Abstract
Voluntary actions are typically preceded by the readiness potential (RP), a negative midfrontal electroencephalography (EEG) deflection that begins ~2 s before movement. What cognitive and neural process the RP reflects and how it relates to conscious intention remain unclear due to conflicting findings. We investigated the neural basis and cognitive significance of the RP in a novel probe-based paradigm. Contrary to prior reports, we found that pre-probe RP buildups were not related to reported awareness of motor preparation. Computational modeling suggested that the best explanation for these results is via metacognitive access to stochastic accumulation. Reported preparation was also related to beta desynchronization over contralateral motor cortex shortly before probe onset. We conclude that the RP may be metacognitively accessible in response to external task demands but does not reflect the onset of a conscious intention. We discuss implications of these findings for voluntary action initiation and intention awareness.
1 Introduction
Voluntary actions are typically preceded by a slow buildup of aggregate neural activity in midfrontal regions (Haggard & Eimer, 1999; Kornhuber & Deecke, 1965; Libet et al., 1983; Sakata et al., 2017; Soon et al., 2008). Single-unit activity in the supplementary motor area (SMA) and anterior cingulate cortex that ramped up or down in advance of movement onset was also demonstrated before voluntary action (Fried et al., 2011). The most well studied of these slow-ramping signals is the readiness potential (RP), an event-related potential that consists of a negative deflection in electroencephalography (EEG) that begins around 2 s before self-initiated movements (Kornhuber & Deecke, 1965; Shibasaki & Hallett, 2006). Despite the decades that have passed since the RP’s discovery, the kind of process that this signal reflects, its cognitive significance, and its role in action initiation remain contested.
Traditionally, the onset of the RP was thought to reflect a discrete neuro-cognitive event such as the beginning of motor preparation (Kornhuber & Deecke, 1965), or the onset of an unconscious decision to move (Libet, 1985; Libet et al., 1983; Shibasaki & Hallett, 2006). In contrast, more recent work suggested that the RP may instead reflect an integration-to-bound process of autocorrelated noise, which then resembles a gradual deflection when aligned to threshold crossing and back averaged (Khalighinejad et al., 2018; Maoz et al., 2019; Moutard et al., 2015; Schmidt et al., 2016; Schurger, 2018; Schurger et al., 2012; see Schurger et al., 2021 for a review). The so-called stochastic accumulator model thus parsimoniously accounts for the early rise and for the shape of the RP as well as for the distribution of wait times in unconstrained self-initiated action. It has further made novel predictions that were empirically confirmed (Schurger et al., 2012), and we recently demonstrated that a spiking neural network extension of stochastic accumulation can parsimoniously explain pre-movement activity at the single-neuro and network levels (Gavenas et al., 2024). However, Bogler et al. (2023) suggested that evidence specific for stochastic accumulation is still lacking, because other models can account for many of the same phenomena as stochastic accumulation. For instance, they note that linear ballistic accumulation—a process in which a monotonic linear buildup with a random slope triggers movement upon crossing a threshold (Brown & Heathcote, 2008)—can also account for the RP’s shape and waiting-time distributions. Unlike the traditional model of the RP, the onset of ramping in stochastic accumulation and linear ballistic-accumulation models does not reflect a specific event that initiates the ramping. The three models entail different implications for the neural bases of spontaneous voluntary action, the implementation of conscious control, and the utility of the RP and similar signals for movement prediction such as for brain–computer interfaces. Distinguishing between them is, therefore, of paramount importance.
Efforts to establish what kind of process underlies the RP are further complicated by conflicting results concerning the RP’s cognitive significance (Dominik et al., 2024). Studies of self-initiated action often have participants report the onset of the conscious urge or intention to move (referred to as W-Time) using the position of a rotating clock (Dominik et al., 2017; Haggard & Eimer, 1999; Libet et al., 1983). Though several studies have found that the onset of the RP does not correlate with W-Time as reported using the clock method (Bredikhin et al., 2023; Haggard & Eimer, 1999; Schlegel et al., 2013), recent studies employing an alternative method of assessing participants’ subjective state found that the reported awareness of conscious intention may indeed be related to RP onset (Parés-Pujolràs et al., 2019; Schultze-Kraft et al., 2020; though see Parés-Pujolràs et al., 2023 for a contrary report). For instance, Parés-Pujolràs et al. (2019) utilized an online-reporting paradigm in which participants were randomly probed before they moved (Matsuhashi & Hallett, 2008; Verbaarschot et al., 2016) and asked whether they had already begun preparing to move. Using this method, Parés-Pujolràs et al. (2019) found RP-like deflections specifically before probes when the participants reported having begun preparing to move. Those deflections were absent when the participants did not report preparing to move. Similarly, Schultze-Kraft et al. (2020) used an online decoding algorithm based on the RP to deliver probes when the RP was present or absent. They found that reported awareness of preparation was more likely when their decoder indicated an RP was present compared with absent.
These results suggest that the RP may specifically relate to the conscious intention to move. Considering recent criticisms of the clock method (Banks & Isham, 2009; Dominik et al., 2017; Lau et al., 2007; Maoz et al., 2015; Triggiani et al., 2023), it may be the case that probe methods, rather than clock methods, are the more appropriate avenue for investigating the onset of conscious intention. Nevertheless, it is unclear how these results influence the debate regarding what kind of process underlies the RP. Furthermore, probe methods inherently disrupt the movement-generation process, so they cannot determine when probes are delivered in relation to RP onset or to the onset of the movement that would have taken place had the probe not been delivered. It is, therefore, unclear whether the apparent link between the RP and reported awareness of intention applies to the early RP or only to its latest stages. Furthermore, decreases in beta power (desynchronization) in contralateral motor cortex also occur prior to self-initiated actions and appear to reflect motor preparation (Pfurtscheller & Aranibar, 1977; Pfurtscheller & Lopes da Silva, 1999; Shibasaki & Hallett, 2006). It was further claimed that beta desynchronization—rather than the RP—is related to awareness of motor preparation (Parés-Pujolràs et al., 2023), although evidence supporting that hypothesis remains tentative.
Here we developed a novel probe paradigm to investigate (1) whether the RP actually reflects the onset of conscious intention and (2) what neuronal processes the RP reflects. Participants were probed while they made self-paced movements and were instructed to inhibit their movement and report whether they were already preparing to move when the probe occurred. Importantly, our design enabled us to tease apart failures to inhibit—that is, when motor commands were already ballistic and participants could no longer inhibit their movement—from earlier stages. If the early RP is related to conscious intentions, we would predict an RP-like buildup before probes on trials when participants successfully inhibit their movement and then report that they were indeed preparing to move. However, if pre-probe EEG amplitude does not co-vary with participants’ subjective reports, after controlling for failures to inhibit, it would suggest that the early RP does not relate to subjective awareness of motor preparation.
2 Methods
2.1 Participants
Twenty-four participants were recruited from the student population of Chapman University (Orange, CA) (mean age 19.5 years, range: 18–22, 71% female, 92% right handed). Written informed consent was obtained from participants before they participated in the experiment. All methods were approved by the local ethics committee (Chapman University IRB-20-122).
2.2 Paradigm
We opted for a trial-based paradigm in line with computational models of self-initiated action. In every trial, participants were instructed to wait approximately 3 s and then press spacebar whenever they wanted. Participants always used their right hands regardless of handedness. Importantly, they were further instructed to fully inhibit their movement if they heard an auditory tone (the probe; 0.1 s at 500 Hz; randomly timed via a shifted gamma distribution (4, 0.75) that began after 2 s; these parameters were selected to result in roughly 50% probe trials). There were, therefore, five types of possible outcomes in every trial (Fig. 1). (1) If participants moved before 2.5 s had elapsed from trial onset, they were warned that they moved too quickly. These were designated Error Trials. (2) If participants moved before probe onset, the experiment would automatically advance to the next trial. These were designated Movement Trials. (3) If the probe went off before participants moved, they were asked whether they were already preparing to move when the probe occurred. If they answered “yes,” these probe trials were designated Prep Trials. (4) If they answered “no,” these were designated No Prep Trials. Finally, (5) if the participants pressed the button within 200 ms after probe onset, the trial was marked as “failure to inhibit” (FTI; see Section 2.5 below).
Paradigm. Participants were instructed to move whenever they felt like it after around 3 s had elapsed. At the beginning of each trial, the timing of an auditory probe was randomly generated according to a gamma distribution (top). Participants were instructed to inhibit movement following the probe and then indicate whether they were preparing to move at probe onset. These left four outcomes on each trial: movement trials, in which participants move before designated probe time; error trials, in which participants did not wait at least 2.5 s to move; probe trials, in which participants hear the probe and do not move following the probe; and failure-to-inhibit (FTI) trials, in which participants are unable to inhibit their movement following probe (presumably because the probe came too close to when they would have moved. Following the probe, participants were asked whether they were preparing to move. If they responded “yes,” the trial was designated a Preparation (Prep) trial, and if they responded “no,” the trial was designated a No Preparation (No Prep) trial. The rare trials where the participants responded “unsure” were omitted.
