRN, F(1, 53) = 45.08, p < .001, ƞ_p² = .46. The main effect of load and the interaction between load X response did not reach significance, all ps > .166.

Figure 2.

The ERN/CRN. A) Grand-averaged waveforms for error and correct responses at a pooling of Fz and Cz where the ERN/CRN were scored, shown separately for low-load and high-load trials. B) Headmaps depicting the voltage difference distribution for error minus correct responses, shown separately for low-load and high-load trials.

3.2. Pe/Pc

Figure 3A depicts grand-averaged waveforms for error and correct responses at Pz where the Pe and Pc were scored and Figure 3B shows headmaps depicting the voltage difference distribution for error minus correct responses, shown separately for the low-load and high-load trials. As expected, the Pe was significantly more positive than the Pc, F(1, 53) = 178.05, p < .001, ƞ_p² = .77. A main effect of load, F(1, 53) = 15.71, p < .001, ƞ_p² = .23, was modulated by an interaction between load X response, F(1, 53) = 4.96, p = .030, ƞ_p² = .09, which indicated that the difference between the Pe and Pc was smaller on high compared to low WM load trials.¹

Figure 3.

The Pe/Pc. A) Grand-averaged waveforms for error and correct responses at Pz where the Pe/Pc were scored, shown separately for low-load and high-load trials. B) Headmaps depicting the voltage difference distribution for error minus correct responses, shown separately for low-load and high-load trials.

3.3. Bayesian Analyses of the ERN/CRN and Pe/Pc

To further evaluate the probability of the observed effects and interactions, we conducted Bayesian repeated measures ANOVAs for the ERN/CRN and Pe/Pc. Bayes Factors were computed in JASP (jasp-stats.org) using the default priors. Bayes Factors less than 1 indicate more support for the null than the alternative hypothesis, and Bayes Factors less than 1/3 suggest moderate support for the null hypothesis (van Doorn et al., 2021).

From the analysis of the ERN/CRN, we observed moderate evidence against including the load X response interaction, BF_inclusion = 0.348, and main effect of load, BF_inclusion = 0.321, in the model, whereas we observed strong evidence for including the main effect of response, BF_inclusion = 7.28 × 10⁵. Therefore, Bayesian statistics supported the notion that WM load did not affect the magnitude of the Ne/ERN.

For the analysis of the Pe/Pc, we observed strong evidence for including the main effects of load, BF_inclusion = 73.60, and response, BF_inclusion = ∞, as well as the load X response interaction, BF_inclusion = 10.27. Therefore, Bayesian statistics supported the notion that WM load reduced the Pe.

3.4. Error rate

Error rates did not differ for low-load (M = 12.30%, SD = 7.10) compared to high-load (M = 13.20%, SD = 7.90) blocks, p = .110.

3.5. Reaction time

As was expected, reaction times were slower for correct (M = 384.46, SD = 47.78) versus error (M = 326.69, SD = 64.07) trials, F(1, 53) = 118.78, p < .001, ƞ_p² = .69. In addition, participants were slower to respond to flankers presented on high-load (M = 363.21, SD = 56.45) compared to low-load (M = 347.94, SD = 55.75) trials, F(1, 53) = 9.46, p = .003, ƞ_p² = .15. However, there was no significant interaction between load X response, F(1, 53) = 2.14, p = .150, ƞ_p² = .04.

3.6. Bayesian Analyses of Error rate and Reaction time

Analyzing the effect of load on error rate using a Bayesian paired samples t-test, we observed little evidence in favor of the null hypothesis, BF₁₀ = 0.507 (i.e., the data was 0.507 times more likely to occur under the null hypothesis).

Analyzing reaction time using a Bayesian repeated measures ANOVA, we observed moderate to strong evidence for including the main effects of load, BF_inclusion = 9.42, and response, BF_inclusion = 2.12 × 10¹², but only weak evidence for including the load X response interaction, BF_inclusion = 1.92. Therefore, Bayesian statistics for error rate and reaction time were in line with frequentist analyses.

3.7. Additional Analyses

The supplementary information section includes multilevel model analyses that evaluate the effect of WM load (low, high) and response (correct, error) for single-trial ERPs while controlling for RT. By including RT as a covariate, these models controlled for the increased potential for stimulus-locked potentials to overlap with the ERN/CRN and Pe/Pc on certain trial types. In other words, controlling for RT should help account for the possibility that if participants were faster to respond to the flankers on certain trials (e.g., low-load trials), stimulus-locked ERPs (like the P300) may have been more or less likely to overlap with subsequently elicited response-related ERPs.

