The time course of neural activity predictive of self-initiated movement

Reviewer #2 (Public review):

Summary:

The authors set out to investigate how well the onset of a self-initiated movement could be predicted at different times prior to action onset. To do so, they collected EEG and MEG data across 15 human participants who watched natural landscape images on a screen. These participants performed active self-initiated movements, or observed passive actions, to have a new image appear. By comparing the neural activity prior to active and time-matched passive actions, the authors found that even though a build-up of neural activity is visible close to 1s prior to action, action onset could only be reliably predicted around 100ms prior to action. These results confirm what was already suggested in previous literature: the commitment to action is only clear from the late stages in the visible neural ramp-up to action onset.

Strengths:

(1) The paper presents a well thought out methodology to assess the predictive value of neural activity prior to a self-initiated movement and passively observed action, while keeping all other experimental factors identical. This methodology can be applied outside the specific scope of this paper as well, in efforts to assess the correspondence of a neural signature with an observed behavior.

(2) The results are a strong confirmation of what was suggested less clearly in previous research (Trevena & Miller, 2010, Consciousness & Cognition; Schmidt et al., 2016, Neuroscience & Biobehavioral Reviews; Travers et al., 2020, NeuroImage).

Comments on revised version.

I thank the authors for addressing my concerns in the Introduction and Discussion of their paper. I don't require any further edits at this point.

Author response:

The following is the authors’ response to the original reviews.

Public Reviews:

Reviewer #1:

Summary:

Jeay-Bizot and colleagues investigate the neural correlates of the preparation of, and commitment to, a self-initiated motor action. In their introduction, they differentiate between theoretical proposals relating to the timing of such neural correlates relative to the time of a recorded motor action (e.g., a keypress). These are categorised into 'early' and 'late' timing accounts. The authors advocate for 'late' accounts based on several arguments that align well with contemporary models of decision-making in other domains (for example, evidence accumulation models applied to perceptual decisions). They also clearly describe prevalent methodological issues related to the measurement of event-related potentials (ERPs) and time-frequency power to gauge the timing of the commitment to making a motor action. These methodological insights are communicated clearly and denote potentially important limitations on the inferences that can be drawn from a large body of existing work.

To attempt to account for such methodological concerns, the authors devise an innovative experiment that includes an experimental condition whereby participants make a motor action (a right-hand keypress) to make an image disappear. They also include a condition whereby the stimulus presentation program automatically proceeds at a set time that is matched to the response timing in a previous trial. In this latter condition, no motor action is required by the participant. The authors then attempt to determine the times at which they can differentiate between these two conditions (motor action vs no motor action) based on EEG and MEG data, using event-related potential analyses, time-frequency analyses, and multivariate classifiers. They also apply analysis techniques based on comparing M/EEG amplitudes at different time windows (as used in previous work) to compare these results to those of their key analyses.

When using multivariate classifiers to discriminate between conditions, they observed very high classification performance at around -100ms from the time of the motor response or computer-initiated image transition, but lower classification performance and a lack of statistically significant effects across analyses for earlier time points. Based on this, they make the key claim that measured M/EEG responses at the earlier time points (i.e., earlier than around -100ms from the motor action) do not reliably correlate with the execution of a motor action (as opposed to no such action being prepared or made). This is argued to favour 'late' accounts of motor action commitment, aligning with the well-made theoretical arguments in favour of these accounts in the introduction. Although the exact time window related to 'late' accounts is not concretely specified, an effect that occurs around -100ms from response onset is assumed here to fall within that window.

Importantly, this claim relies on accepting the null hypothesis of zero effect for the time points preceding around -100ms based on a somewhat small sample of n=15 and some additional analyses of individual participant datasets. Although the authors argue that their classifiers are sensitive to detecting relevant effects, and the study appears well-powered to detect the (likely to be large magnitude) M/EEG signal differences occurring around the time of the response or computer-initiated image transition, there is no guarantee that the study is adequately sensitive to detect earlier differences in M/EEG signals. These earlier effects are likely to be more subtle and exhibit lower signal-to-noise ratios, but would still be relevant to the 'early' vs 'late' debate framed in the manuscript. This, along with some observed patterns in the data, may substantially reduce the confidence one may have in the key claim about the onset timing of M/EEG signal differences.