Paradigm. Participants were instructed to move whenever they felt like it after around 3 s had elapsed. At the beginning of each trial, the timing of an auditory probe was randomly generated according to a gamma distribution (top). Participants were instructed to inhibit movement following the probe and then indicate whether they were preparing to move at probe onset. These left four outcomes on each trial: movement trials, in which participants move before designated probe time; error trials, in which participants did not wait at least 2.5 s to move; probe trials, in which participants hear the probe and do not move following the probe; and failure-to-inhibit (FTI) trials, in which participants are unable to inhibit their movement following probe (presumably because the probe came too close to when they would have moved. Following the probe, participants were asked whether they were preparing to move. If they responded “yes,” the trial was designated a Preparation (Prep) trial, and if they responded “no,” the trial was designated a No Preparation (No Prep) trial. The rare trials where the participants responded “unsure” were omitted.
Regarding the movement, participants were instructed to press the spacebar with their right index finger spontaneously, without pre-planning, whenever they felt like it after waiting for around 3 s from the onset of the trial. However, they were further instructed not to count the seconds but to just estimate the time. For the first 40 trials, we let the participant get used to the paradigm, so we did not include probes. Hence, participants just carried out self-initiated action, while fixating on a white cross on a gray screen. After those initial trials, participants completed 20 practice trials of the main experiment (Fig. 1). After participants confirmed they understood the task instructions and required no more practice, we began the main experiment. We specifically instructed the participants not to let the possibility of a probe change how they generated actions throughout the experiment. In probe and FTI trials, 1 s after probe onset, participants were presented with a new screen and asked, “Were you preparing to move when the tone occurred?” Participants responded by pressing 1 (no), 2 (unsure), or 3 (yes) with their left hand. “Yes” responses were counted as “preparation” trials, and “No” responses were counted as “no preparation” trials. “Unsure” trials were not included in further analyses (see Section 2.5). After practicing for 20 trials, participants were asked whether they understood the experiment protocol, had the instructions repeated to them, and then they continued to the main experiment.
The actual experiment was broken into blocks of 20 trials each. Participants were given a break between blocks. Our IRB approval set a limit of 2 h on our experimental session duration. Hence, after EEG setup and practice trials, we ran as many blocks as we could fit into the 2 h session. Participants completed between 8 and 10 blocks, and hence between 160 and 200 trials.
We were concerned that prior results linking the RP to access to consciousness were confounded by trials where participants were just about to move when the probe went off (Parés-Pujolràs et al., 2019) and thus could no longer inhibit their action. To account for such cases, we instructed participants to fully inhibit all actions upon hearing the probe. Hence, trials where the motor process had become ballistic would result in button presses after probe onset, which we termed failures to inhibit (FTI). Thus, in our paradigm, we could tease apart probe trials that included ballistic button presses (before which RPs would be expected) from probe trials in which participants’ responses were based on metacognitive access to the underlying neural process.
2.3 EEG recording and preprocessing
EEG was recorded online at 2048 Hz using a BioSemi 64-channel system and then read into MATLAB (MathWorks) R2019B and preprocessed using the FieldTrip software package. Data for each trial were epoched between 3.5 s before and 2 s after movement onset (movement trials) or tone onset (probe trials). First, all trials were visually inspected for artifacts, keeping the type of trial blind; trials containing artifacts were rejected (3.3 ± 3.7% of all trials, range: 0–14%). We also removed electrodes that had poor signal-to-noise ratio based on visual inspection. Data were then downsampled to 200 Hz and bandpass filtered between 0.1 and 35 Hz (filtered using Fieldtrip Defaults; Butterworth IIR filter). Using a 0.1 Hz high-pass filter has recently been suggested to reduce the impact of infra-slow oscillations (<0.1 Hz) when investigating the RP and other slow cortical potentials (Garipelli et al., 2013; Park et al., 2020). A non-filtered set of data was retained for time–frequency analysis. Independent Components Analysis (RunICA algorithm; FieldTrip implementation) was then used to visually remove artifacts corresponding to blinks and eye movements. The ICA procedure results in an unmixing matrix, which was then also used to remove those components from the unfiltered data. Finally, we re-referenced the EEG data to the overall mean of the remaining electrodes. For ERP analysis, we baseline corrected the data by subtracting average EEG amplitude 2.5 to 2 s before movement, following prior studies (Parés-Pujolràs et al., 2019).
2.4 Time–frequency decomposition
Time–frequency analysis was conducted using custom scripts (Cohen, 2014) in MATLAB. We decomposed the data into the time–frequency domain using complex Morelet wavelets, ranging from 0.5 to 40 Hz. Wavelets started at 4 cycles and increased to 13 cycles as the frequency increased. Power was taken as the squared absolute value of the resulting complex time series. For power in specific frequency bands, the frequencies that fell within that band (delta: 0.5–4 Hz; theta: 4–7 Hz; alpha: 7–12 Hz; low beta: 12–20 Hz; high beta: 20–30 Hz; gamma: 30–40 Hz) were averaged together.
2.5 Data retention
For EEG analyses, we omitted trials according to the following determined in-advance criteria: (1) we removed probe trials on which the participants reported that they were “unsure” whether they were preparing to move (3.6 ± 3.5% of all trials, range: 0–11%), (2) we removed “error” trials on which participants pressed the button less than 2.5 s after trial onset (3.4 ± 2.9% of all trials, range: 0–11%), and (3) we removed “failure-to-inhibit” trials on which participants pressed the button more than 200 ms after probe onset (4.0 ± 5.4% of all trials, range: 0–19%). Two hundred milliseconds before movement has been identified as the “point of no return” for self-initiated movements (Schultze-Kraft et al., 2016). Therefore, movements following the probe by more than 200 ms likely reflect lapses of attention rather than actual failures of inhibition. We used 200 ms as our cutoff to be conservative, even though the distribution of post-probe button presses in our analysis seemed to fall after about 250 ms (see Fig. 2C). We retained the remaining data (besides those omitted from visualizing EEG; see Section 2.3 for details) for statistical analyses.
Behavioral results. (A) Histogram of waiting times pooled across participants (for visualization purposes only). Dotted line: 2.5 s cutoff for excluding responses from the analysis. Solid line: mean of waiting times (after excluding waiting times <2.5 s). Inset: Histogram of probe delivery times (in gray; density; also pooled across participants) and the gamma distribution used for random probe delivery schedule (in red). (B) Frequency of experimental conditions following the probe. Participants could answer the question about preparing to move with “Yes” coded as Preparation (Prep), “No” coded as No Preparation (No Prep), and “Unsure.” Or, if they moved within 200 ms of probe onset, the trial was categorized as failure to inhibit (FTI; see Section 2 for more information). (C) Histogram of button-press times when participants moved following the probe pooled across all participants. About 65.4% of movements were within 250 ms of the probe, suggesting actions were already ballistic and participants were not able to inhibit their movements (a uniform distribution across 1 s after probe onset would predict 25% rather than 65.4%). However, we retained only those button presses within 200 ms of probe onset, as this was found to be the point of no return (Schultze-Kraft et al., 2016).
Behavioral results. (A) Histogram of waiting times pooled across participants (for visualization purposes only). Dotted line: 2.5 s cutoff for excluding responses from the analysis. Solid line: mean of waiting times (after excluding waiting times <2.5 s). Inset: Histogram of probe delivery times (in gray; density; also pooled across participants) and the gamma distribution used for random probe delivery schedule (in red). (B) Frequency of experimental conditions following the probe. Participants could answer the question about preparing to move with “Yes” coded as Preparation (Prep), “No” coded as No Preparation (No Prep), and “Unsure.” Or, if they moved within 200 ms of probe onset, the trial was categorized as failure to inhibit (FTI; see Section 2 for more information). (C) Histogram of button-press times when participants moved following the probe pooled across all participants. About 65.4% of movements were within 250 ms of the probe, suggesting actions were already ballistic and participants were not able to inhibit their movements (a uniform distribution across 1 s after probe onset would predict 25% rather than 65.4%). However, we retained only those button presses within 200 ms of probe onset, as this was found to be the point of no return (Schultze-Kraft et al., 2016).
In addition, for our ERP analysis, we were specifically interested in the RP. Therefore, to account for differences in underlying brain morphologies, we inspected the obtained ERPs aligned to movement onset for movement trials at electrodes Cz, CPz, FCz, and Fz. For each participant, we used data from the electrode where the EEG signal was most negative at the time of movement (using Cz for every participant reduced RP amplitude but did not meaningfully change our results). We omitted 4 participants (out of 24, 16.7%) who did not exhibit a visible RP at any of these locations. Therefore, our main inclusion criteria had 20 participants (Cz n = 9, FCz n = 9, Fz n = 2), which is comparable to similar studies.
2.6 Statistical analyses
The number of trials was not the same across participants (see Paradigm). Due to stochastic probe delivery and the dependency of reports on participants, our methods frequently result in different number of trials for different conditions (we split the probe-trial data into four cases: preparation reported, unsure, no preparation reported, and failure to inhibit). So, rather than traditional cluster-based ERP analysis, we opted to analyze single-trial data using mixed-effects models (LMEs, also called linear mixed-effects models). LMEs were implemented using the Pymer4 package in Python, which is a port of the commonly used Lme4 package in R. In some cases, we also wanted to assess how data affected a binary variable (e.g., response to a probe being a “yes” or a “no”). For these cases, we used a logistic mixed-effects model to assess the degree to which a given variable affects the log-odds of a binary outcome variable.