Additionally, the supplementary information section includes both repeated measures ANOVA and MLM analyses for only incongruent trials. The incongruent ANOVA showed a load X response interaction for the ERN/CRN, such that the CRN was larger for high versus low load trials, but not the ERN. However, this interaction was not significant in the MLM analysis. Moreover, both the incongruent ANOVA and MLM analyses failed to show a load X response interaction for the Pe/Pc, but still showed an overall decrease in the Pe/Pc on high versus low load trials (i.e., main effect of load). Trial counts, mean ERP amplitudes, grand-averaged waveforms with headmaps, congruency analyses in a subsample of participants with sufficient trials and difference waveforms can also be found in the supplement.

4. Discussion

Error monitoring is potentially just as important when individuals must divert attention and processing resources away from the task at hand as when they can devote all their attentional resources to a central task. Here, we assessed the effect of WM load on early and late stages of error monitoring, as measured using the Ne/ERN and the Pe. Results showed that WM load did not affect the Ne/ERN, but reduced the Pe. Therefore, error monitoring might be a relatively automatic process that is robust to the consumption of domain-general processing resources. On the other hand, more elaborative error processing, as measured by the Pe, appears to rely more on the availability of these resources. Additionally, WM load did not increase flanker task error rate, but instead increased overall response times, suggesting a compensatory process.

Errors can result in harmful consequences and may even be fatal. Therefore, from an evolutionary perspective, there should be an advantage to the rapid and automatic detection of errors. In support of this, evidence suggests that error detection and behavioral corrections can occur outside of conscious awareness. For example, motor corrections following errors on a pointing task with fast moving targets have been shown to occur automatically, even when participants are told to inhibit these corrections (Pisella et al., 2000). Similarly, rapid behavioral corrections following errors can occur before participants consciously recognize that they have made an error (Rabbitt, 2002). These results corroborate those observed here, to suggest that at least to some degree, aspects of error detection may proceed without the need for conscious executive control.

In contrast to results observed here, some prior work had found that WM load either increased (Moser et al., 2013) or decreased (Klawohn et al., 2016) the Ne/ERN. Nonetheless, a closer look at Klawohn and colleagues (2016) suggests some potential consistency with the current results. That is, the Ne/ERN was not reduced on 2-back versus 1-back WM load trials. Instead, the Ne/ERN was only reduced for WM load trials compared to trials without any WM load. Therefore, similar to the current study, null effects were observed when comparing high versus low WM load. While the Klawohn and colleagues (2016) results could additionally have been affected by a fixed task order (i.e., standard task, followed by 1-back and then 2-back), the current results are not subject to this potential confound, and therefore go farther to suggest that the Ne/ERN is not affected by WM load. Nonetheless, inclusion of a control condition without WM load would have further solidified this interpretation.

The current results most resemble findings from LoTemplio and colleagues (2023), who also found that WM load did not modulate the Ne/ERN but that it did reduce the Pe. Given that we did not employ a mood manipulation and also used shorter blocks of flanker trials (i.e., 15 instead of 50), our results help rule out other potential explanations, such as the possibility that letters were transferred to long-term memory/findings were not due to WM load. Additionally, Bayesian statistics employed in the current study increase confidence in the absence of an effect of WM load on the Ne/ERN. Therefore, based on results from LoTemplio and colleagues (2023) and those observed here, it seems likely that while WM load does not interfere with error monitoring, it does reduce error salience, or the evaluation of error significance (as measured by the Pe). Along these lines, prior work had found that individuals with lower WM capacity, which could mirror capacity limitations induced under high WM load, have smaller Pe’s (Coleman et al., 2018). Similarly, children and adults with attention deficit hyperactivity disorder (ADHD), which may be characterized by reduced WM capacity (Kofler et al., 2010) also have smaller Pe’s (compared to those without ADHD), but not smaller Ne/ERNs (Wiersema et al., 2005, 2009). Therefore, reduced availability of WM resources – whether due to individual differences or task design – appears to compromise the elaborated processing of error salience.