2.6.1 Behavioral analysis
We also used logistic mixed-effects models to assess whether participants were more likely to report that they were preparing to move if more time had elapsed since trial onset. To do so, we used the following regression equation with random slopes:
where in the above, isprep is a binary response variable (1 for prep reported, 0 for prep not reported), and we regress time of the probe since trial onset (timeprobe) with a random intercept and effect of time for each participant. We complemented this analysis by regressing condition (Prep vs. No Prep) on probe timing:
2.6.2 Event-related potential analysis
In Figure 3 we assessed the means of each ERP using linear mixed-effects models. For Figure 3A, we assessed grand-average amplitude at each time point using the below equation with random intercepts (to avoid singularities in fitting, we omitted random slopes):
ERP Results: Pre-probe RP amplitude is not associated with subjective reports of motor preparation. (A) ERPs for Movement (green), Prep (red), No prep (blue), and FTI (black) trials. Note that the ERPs are aligned to button-press onset for the Movement trials and to probe onset for other conditions, so t = 0 is not the same for all conditions. Solid lines are means and shaded regions are standard error (obtained via Linear Mixed Effects; LME). (B) RP amplitude (mean and 95% confidence interval; LME analysis) in the 250 ms before probe onset (Prep, No Prep, FTI) or before movement onset (Movement). Statistical results from LME analysis post hoc comparison (FDR corrected) and Bayesian ANOVA shown. (*, **, and *** indicate p < 0.05, 0.01, and 0.001, respectively). (C) Zoom on pre-probe/movement period for probe trials (means and SE; LME analysis). Separating FTI and Prep trials results in negative deflection only on FTI trials (left). But mixing FTI and Prep trials creates a downward, RP-like, deflection on “prep” trials (right). (D) Post-probe ERPs aligned to probe onset (means and SE; LME analysis; baselined using period [-0.2, 0.0] s relative to probe onset). All trials showed a negative deflection of similar magnitude followed by a positive deflection that reflected both whether participants failed to inhibit their movement and their later report.
ERP Results: Pre-probe RP amplitude is not associated with subjective reports of motor preparation. (A) ERPs for Movement (green), Prep (red), No prep (blue), and FTI (black) trials. Note that the ERPs are aligned to button-press onset for the Movement trials and to probe onset for other conditions, so t = 0 is not the same for all conditions. Solid lines are means and shaded regions are standard error (obtained via Linear Mixed Effects; LME). (B) RP amplitude (mean and 95% confidence interval; LME analysis) in the 250 ms before probe onset (Prep, No Prep, FTI) or before movement onset (Movement). Statistical results from LME analysis post hoc comparison (FDR corrected) and Bayesian ANOVA shown. (*, **, and *** indicate p < 0.05, 0.01, and 0.001, respectively). (C) Zoom on pre-probe/movement period for probe trials (means and SE; LME analysis). Separating FTI and Prep trials results in negative deflection only on FTI trials (left). But mixing FTI and Prep trials creates a downward, RP-like, deflection on “prep” trials (right). (D) Post-probe ERPs aligned to probe onset (means and SE; LME analysis; baselined using period [-0.2, 0.0] s relative to probe onset). All trials showed a negative deflection of similar magnitude followed by a positive deflection that reflected both whether participants failed to inhibit their movement and their later report.
The estimated average values obtained from the fitted LME models and their standard errors were used for Figure 3A. Figure 3C was obtained using the same model except re-coding FTI trials as Prep trials. Figure 3D was obtained using the same equation except with the EEG data baselined at (-0.2, 0) instead of (-2.5, -2).
For assessing differences in pre-probe amplitude between No Prep, Prep, and FTI trials, we averaged single-trial EEG amplitudes in the 250 ms before probe onset and used the following equation with random intercepts to assess differences between conditions.
For the analysis of slope mentioned in text, we used a similar equation except we did not average amplitude, instead fitting on the preprocessed EEG. We also included a random effect of condition to account for the relatively lower number of prep trials present in our data (model would not converge with a random effect of time or the interaction term):
2.6.3 Time–frequency analysis
For the analysis of power in the time–frequency domain in Figures 5A and 6, we fitted single-trial power data at every time–frequency pair using LMEs with a fixed effect of condition and random effect of participant (we did not use random slopes because of model fitting issues):
Then, the reported power in Figure 5A was taken as the marginal estimate of power for each condition. For Figure 6, we restricted the conditions to those specified and then conducted the same analysis and reported estimated differences using post hoc testing. In Figure 6 we omitted time–frequency pairs that had an uncorrected p-value greater than 0.05 because that analysis was exploratory. Because these analyses entailed 1200 tests (one for each time–frequency pair) for each of 4 channels of interest, there are likely some spurious findings.
For specific analysis of pre-movement and pre-probe beta power (Fig. 5B), we averaged beta power across the low- (12–20 Hz) and high- (20–30 Hz) beta bands in the 250 ms before probe/movement to get single-trial beta power values. Then, we fit an LME to these data using the following equation, including condition as a fixed effect and a random effect of participant:
We established significance using post hoc tests implemented in the pymer4 python library. For Bayesian analysis we ran a Bayesian ANOVA with the same structure in JASP (JASP Team (2023)) version 0.6.13 and reported Bayes factors obtained from post hoc tests.
2.7 Computational modeling
We also conducted computational modeling of the action-generation and metacognitive process to explain the basis of our ERP results. Note that all units are arbitrary and inverted for plotting to match the ERPs. In some cases, model parameters are taken from prior studies (e.g., stochastic accumulation). In those cases, we note the source of parameters. In other cases, model parameters were tuned manually to simultaneously obtain an RP shape and response time distribution close to real data on movement trials (i.e., without investigating differences based on Prep or No Prep trials).
2.7.1 Classic RP model
The so-called classic interpretation of the RP has not been given a formal model before, so we began by doing so. The interpretation posits that RP onset corresponds to an “unconscious decision” followed by a monotonic rise in activity. We implemented this in an accumulator framework as a noisy fluctuation process that does not cross threshold itself. After some time has passed (modeled as a gamma distribution), a strong input turns on (reflecting an unconscious decision) that drives the accumulator to the threshold in roughly 1 s.
where 1A(u) is the indicator function for u (i.e., gives 1 if u is true, 0 if u is false), and ton is a randomly generated time at which a strong input drives the accumulator to the threshold. In our simulations, the threshold is 4/3 and dt = 0.001. Individual runs, threshold-crossing distribution, and threshold-locked averages are shown in Supplementary Figure S2. Probe delivery times were modeled as a shifted gamma function:
We first simulated 10,000 runs of 20 s, and then for each run calculated when tprobe occurred relative to ton. If tprobe occurred before ton, then the trial was categorized as a No Prep. Otherwise, if tprobe occurred after ton but before threshold crossing, the trial was categorized as a Prep. If tprobe occurred within 200 ms of threshold crossing, the trial was categorized as FTI. Otherwise, the trial was categorized as a Movement trial. Figure 5A right gives the average traces of these four conditions.
2.7.2 Stochastic accumulator model
In the stochastic accumulator model, movement is triggered upon autocorrelated fluctuations (implemented as drift–diffusion process) crossing a threshold. We implemented the model following Schurger et al. (2012), using the equation:
where I is the drift rate, k is the leak, x is the accumulator amplitude, and c is a noise scaling term. We used the parameters (I = 0.11, k = 0.5, c = 0.1, dt = 0.001, and a threshold of 0.298, from Schurger et al. (2012)). Individual runs, threshold-crossing distribution, and threshold-locked averages are shown in Supplementary Figure S2.
2.7.3 Linear ballistic accumulator model
Recently, Bogler et al. (2023) suggested that linear ballistic accumulation dynamics (Brown & Heathcote, 2008) could explain the early rise and shape of the RP. In linear ballistic accumulation, movement is triggered when a linearly increasing accumulator crosses a threshold. We, therefore, also simulated these dynamics to test whether our study’s findings could distinguish between the two underlying processes. In order to more accurately recreate the shape of the RP, we included a fixed delay of 2500 ms, such that the LBAc would only begin 2500 ms into the trial. Note that without this nonlinearity, the average trace of the LBA before a threshold crossing would be a straight line. Also note that we tested a variety of parameter combinations, and they largely did not change our results. Our simulation was governed by the following equation (adapted from Bogler et al. (2021)):
where I is the slope (normally distributed) and x0 is the starting point of the accumulator. We used a normal distribution with mean = 2, standard deviation = 1 for slopes in our model, whereas Bogler et al. (2023) used a normal distribution with mean = 1, standard deviation = 2. We made this change because, in simulating the model runs, we found that using their distribution would sometimes result in very shallow or even negative slopes on some trials. This led to long waiting times and positive-trending pre-probe EEG on No Prep trials, so we modified parameters to better match behavioral and EEG data.
2.7.4 Pink-noise accumulator model
Schurger (2018) hypothesized and found evidence that the RP might reflect a noisy autocorrelated input to an accumulator. This claim was evidenced by that RP amplitude being higher for trials that had shorter waiting times in comparison with longer waiting times. However, we did not find such a difference (Supplementary Fig. S1B). Nevertheless, we included this model here as an alternative to the stochastic accumulator model. Briefly, we generated traces of pink noise with a 1/f slope of 1.5 using the “colorednoise” package in Python. If we let such traces be denoted P, our accumulators obeyed the following equation:
where Pt denotes the tth entry in P. We used the same parameters as the stochastic accumulator above, except we used the noisy input P as the traces for further analyses.
2.7.5 Single-stage reporting model
In our single-stage reporting model, whether a trial was designated Prep or No Prep was based on accumulator amplitude at the time of the probe. Probe delivery times were generated using the following distribution:
We first simulated 10,000 runs of 20 s, and then for each run calculated when tprobe occurred relative to threshold crossing. Trials where tprobe occurred before threshold crossing were counted as probe trials. Otherwise, the trial was counted as a regular press trial.
The amplitude at the time of the probe was then extracted for each probe trial and categorized mirroring the percentages present in our behavioral data; the lowest 53.2% were categorized as No Prep trials, the next 11.4% were counted as Unsure and discarded, and finally the highest remaining 35.4% were counted as Prep trials. Using a median split to define Prep versus No Prep, or using an 40–20–40 split for Prep, Unsure, and No Prep left results unchanged. Model runs were then grouped according to this classification and averaged for Figure 4C top row.
Computational modeling results: dual-stage metacognitive access to stochastic accumulation can explain ERP results. (A) Classic model simulations. Left: model schematic. If the probe occurs before (after) an RP onset, the trial is categorized as a No Prep (Prep) trial. Right: modeling results. Prep trials show moderate deflection while No Prep trials do not. Move trials show stronger deflection. (B) Modeling framework. Top: Schematic of two accumulator models; Left: stochastic accumulator: a noisy diffusion process triggers movement upon crossing a threshold (dashed line). Right: linear ballistic accumulator: linearly accumulating process triggers movement upon crossing the threshold. Bottom: Schematic of metacognitive reporting demonstrated on stochastic accumulator dynamics. Left: single-stage model, wherein reports are based on the magnitude of the accumulator process at the time of the probe (roughly, above or below the dashed gray threshold; see Section 2). Right: dual-stage model, where participants’ reports are based on the slope of the accumulator immediately after probing. (C) Modeling results. We tested two types of underlying accumulation dynamics each with two types of reporting models. With single-stage metacognition, both stochastic and linear ballistic accumulation predict differences between Prep and No Prep trials. Only dual-stage metacognition with stochastic accumulator dynamics matches ERP results, and also recreates the slight negative deflection before both Prep and No Prep trials.
Computational modeling results: dual-stage metacognitive access to stochastic accumulation can explain ERP results. (A) Classic model simulations. Left: model schematic. If the probe occurs before (after) an RP onset, the trial is categorized as a No Prep (Prep) trial. Right: modeling results. Prep trials show moderate deflection while No Prep trials do not. Move trials show stronger deflection. (B) Modeling framework. Top: Schematic of two accumulator models; Left: stochastic accumulator: a noisy diffusion process triggers movement upon crossing a threshold (dashed line). Right: linear ballistic accumulator: linearly accumulating process triggers movement upon crossing the threshold. Bottom: Schematic of metacognitive reporting demonstrated on stochastic accumulator dynamics. Left: single-stage model, wherein reports are based on the magnitude of the accumulator process at the time of the probe (roughly, above or below the dashed gray threshold; see Section 2). Right: dual-stage model, where participants’ reports are based on the slope of the accumulator immediately after probing. (C) Modeling results. We tested two types of underlying accumulation dynamics each with two types of reporting models. With single-stage metacognition, both stochastic and linear ballistic accumulation predict differences between Prep and No Prep trials. Only dual-stage metacognition with stochastic accumulator dynamics matches ERP results, and also recreates the slight negative deflection before both Prep and No Prep trials.
2.7.6 Dual-stage reporting model
In our dual-stage reporting model, a trial was designated Prep or No Prep according to the accumulator slope immediately following the probe, as opposed to the amplitude at the time of probe delivery. This model was developed following post-decision metacognitive models in perceptual decision making (Desender et al., 2021; Moran et al., 2015; Pleskac & Busemeyer, 2010). In such models, confidence reports are based on whether evidence continues to accumulate in the same direction after reaching a decision threshold—if the accumulation continues, the participant is confident, but if the evidence stops accumulating, the participant is less confident (some variants include whether the post-decision accumulation hits a secondary, higher threshold). We used a simplified version of such models by taking the slope of the post-probe accumulation. Probe delivery times were generated using the same equation as the single-stage model, and the simulations followed the same protocol. For trials categorized as probe trials, we fit a linear regression on the 200 ms following the probe. We tested a variety of window sizes—100, 200, 300, and 400 ms—and the choice of window size did not substantively affect our results. So, we report results from the 200 ms window below, a duration that is in line with the perceptual metacognition literature. Then, we categorized trials using the same percentiles as the single-stage model (i.e., highest post-probe slope percentile gets counted as Prep, lowest as No Prep, and medium as Unsure). Model runs were then grouped according to this classification and back averaged for Figure 4C bottom row.
3 Results
Participants carried out a self-paced movement task, in which they were sometimes interrupted by an auditory tone that was scheduled to be delivered at a random time in each trial (the probe; see Fig. 1 and Section 2.2). Participants were instructed to move whenever they felt like it after around 3 s had elapsed from trial onset, but to abstain from moving if the probe occurred. If they moved before 2.5 s had elapsed, they were given a warning, and the trial was aborted and designated an Error trial. Alternatively, if they moved after 2.5 s elapsed but before the probe went off, the trial was designated a Movement trial and participants immediately advanced to the next trial. Otherwise, the probe occurred before participants spontaneously moved. Despite being instructed not to move if the probe went off, participants sometimes pressed the button after probe onset. These presumably reflected failures to inhibit the movement (FTI) for those trials—that is, trials where the participants could no longer inhibit the movement at probe onset (or, in other words, the movements were already ballistic). We designated trials where the button presses took place within 200 ms of probe onset as FTI trials because 200 ms before movement marks the “point of no return” at which action-initiation processes become ballistic (Schultze-Kraft et al., 2016) (see Section 2.5). Finally, if no movement followed the probe, participants were asked whether they were preparing to move when the probe went off. They could respond Yes, in which case the trial was designated a Preparation trial; Unsure, in which case the trial was discarded; or No, in which case the trial was designated a No preparation trial. The primary trials of interest were (1) Movement Trials (Move), (2) FTI trials, (3) Preparation (Prep) trials, and (4) No Preparation (No Prep) trials. Other trials were not analyzed further.
3.1 Behavioral results
On average (grand averaged across participant-specific averages), participants waited 3.98 s before moving (STD: 0.96, range: 3.38–5.91; see Fig. 2A for all waiting times pooled across participants) and were interrupted by probes on 39.2% of trials (STD: 18.2%, range: 14.0–85.0%; see Fig. 2A insert for all probe delivery times). The heterogeneity across participants in the frequency of being probed is likely because some participants tended to move earlier (later) in the trial than others. Such movements were less (more) likely to encounter a probe before they moved. This also explains why the empirical distribution of probe delivery times was skewed earlier than the distribution used for probe scheduling (red line in Fig. 2A insert): later probes were more likely to be preempted by a participant’s movement and hence less likely to go off.
Among probe trials, participants reported preparation in 26.9% of trials, no preparation in 40.4%, and being unsure in 8.7%; a further 24.1% of trials were FTI (Fig. 2B). Thus, in 88.5% of non-FTI trials, participants were confident in their ability to report having prepared or not prepared to move at probe onset, suggesting that they were largely capable of performing the task as instructed. Overall, averaged across participants, we obtained 117.2 ± 35.4 (mean ± STD) Movement trials, 20.2 ± 17.2 Prep trials, 33.3 ± 25.4 No prep trials, 7.0 ± 7.0 trials where participants were unsure about preparing to move, and 15.7 ± 11.6 FTI trials. Notably, most button presses following the probe occurred within 250 ms of probe onset (65% compared with 25% at chance level assuming a uniform distribution across 1 s; Fig. 2C), suggesting that participants strove to follow instructions and inhibit movements after probe onset.
Removing FTI and unsure responses from probe trials, we found that the probability of reporting preparation significantly increased with time from trial onset (βtime = 0.322, p = 0.001; Logistic Mixed Effects). Prep trials also tended to have longer wait times than No prep trials (p = 0.023; Linear Mixed Effects), providing further evidence that participants were able to accomplish the task successfully. Similarly, among all probe trials, the probability a participant would fail to inhibit a movement also increased significantly with time from trial onset (βtime = 0.447, p < 0.001; Logistic Mixed Effects).
3.2 EEG and modeling results
3.2.1 The RP does not reflect awareness of motor intention
We investigated pre-probe EEG during Movement trials, FTI, Prep, and No prep trials. We used linear mixed-effects (LME) analysis to estimate voltage at each time point for each condition (Fig. 3A; see Supplementary Fig. S1C–D for other estimation methods). A clear RP was visible for Movement trials (Fig. 3A in green; see individual participants in Supplementary Fig. S1A). FTI trials were also preceded by a significant negative deflection compared with Prep and No Prep trials, but we did not observe a difference between pre-probe EEG when comparing Prep and No Prep trials. To investigate the significance of these results, we investigated single-trial amplitudes in the 250 ms before probe onset (or button press for movement trials) using linear mixed-effects models (LME; implemented in python using the pymer4 library version 0.7.7 (Jolly, 2018)) and a Bayesian ANOVA (implemented in JASP (JASP Team (2023)) version 0.6.13). Critically, we found no difference between pre-probe EEG amplitude for Prep and No Prep trials, with Bayes factors providing strong evidence that the two trial types were equivalent (Fig. 3B; post hoc tests from LME analysis: No Prep vs. Prep: t(3305.9) = 0.874, ptukey = 0.004, BF10 = 0.091; No Prep vs. Movement: t(2345.7) = 5.140, ptukey < 0.001, BF10 = 815.4; No Prep vs. FTI: t(3206.6) = 3.260, ptukey = 0.006, BF10 = 21.0; Prep vs. Movement: t(2903.4) = 2.948, ptukey = 0.017, BF10 = 2.127; Prep vs. FTI: t(2.994) = 2.059, ptukey = 0.167, BF10 = 1.584; Movement vs. FTI: t(3440.4) = -0.223, ptukey = 0.996, BF10 = 0.072).