Compared to the Ne/ERN, the Pe appears to be especially involved in error awareness and the elaborated processing of error significance (Nieuwenhuis et al., 2001). The Pe may also be involved in the updating of cognitive strategies for remedial action, which could rely on WM resources (Coleman et al., 2018; Falkenstein et al., 2000; A. E. Miller et al., 2012). Along these lines, while inhibitory TMS of the dlPFC, which is recruited during WM tasks, has been shown to reduce the Pe (Masina et al., 2019), excitatory tDCS of the dlPFC has been found to enhance error awareness, as measured in a Go/No-go response inhibition task in which participants pressed a button after recognizing they had made an error (Harty et al., 2014). Therefore, though evidence has suggested that the Pe stems from the dACC and rostral cingulate zone rather than the dlPFC (Hester et al., 2005; Klein et al., 2007), the dlPFC might nonetheless indirectly affect the magnitude of the Pe via its involvement in the evaluation and elaborated processing of error significance. In particular, the dlPFC may be less available to evaluate error significance when occupied with other tasks, such as holding a string of letters in memory.

Similarities between the Pe and ERP component, the P3, inspired our hypotheses and may also provide context for the current results. For instance, context updating theory proposes that the P3 plays a role in updating WM and in revising models of the environment. In turn, WM updating is thought to lay the groundwork for more effective reactions to future encounters with stimuli (Donchin, 1981; Donchin & Coles, 1988). Therefore, like the P3, the Pe might use WM resources to update models of the environment. In this case, reductions in the Pe induced by another, ongoing task could indicate interference with WM updating on a central task (i.e., here, the flanker task), that could negatively affect behavioral adjustments to reduce future errors.

Interestingly, however, WM load did not increase flanker error rate in the current study. Instead, participants responded more slowly on high-load blocks (Klawohn et al., 2016; Maier & Steinhauser, 2017), suggesting that they may have been compensating for reduced availability of processing resources on high-load trials. That is, like LoTemplio and colleagues (2023), who found the same behavioral results with longer blocks of trials and a mood manipulation, our results suggest that WM load may lead to a shift in strategy rather than a deterioration in performance. Though we did not observe an effect of WM load on error rates in the current study, it is possible that this would emerge with modified task designs that could be examined in future work. For instance, differences in flanker accuracy under high versus low WM load might be observed if faster response times were emphasized and/or if the flanker task was more difficult.

The current study does come with some qualifications. For instance, additional analyses examining only incongruent trials showed somewhat different findings than reported in the main manuscript. That is, the interaction between load X response trended toward significance for the ERN/CRN. This effect was driven by larger CRNs on high-load compared to low-load trials (WM load did not affect the ERN). Prior literature suggests that the CRN may be an indicator of trial-to-trial task engagement and strategy adaption (Bartholow et al., 2005; Hajcak et al., 2005; Ridderinkhof et al., 2003). Therefore, on incongruent trials, larger CRNs on high-load trials could reflect increased effort needed to maintain adequate task performance. That is, increased effort to maintain correct responses may have been especially necessary on incongruent trials, which might explain why we did not observe an effect of WM load on the CRN when collapsing across congruent and incongruent trials.

For incongruent trials, we also failed to observe a load X response interaction for the Pe/Pc, while still finding that WM load decreased overall amplitudes for the Pe and Pc. An overlapping stimulus-locked P300, which has been shown to be reduced under higher WM load (Pratt et al., 2011; Wei & Zhou, 2020), could potentially explain this global blunting. Altogether, these additional analyses indicate that flanker congruency may affect findings in ERN/CRN and Pe/Pc studies, and could potentially account for mixed findings in prior work that has investigated the automaticity of the ERN.

Limitations of the current study include a sample comprised of college students with a mean age of around 19 years old. Therefore, it is unclear whether results would generalize to older individuals, some of whom might have reduced WM capacity. As such, future work may aim to include a larger and more diverse sample. Additionally, stronger conclusions regarding the role of WM load in reducing the Pe could be drawn if we had provided evidence of an association between a direct measure of WM load and the Pe (e.g., slow-wave potential; Maier & Steinhauser, 2017; Moser et al., 2013).

Taken together, our results suggest that on the whole, early error monitoring is a robust process that is uncompromised by an ongoing, cognitively demanding task. On the other hand, WM load may interfere with the later, more elaborative stages of error processing, potentially compromising the evaluation of error significance and WM updating following errors.