To account for concerns that the choice of baseline can artificially introduce differences in amplitude, we also investigated the EEG slope in the second before movement, because slopes are not affected by the choice of baseline. To do so, we fit LME models to the EEG data in the second before probe onset/movement. Consistent with the above analysis, pre-probe EEG slope was more negative in FTI and Movement trials than in Prep and No Prep trials, whereas the slope for Prep and No Prep trials was not significantly different (post hoc tests from LME analysis: No Prep vs. Prep: Z = -0.572, ptukey = 0.940; No Prep vs. Movement: Z = 6.657, ptukey < 0.001 No Prep vs. FTI: Z = 3.260, ptukey < 0.001; Prep vs. Movement: Z = 2.948, ptukey < 0.001; Prep vs. FTI: Z = 2.059, ptukey < 0.001; Movement vs. FTI: Z = 0.704, ptukey = 0.896; note that the Movement condition is by necessity aligned to a different event—movement—than the other conditions, which are aligned to probe onset). However, pre-probe slope was indeed slightly negative on both Prep and No Prep trials (Prep: βslope = -0.509 μV/s, 95% CI: [-0.935, -0.083]; No Prep: βslope = -0.661 μV/s, 95% CI: [-0.962, -0.361]). Notably, these findings were similar when we tested our data with more stringent exclusion criteria for participants (see Supplementary Table S1).
Furthermore, we hypothesized that combining EEG signals from trials that included ballistic button presses (i.e., FTI trials) with EEG signals from trials that include presses stemming from actual metacognitive decisions that the participant was preparing to move (i.e., Prep trials) would result in an average ERP trace that would show a more negative deflection than No Prep trials. We tested this in our dataset and indeed found that combining Prep and FTI trials in this manner resulted in greater pre-probe negativity (Fig. 3C). Additionally, when investigating post-probe ERPs, we observed a negative deflection that began around 150 ms after probe onset (Fig. 3D). This deflection was of similar magnitude across Prep, No Prep, and FTI trials. This initial deflection was followed by a positive deflection that was visibly largest and peaked latest for FTI trials, followed by Prep trials then No Prep trials.
3.2.2 Reports based on stochastic accumulation can explain ERP results
Our ERP results suggest that there is no difference in pre-probe RP amplitude across trials between the condition where preparation was reported and the one where it was not, though our behavioral results suggest that participants were able to report on their state of preparation. For this to be possible, whichever feature of neural activity is accessed for these reports must be statistically independent (or only weakly dependent) from pre-probe RP buildup. In this section, we use computational models of the putative action-initiation neural mechanisms to investigate what processes could underlie our ERP results.
First, we looked at the classic RP interpretation—that its deflection reflects motor preparation following an unconscious decision to move (Libet, 1985). Under that interpretation, probing after (before) the onset of the RP will result in the reported presence (absence) of motor preparation (Matsuhashi & Hallett, 2008; Schultze-Kraft et al., 2020). We developed a simple computational model to capture the essential components of this interpretation. Briefly, activity fluctuates noisily (random walk) until the onset of an “implicit decision,” after which a strong input drives activity up toward a threshold. Movement is initiated when that threshold is crossed. Although the RP usually has an exponential shape, we used a linear input for simplicity (see Classic RP model in Section 2.7; results are the same if we use an exponential buildup). In this model, if a randomly timed probe is delivered before (after) the implicit decision, the trial would be categorized as No Prep (Prep) (Fig. 4A left). If the probe was delivered after threshold crossing, the trial was designated a Movement trial (we did not model FTI trials for simplicity). In this framework, pre-probe activity is statistically dependent on report (by definition), because the report explicitly depends on whether or not the “implicit decision” to move has been made. Accordingly, we predicted that pre-probe activity would differ across Prep and No Prep trials for the classic RP model.
We simulated 10,000 trials under these conditions (Fig. 4A; see Section 2.7.1 for details; see Supplementary Fig. S2 top for example runs). Consistent with our predictions, we observed a flat ERP on No Prep trials (blue), a slight negative pre-probe deflection in Prep trials (red), and a stronger negative deflection before the probe or movement on FTI and movement trials (black and green, respectively). Comparing the model’s predictions with our empirical results (Fig. 3A) suggests that they are qualitatively not a good fit. In particular, the model predicts that there would be a difference between Prep and No Prep trials that grows with time, which we did not observe in the empirical data. Hence, the classic RP model is incompatible with our empirical results.
We next considered two prominent computational models of action initiation—stochastic accumulation (Khalighinejad et al., 2018; Schurger et al., 2012) (Fig. 4B top left) and linear ballistic accumulation (Bogler et al., 2021; Brown & Heathcote, 2008) (Fig. 4B top right). In stochastic accumulation, movement is initiated when a noisy diffusion process crosses a threshold (see Section 2.7 for details). Critically, that process is not monotonic—it can rise and fall until threshold crossing—and only resembles an exponential rise when each run is aligned to its threshold crossing and many trials are then back averaged (Schurger et al., 2012). In our linear ballistic accumulator model, movement is initiated when a monotonic linear signal that begins increasing at trial onset crosses a threshold. The accumulator’s slope is drawn from a half-normal distribution at the onset of the accumulation process, and the accumulation then proceeds linearly and monotonically with that slope (see Supplementary Fig. S2 for individual model runs).
We combined these two action-initiation models with two models of the reporting process. In the first reporting model, reports were based on the accumulator amplitude at the time of the probe: if the accumulator is relatively high (low) at probe onset—that is, if the accumulator was closer to (farther from) the threshold—the trial was categorized as Prep (No Prep) (Fig. 4B bottom left, where—schematically—the gray dashed line marks the threshold between preparing and not preparing; see Section 2.7.5 for details). We termed this a single-stage reporting model. Notably, both stochastic and linear ballistic accumulation are autocorrelated processes, and thus amplitude at a particular time (e.g., at probe onset) is statistically dependent on the activity immediately prior. Therefore, we predicted differences in pre-probe RP for both models when accumulator amplitude served as the basis for reports. Hence, we also considered a second reporting model based on the direction of the post-probe accumulation. If the slope of the accumulator was relatively high (low) in a short window following probe onset, the trial was categorized as Prep (No Prep) (Fig. 4B lower right; we used 200 ms for presented simulations but tested a variety of windows and window size did not qualitatively change our results). We termed this a dual-stage reporting model (see Sections 2.7.5 and 2.7.6 for details). Perceptual metacognition is sometimes better explained by dual-stage models, where reports are based on whether evidence continues to accumulate after crossing the decision threshold (Desender et al., 2021; Moran et al., 2015), suggesting such a metacognitive model may be appropriate in the domain of self-initiated actions. Critically, because stochastic accumulation is a summation over white noise, whether the accumulator increases or decreases at any time point is only weakly independent from prior activity (with said weak dependence emerging due to accumulator input and leak). Therefore, we predicted no difference between pre-probe activity for Prep and No Prep trials when combining stochastic accumulation with dual-stage reporting. During linear ballistic accumulation, however, accumulator trajectory is linear, so slope at any time is not independent from previous activity, and thus we predicted differences in pre-probe RP for this model.
We, therefore, simulated two accumulation models, each with two reporting models, hence four models overall: (1) stochastic accumulation with single-stage reporting, (2) stochastic accumulation with dual-stage reporting, (3) linear ballistic accumulation with single-stage reporting, and (4) linear ballistic accumulation with dual-stage reporting. We simulated 10,000 runs of each model type and contrasted average traces before threshold crossings and simulated probe onset. Both stochastic accumulation and linear ballistic accumulation recreated the shape of the RP in Movement trials (inverted to match the negative trajectory of EEG; Fig. 4C). Consistent with our predictions, model (2), stochastic accumulation with dual-stage reporting (bottom left in Fig. 4C), was the only model that was able to recreate our finding that pre-probe EEG traces on Prep and No Prep trials were largely overlapping (Fig. 3 top). This model also predicted a slight negative slope in pre-probe EEG on both Prep and No Prep trials, which was also observed in experimental data (Fig. 3). All other model combinations led to large visible differences in pre-probe activity when comparing Prep and No Prep trials (Fig. 4C). Another version of the stochastic accumulator model, which we term the pink-noise accumulator, has been suggested to underly self-initiated action and the RP (Schurger, 2018). In this model, the RP reflects an autocorrelated noisy input to stochastic accumulation, rather than the state of the accumulator itself (Schurger, 2018). Interestingly, this variant of the stochastic accumulator predicted differences between Prep and No Prep trials for both single- and dual-stage models (Supplementary Fig. S3A). We also tested models where the metacognitive decision was made after a delay (to account for auditory processing and task shifting, among other possibilities), and the results were largely the same as the models discussed here (Supplementary Fig. S3B).
Of the computational models we tested, stochastic accumulation with dual-stage reporting was the closest to recreating our ERP results. In stochastic accumulation, this is because slope following probe onset (the feature used in dual-stage reporting) has a weak statistical relationship with pre-probe activity, whereas the features used for reporting in all other models were strongly related to pre-probe activity. Hence, this is the only model that shows little-to-no difference in pre-probe activity when comparing Prep and No Prep trials. However, another possibility is that reports of preparation are made based on a neural feature unrelated to the RP, as has recently been suggested (Parés-Pujolràs et al., 2023) in opposition to prior probe method studies (Matsuhashi & Hallett, 2008; Parés-Pujolràs et al., 2019; Verbaarschot et al., 2016). We assessed this possibility by repeating the previous modeling and classifying probe trials as Prep or No Prep trials randomly (as would be the case if the neuronal basis of reports was unrelated to the RP). As expected, we found no difference in the pre-probe RP amplitude for all tested models under this assumption (Supplementary Fig. S3C).
3.2.3 Pre-probe beta power may track reported state of preparation
We next investigated changes of beta power in contralateral motor cortex (electrode C3) before movement and probe onset. We observed beta desynchronization on Movement and FTI trials (Fig. 5A). We also observed visibly less beta power prior to probe onset for Prep compared with No Prep trials. We next computed average beta power in the 250 ms (as in our RP analysis) prior to probe onset/movement to investigate whether differences between conditions were significant. We separately analyzed low beta (12–20 Hz) and high beta (20–30 Hz) because pre-movement desynchronization was most prominent in the low-beta band in our data (analyzing beta as a whole, 12–30 Hz, led to similar but weaker results as analyzing the low-beta band). Furthermore, pre-movement desynchronization occurs at different times for low and high-beta band activity (Shibasaki & Hallett, 2006), suggesting separate analysis of these bands is appropriate. We fit LME models and conducted post hoc tests to investigate differences and used a Bayesian ANOVA in JASP to estimate Bayes Factors. Pre-probe power in the low-beta band significantly differed across trial type (Fig. 5B upper; No Prep vs. Prep: t(3979.6) = 3.421, ptukey = 0.004, BF10 = 0.711; No Prep vs. Movement: t(3987.4) = 6.459, ptukey < 0.001, BF10 > 1000; No Prep vs. FTI: t(3974.6) = 5.591, ptukey < 0.001, BF10 > 1000; Prep vs. Movement: t(3982.4) = 1.467, ptukey = 0.458, BF10 = 0.876; Prep vs. FTI: t(3975.9) = 2.857, ptukey = 0.022, BF10 = 60.2; Movement vs. FTI: t(3972.8) = 2.287, ptukey = 0.101, BF10 = 0.606). However, we found no differences in pre-probe power in the high-beta band across conditions (Fig. 5B lower; No Prep vs. Prep: t(3974.3) = 1.987, ptukey = 0.193, BF10 = 0.066; No Prep vs. Movement: t(3978.0) = 1.003, ptukey = 0.748, BF10 = 0.046; No Prep vs. FTI: t(3972.3) = 1.233, ptukey = 0.606, BF10 = 0.097; Prep vs. Movement: t(3975.5) = -1.442, ptukey = 0.473, BF10 = 0.057; Prep vs. FTI: t(3972.8) = -0.222, ptukey = 0.996, BF10 = 0.106; Movement vs. FTI: t(3971.6) = 0.750, ptukey = 0.877, BF10 = 0.180).
Analysis of pre-probe beta desynchronization. (A) Spectrograms of beta-range activity (12–30 Hz) in the 2 s prior to probe onset (Prep, No Prep, failure to inhibit) or movement. Grand-average power estimates obtained by fitting an LME model at each time X frequency pair. (B) Estimated beta power (mean and 95% CI, LME) in the 250 ms before probe onset or movement for the low-beta (12–20 Hz) and high-beta (20–30 Hz) bands. No significant modulation was found in the high-beta band, but the condition was significantly associated with pre-probe power in the low-beta range. In particular, Prep trials were associated with lower pre-probe beta power than No Prep trials. (*, **, and *** indicate p < 0.05, 0.01, and 0.001, respectively).
Analysis of pre-probe beta desynchronization. (A) Spectrograms of beta-range activity (12–30 Hz) in the 2 s prior to probe onset (Prep, No Prep, failure to inhibit) or movement. Grand-average power estimates obtained by fitting an LME model at each time X frequency pair. (B) Estimated beta power (mean and 95% CI, LME) in the 250 ms before probe onset or movement for the low-beta (12–20 Hz) and high-beta (20–30 Hz) bands. No significant modulation was found in the high-beta band, but the condition was significantly associated with pre-probe power in the low-beta range. In particular, Prep trials were associated with lower pre-probe beta power than No Prep trials. (*, **, and *** indicate p < 0.05, 0.01, and 0.001, respectively).
3.2.4 Exploratory analysis of what neural activity relates to reporting and report content
We next investigated other potential relationships between neural activity and the reported awareness of motor preparation. We focused on electrodes above brain regions that have previously been associated with intention awareness—frontal (F1), central (Cz), contralateral motor (C3—as participants always moved their right hand), and parietal (P1) sites. We extracted power at different frequencies and times using a Morelet wavelet decomposition (Cohen, 2014) (see Section 2.4). We assessed whether spectral power depended on trial type by fitting an LME model to each time–frequency pair. Figure 6 shows the results of this exploratory analysis, with time–frequency pairs that had p > 0.05 (uncorrected) whited out. These analyses were exploratory and, in the latter case, likely confounded by factors such as auditory processing, yet we thought they may be of interest to the field and opted to include them.
Results from exploratory time–frequency analysis. (A) Time–frequency plots comparing report content at midfrontal, motor, frontal, and parietal sites (estimated difference in power between Prep vs. No Prep trials estimated at each time X frequency pair using LME models). Red (blue) signifies greater power on Prep (No Prep) trials. Time–frequency pairs resulting in non-significant modulation are whited out (p > 0.05 uncorrected, less conservative because this is an exploratory analysis). Notably, report content significantly varies with activity in the second following probe onset. (B) Time–frequency activity associated with making probe reports compared with not (estimated difference in power on Prep and No Prep trials compared with movement trials). Analysis conducted as in A. Reporting is associated with broad changes in activity in theta and alpha bands following probe onset.
Results from exploratory time–frequency analysis. (A) Time–frequency plots comparing report content at midfrontal, motor, frontal, and parietal sites (estimated difference in power between Prep vs. No Prep trials estimated at each time X frequency pair using LME models). Red (blue) signifies greater power on Prep (No Prep) trials. Time–frequency pairs resulting in non-significant modulation are whited out (p > 0.05 uncorrected, less conservative because this is an exploratory analysis). Notably, report content significantly varies with activity in the second following probe onset. (B) Time–frequency activity associated with making probe reports compared with not (estimated difference in power on Prep and No Prep trials compared with movement trials). Analysis conducted as in A. Reporting is associated with broad changes in activity in theta and alpha bands following probe onset.
To investigate what activity reflected the content of reports, we compared Prep with No Prep trials. Activity in several sites seemed to associate with report content (Fig. 6A). Notably, much of the modulation we found was present after probe onset. Most prominently, we found midfrontal delta (0.5–4 Hz) and theta (4–7 Hz) power was higher (red) on Prep than on No Prep trials immediately after probe onset (Fig. 6A top left; peak modulation at 2.52 Hz, 350 ms after probe onset, p = 3.44 x 10-28 uncorrected). We also found a cluster of decreased (blue) parietal alpha (7–12 Hz) power (Fig. 6A top right; most extreme p = 4.04 x 10-8 at 9.6 Hz, 350 ms after probe onset). Notably, C3 was the only site we investigated where we observed differences prior to probe onset, likely reflecting the pre-probe beta differences we investigated in Figure 5.
To investigate what activity reflected the metacognitive processes involved in reporting, we compared Prep and No Prep trials together with Movement trials. Once again, activity in several sites seemed to associate with the act of reporting (Fig. 6B), while modulations largely occurred after probe onset, excepting beta band activity over contralateral motor cortex, which likely reflects the stronger beta desynchronization on movement trials than Prep and No Prep trials (hereafter just referred to as “Probe trials”). We found a cluster of broadly increased theta power following probe onset on Probe trials compared with Movement trials (Fig. 6B top left; peak modulation at 2.52 Hz, 350 ms after probe onset, p = 3.44 x 10-28 uncorrected), which may be an artifact of hearing the probe. We also found a cluster of decreased alpha power beginning several hundred milliseconds after probe onset, most prominently over Parietal sites (Fig. 6B bottom right; peak modulation at 13.67 Hz, 600 ms after probe onset, p = 4.48 x 10-23 uncorrected). Again, these analyses are highly exploratory and it is not clear whether these results reflect the reporting process per se or another process, due to our design.
4 Discussion
We set out to investigate what process the readiness potential (RP) reflects and how it relates to the awareness of the preparation to move. We ran a variant of a self-initiated action task, where we sometimes interrupted participants before they moved with an auditory probe. We instructed participants to inhibit movement in response to the probe and only then report whether they were preparing to move when the probe arrived. This modification allowed us to separate cases of late-stage action initiation, when movement was already ballistic and participants failed to inhibit it, from earlier stages. Contrary to prior results (Parés-Pujolràs et al., 2019; Schultze-Kraft et al., 2020), we found no relationship between reported awareness of motor preparation and pre-probe RP-like buildups after controlling for trials where the participants failed to follow the instructions and inhibit their movement in response to the probe. As we showed, such ballistic actions were preceded by an RP. This finding suggests that prior reports that pre-probe RP-like deflections were associated with reported awareness of motor preparation reflect the late but not early RP. Instead, in line with recent reports (Parés-Pujolràs et al., 2023), we found that beta power—and more specifically low-beta power, in the 12–20 Hz range—over contralateral motor cortex prior to probe onset covaried with reported awareness of motor preparation.
Our findings suggest that the early buildup of the RP does not reflect a phenomenal experience of preparing or initiating an action. These results are, therefore, in line with findings that the timing of the RP’s buildup does not relate to reported timing of the conscious intention to move timed using a clock (Bredikhin et al., 2023; Haggard & Eimer, 1999; Schlegel et al., 2013). Our results are also in line with another recent probe study, which found similar results regarding the RP (Parés-Pujolràs et al., 2023). In the same vein, our findings also suggest that prior claims that the pre-probe RP amplitude relates to awareness of motor preparation are based on the inability of some probe paradigms to tease apart ballistic movements (when the probe was delivered when the participants could no longer refrain from moving) from trials where the participants reported preparing to move. More specifically, Parés-Pujolràs et al. (2019) found an RP before probes when participants reported they were preparing to move. However, in their paradigm, participants were instructed to move immediately if the probe occurred when they were already preparing to move (in contrast to our instruction to participants to inhibit movement in such cases). Hence, due to their experimental design, it was impossible for them to tease apart trials where the probe arrived when movement was already ballistic from trials where the participants moved in response to the probe to indicate they were preparing to move. Similarly, Schultze-Kraft et al. (2020) found that participants were more likely to report having an intention upon being probed when an online, real-time, closed-loop decoding system determined that an RP was present. However, predicting when someone will make a spontaneous movement from EEG is notoriously difficult. For instance, Bai et al. (2011) were able to predict an upcoming movement only around 0.62 s before its onset on average (with an average accuracy of 78% correct; other studies found similar results e.g., Lew et al., 2012; Lew et al., 2014), whereas RP onset occurs more than 1.5 s before the movement. Notably, those studies achieved their results by simultaneously analyzing tens of electrodes. Therefore, Schultze-Kraft and colleagues’ finding that probes delivered when a BCI detected an RP were more likely to elicit reported awareness of motor preparation was also likely based on the later stages of the RP.
Parés-Pujolràs et al. (2023) recently ran another probe method study that controlled for ballistic button presses by having participants report whether they were preparing to move at probe delivery using the hand opposite the one used for the self-paced action task. With this change in task design, they found no difference in pre-probe RP amplitude when comparing trials in which participants reported preparing versus not preparing to move. They hypothesized that their prior results (Parés-Pujolràs et al., 2019) were “confounded by some trials where self-paced action preparation was interrupted prior to awareness.” We found that the pre-probe EEG waveform on failure-to-inhibit trials was nearly identical in morphology and amplitude to the RP on movement trials (Fig. 3A–B; although it is important to note that these two waveforms are aligned to different events—probe onset in the former and movement in the latter). These results suggest that participants were at the very latest stages of action initiation when the probe was delivered and could thus no longer refrain from moving. Paradigms that cannot distinguish between these kinds of trials and those in which participants are probed earlier in the movement-generating process may, therefore, end up conflating these two trial types, which could lead to misleading results.
We next contrasted the ability of the classic RP model, stochastic accumulation, and linear ballistic accumulation to explain our RP results. Only stochastic accumulation, when combined with a dual-stage metacognition model, could successfully explain our EEG findings. Among other things, all other models predicted differences in pre-probe RP amplitude between Prep and No Prep trials, contrary to our results. Critically, our ERP results directly contradict conceptual models wherein RP onset corresponds to a specific neurocognitive event, after which participants (if probed) will report that they were preparing to move (Matsuhashi & Hallett, 2008; Parés-Pujolràs et al., 2019; Schultze-Kraft et al., 2020; Verbaarschot et al., 2016).
More generally, our computational modeling results support the view that the RP reflects stochastic accumulation to a threshold that is aligned to and backward averaged from threshold crossings, and, furthermore, that subjective reports reflect dual-stage metacognitive access to this process in response to hearing the probe. Only this combination of action-initiation and reporting models successfully explained our ERP results. Furthermore, our model also predicted the slight negative deflection in pre-probe ERP, which we observed on probe trials regardless of later reported awareness. Stochastic accumulation is a more recent yet influential model of self-initiated action and the RP (Khalighinejad et al., 2018; Maoz et al., 2019; Schurger, 2018; Schurger et al., 2012). Our results support stochastic accumulation over classic RP models (Libet et al., 1983; Matsuhashi & Hallett, 2008) or linear ballistic accumulation (Bogler et al., 2021). Surprisingly, all the models that were based on single-stage reporting—that is, on accumulator activity at the time of the probe—failed to recreate our empirical results. Instead, our results suggest that dual-stage reporting underlies reports of awareness of motor preparation. Hence the findings are reminiscent of work on perceptual metacognition, which suggests that post-decision evidence accumulation drives confidence reports (Desender et al., 2021; Moran et al., 2015). Dual-stage reporting could conceivably be implemented by a separate population of neurons that has the capacity to monitor the accumulator, but only does so when triggered by the probe. Notably, stochastic accumulation with dual-stage metacognition (which was simulated using 10,000 trials—many more than is feasible to collect empirically) suggests that there may indeed be a slight difference between pre-probe neural activity when comparing Prep with No Prep trials (see Fig. 4C). However, that difference is too small to be resolved using EEG, given the limited number of trials that can be recorded empirically together with EEG’s low signal-to-noise ratio. Furthermore, the dual-stage-reporting model, by design, predicts a difference in post-probe EEG slope when comparing Prep and No Prep trials (with positive and negative slopes immediately after probe onset in Prep and No Prep trials, respectively). However, no such difference was visible in our EEG data, potentially because it was masked by the substantially larger P3-like ERP component (Fig. 3A, D). The greater spatial resolution afforded by intracranial electrophysiology may be required to resolve these two limitations.
Schmidt et al. (2016) propose a general model of the RP (without defining it mathematically), wherein slow fluctuations in scalp EEG signal bias action timing. In their model, movements are more likely during the negative phase of infraslow cortical fluctuations (<0.1 Hz). This model is fundamentally different from stochastic accumulator models, which assume that the RP reflects an accumulation process that begins at trial onset (although the two could possibly be merged, if the ongoing slow fluctuations were fed into an accumulator—though investigating this was beyond the scope of the current study). Schmidt et al. (2016) hypothesized that infraslow fluctuations would be related to the ebb and flow of a subjective feeling of “readiness” to move. On one reading, our results provide evidence against that hypothesis: the lack of difference between pre-probe RP across Prep and No Prep trials suggests that the RP (whether it reflects infraslow cortical fluctuations or another phenomenon) does not reflect the subjective experience of already feeling readiness to move prior to being probed. However, on another reading, our results are compatible with Schmidt et al. (2016) hypothesis due to the specific phrasing we used in our instructions. We asked participants whether they had already begun preparing to move, which is an active process that involves major shifts in neural state space (Shenoy et al., 2011). However, Schmidt et al. (2016) specifically hypothesized that slow cortical fluctuations impact the feeling of readiness to move, which may be more of a passive process. Whether participants would treat these two questions the same or differently is an interesting open question for further study (although see Parés-Pujolràs et al., 2023).
Our findings have major implications for the validity of the probe method for timing intention onset. The probe method was originally introduced as an alternative to Libet’s clock method for timing intention onset (Matsuhashi & Hallett, 2008). “T-Time,” as it was called by Matsuhashi and Hallett (2008), was obtained by instructing participants to inhibit their movements in response to the probe if they were already preparing to move, and to ignore the probe otherwise. Hence, probes would result in vetoed movements when participants were preparing to move. So, the distribution of probe timings relative to movement onsets would have a “dip” in probe frequency at a certain time before movement corresponding to T-Time (see Matsuhashi & Hallett, 2008 for details). Using this method, T-Time (usually 1–2 s before movement onset) was far earlier than estimates of intention onset from the clock method (usually around 200 ms before movement onset). This may even overturn the classic interpretation that the RP emerges before conscious intention (Neafsey, 2021; though see Matsuhashi & Hallett, 2008). However, the validity of this method relies on two crucial assumptions. For T-Time to be an accurate measure of intention onset, it requires that the assessment of whether one is preparing to move can be made (1) instantly and (2) without affecting the movement generating process. If (2) is met but (1) is not, and assessing whether one is preparing to move takes, say, 300 ms, then T-Times would be biased to be 300 ms further from movement than the “true” timing of intention onset. If (2) is not met, the whole movement-generation process might be thrown off when participants report T-Time, making it less likely that T-Time captures the timing of a specific, meaningful event prior to movement onset, let alone specifically tracking the onset of intention. Furthermore, under the stochastic-accumulation interpretation of the RP at least, the process of accumulation over noise leaves no reason to expect events relevant to action initiation to reliably take place 1–2 s before movement onset. Indeed, under this interpretation, the only events accompanying action initiation are trial onset, when the accumulator process is initiated, and threshold crossing, shortly before movement onset. There was an extension of the stochastic-accumulator model that incorporated a sub-threshold value, whose crossing would set of an intention to move (Schurger, 2018). Nevertheless, while probe studies offer valuable insight into the nature of conscious intention, our results and the stochastic accumulator model suggest that care should be taken when designing and interpretating such probe studies.
Our study is limited by our assumption that responses to the probe are made based on a feature of the RP (a common assumption in probe paradigms; Matsuhashi & Hallett, 2008; Parés-Pujolràs et al., 2019; Schultze-Kraft et al., 2020; Verbaarschot et al., 2016, 2019; although see our discussion of beta power below). However, the critical requirement for our results is that the basis of reporting is weakly correlated with pre-probe RP buildup, which could also be satisfied if reports were based on a neural signal unrelated to the RP. Indeed, simulations showed that if trials were randomly classified as Prep or No Prep (as would be the case if they were based on a signal independent from the RP), it could lead to EEG results similar to those we observed under all the models we tested (Supplementary Fig. S3C, though as discussed there, random guessing is unlikely due to low rates of participants reporting being unsure of their preparation, probability of reported preparation increasing with trial duration, and the other differences between Prep and No Prep trials we observed). Our results could thus be stated as the following: either (a) reported awareness of motor preparation is based on a feature of the RP, in which case only stochastic accumulation can explain our ERP results, or (b) reported awareness of motor preparation is not based only on the RP but rather also on another neural signal, such as low-beta power in contralateral motor cortex (Parés-Pujolràs et al., 2023). Despite the field’s historical interest in the RP, when taking our findings together with those of Parés-Pujolràs et al. (2023), we consider the latter more likely. In either case, our study provides strong evidence against the influential conceptual model offered by Matsuhashi and Hallett (2008), wherein participants are initially “latently aware” of an intention to move, where the emergence of latent awareness is related to RP onset.
Recently, Parés-Pujolràs et al. (2023) suggested that beta oscillations, rather than the RP, relate to subjective reports of motor preparation. In line with their results, we found that pre-probe low-beta power over contralateral motor cortex was slightly but consistently higher for Prep trials than for No Prep trials (Fig. 5). Studies employing a stop-signal paradigm have found that beta-band activity is relevant for initiating and stopping actions (Mosher et al., 2021; Raud et al., 2020; Wagner et al., 2018; Wessel, 2020). Given that our study could be considered a modified stop-signal paradigm without a specific go-cue, the pre-probe differences in motor beta power we observed may indeed be related to motor-initiation processes.
Following the above, it is important to remember that probe studies aim to test participants’ insight into their motor-preparatory state on the fly, before they move. With that in mind, we should distinguish between conclusions that can be drawn from neural activity before versus after probe onset, for two reasons. First, post-probe activity might be related, at least in part, to the neural processing in response to the probe stimulus, rather than only to movement. Second, the probe likely interrupts and disrupts the ongoing motor preparation that would have counterfactually taken place had there been no probe. Now, we found no differences between Prep and No Prep trials for RP-like neural processes before probe onset (Fig. 3). In contrast, we found reliable pre-probe differences in low-beta power that has been associated with motor preparation, suggesting reliable differences in pre-probe motor activity (Fig. 5). This thus serves as an alternative explanation for our results, which does not depend on post-probe drift–diffusion fluctuations of the dual-stage model.
Importantly, the fact that pre-probe beta power differs between Prep and No Prep trials may suggest that participants are already aware that they are going to move when beta desynchronization begins. As such, pre-probe differences have categorically different implications than the post-probe differences we observed in the P3 ERP component (Fig. 3D) and that are suggested by the dual-stage reporting model we developed (Fig. 4). But it is also possible that the beta desynchronization we observed is related to motor preparation but not awareness—the state of motor preparation could be accessible upon being probed without that information being present in awareness prior to probe onset (note that this argument could also be applied to other probe studies). Thus, it remains unclear whether pre-probe beta power, post-probe stochastic accumulator activity, or another feature of neural activity altogether is the basis of our participants’ reports. This issue is compounded by the fact that the relationship between the RP and beta desynchronization is not well understood and requires further study.
Taken together, our results, therefore, raise three critical questions for future studies on volition: (1) are reports in probe paradigms made using a feature of the RP or beta desynchronization, (2) can beta desynchronization prior to self-initiated movement be explained by a stochastic accumulator-type process, and (3) how does beta desynchronization in the motor cortex relate to the RP? Although many studies have been conducted on the RP and beta desynchronization, few look at them simultaneously. Furthermore, although they have different cortical sources (Shibasaki & Hallett, 2006), the motor cortex and SMA are strongly interconnected, and it is unlikely the two phenomena are completely independent.
We observed that post-probe oscillatory power in the theta and beta frequency bands was related to reported awareness of motor preparation. Interestingly, oscillations in those frequency bands have been observed during metacognition in other contexts (Soutschek et al., 2021; Tang et al., 2021; Wokke et al., 2017, 2020). But our paradigm differs from metacognition paradigms where participants reported confidence about a perceptual stimulus (Moran et al., 2015) or the quality of a value-based decision (Brus et al., 2021). It is, therefore, unclear whether these oscillations relate to metacognition per se. Theta oscillations have also been associated with cognitive control (Cavanagh & Frank, 2014; Dippel et al., 2017; Pertermann et al., 2019), which may have been required to suppress movement when participants were already preparing to move—that is, in Prep trials in our paradigm—leading to the increased theta power on Prep trials that we observed. Notably, it has been proposed that midfrontal theta oscillations broadcast a need for cognitive control (Cavanagh & Frank, 2014), and then inter-region communication via beta-band synchrony underlies metacognition (Wokke et al., 2020). Our results largely align with these suggestions; although our study was not properly controlled to make inferences about what specific mechanisms these power modulations reflect. Future research on metacognition and motor preparation could use a modified version of our paradigm—for example, by including some trials where probes are delivered but participants are required to make a perceptual decision (such as whether the probe was of high or low pitch) rather than a metacognitive decision—to investigate that hypothesis more rigorously.
Our study was also limited in other ways. First, we required participants to inhibit their movement in response to the probe and only then report whether they were preparing to move. Hence, participants’ reports in our paradigm relied on working memory, which was part of the impetus behind probe studies (Matsuhashi & Hallett, 2008; Parés-Pujolràs et al., 2019). This design, therefore, may introduce memory-associated noise and biases to participants’ reports, which were at least partially removed from fully online probe methods. However, importantly, even in such online probe studies, decisions about one’s underlying state cannot be made immediately: at the very least, one must perceive the probe and then decide whether one was already preparing to move, thus introducing some delays and relying at least to some degree on working memory. Furthermore, in the present study, the timing of participants’ movements was arbitrary. Such actions are often habitual and could involve less overt awareness compared with actions made purposefully. Future research could increase the intentionality behind movements and increase participants’ attention during experiments, for example, by providing reasons to move at one time point or another and rewarding adherence to task demands.
As a whole, our study parsimoniously resolves ongoing debates and conflicting findings regarding what kind of process underlies the RP and how that process relates to the conscious intention to move. Our ERP results suggest that the RP is not directly related to the conscious experience of initiating action. But our computational modeling suggests that the neural process underlying the RP or beta desynchronization may nevertheless be accessible via metacognitive mechanisms. These results highlight the importance of investigating metacognitive mechanisms when studying conscious intentions. Consider that we are not constantly conscious of all our actions, decisions, and intentions when acting voluntarily—we can reach for a doorknob or pick up a cup of coffee without consciously representing, in advance or after the fact, all of the intentions that led to that action. Yet, we can introspectively access our intentions and report them if needed. We suggest that such metacognitive mechanisms are crucially involved in probe paradigms that target motor preparation and may be involved in accessing intention-related information. The version of the probe method that we developed thus offers an opportunity to study conscious intentions more generally, which is a long-standing goal of volition neuroscience.
Data and Code Availability
Data and code for reproducing main results are available at https://github.com/jgavenas42/ProbeMethodEEG.
Author Contributions
Jake Gavenas: conceptualization, methodology, software, formal analysis, investigation, visualization, writing—original draft; Aaron Schurger: methodology, resources, writing—review and editing, supervision; Uri Maoz: conceptualization, methodology, validation, resources, writing—review and editing, supervision, funding acquisition.
Funding
J.G., A.S., and U.M. were supported by the John Templeton Foundation and the Fetzer Institute (Consciousness and Free Will: A Joint Neuroscientific-Philosophical Investigation (John Templeton Foundation #61283; Fetzer Institute, Fetzer Memorial Trust #4189)). J.G. was supported by the NSF (BCS-2219800).
Declaration of Competing Interest
The authors have no competing interests to report.
Acknowledgements
We thank the anonymous reviewers of this manuscript for their insightful comments, which have substantially improved it.
Supplementary Materials
Supplementary material for this article is available with the online version here: https://doi.org/10.1162/imag_a_00465.
References
Author notes
J.G. moved from Chapman University to Cedars-Sinai Medical Center during revision of this manuscript