We thank the reviewer for this thorough summary and for providing the opportunity to clarify our claims in light of the concerns identified by this review. Our key claim we make is not that there are no early signals. Rather it is that the well-known and reliable signals (e.g. the RP) that could be construed as evidence for early commitments do not have the characteristics of early commitment signals (i.e. they are not predictive). Based on our main results (Figure 3) we can classify the presence of those early signals using traditional approaches rather well (time-based approach), but when we control for autocorrelation in the data (task-based approach), this classification performance goes away – indicating that those particular signals might not contain movement or intention specific information. Further, we do have evidence that our paradigm was sensitive enough to detect small differences between the two conditions as is evidenced in Appendix 1, Figure 12 with above chance AUC right after trial start, a period during which participants likely recall the trial type (Active vs. passive). We have now added individual peak AUC to the figure to illustrate that our approach was sensitive to these differences (See Figure S12).

We do however fully agree that absence of evidence is not evidence of absence. We have therefore effected the following changes to mitigate misinterpretations of our results: Abstract: “…it was not useful in predicting movement.” “ “Our results substantially constrain the interpretation that early EEG antecedents, including the RP, reflect a commitment to act, instead indicating that EEG-detectable evidence consistent with such a commitment emerges closer to movement onset, in line with the timing of conscious intention and the capacity for late movement inhibition.

Introduction:

“We show that, once appropriate controls are introduced, there is little evidence that cortical EEG contains reliable predictive information much earlier than ~100 ms before movement. Those results better align with late-decision accounts, which posit that the final neural commitment to initiate movement arises very close to the time of movement onset, in the context of self-initiated movements that are not pre-planned.”

“[…] and at odds with the presence of predictive early signals indicative of a commitment to move – from the vantage point of cortical activity.”

Discussion:

“Instead, the data better align with the more recent late-decision accounts”

“We show here that the data, once properly controlled for, do not support this early neural commitment.”

“Importantly, the significance of these findings lies not in whether premovement neural activity exists, but in what that activity can legitimately be interpreted to reflect. Our results suggest that although early ramping activity precedes self-initiated movement, this ramping activity may not reflect a true commitment to initiate movement. Instead, they may reflect broader preparatory neural dynamics that bias action likelihood without determining action outcome. Importantly, our findings do not establish whether the relationship between premovement activity, intention, and action is causal or merely correlational. Rather, they constrain interpretations that equate the presence of early neural antecedents with evidence that a specific voluntary decision has already been unconsciously determined one or more seconds before movement onset.”

Figure S12

Added individual peak AUCs post slide appearance

Caption: “Individual peaks after the appearance of the slide are displayed as black circular dots in the figure.”

Notably, there is some indication of above-chance (above 0.5 AUC) classification performance at time points earlier than -100ms from the response, as visible in Figure 3A for the task-based EEG analyses (EEG OC dataset, blue line). While this was not statistically significantly above chance for their n=15 sample, these results do not appear to be clear evidence in favour of a zero-effect null-hypothesis. In Figures 2A-B, there are also visible differences in the ERPs across conditions, from around the time that motor action-related components have been previously observed (around -500ms from the response). The plotted standard errors in the data are large enough to indicate that the study may not have been adequately powered to differentiate between the conditions.

We agree with the reviewer that our data does not support the strong claim that there are no early signals. In particular, we agree that the absence of statistically significant decoding does not demonstrate the absence of any underlying signal, especially for potentially subtle effects. However, our central claim is more specific. We do not argue that no early signals exist, rather, we argue that signals previously taken to be evidence of early commitment to move (like the RP) do not exhibit the characteristics expected of a cortical EEG signature of commitment: namely being predictive.

The slow ramping that we observed, despite its early onset, was not reliably predictive of upcoming movement. It is precisely the contrast between the early presence of ERPs without reliable early decoding that supports our main claim that those signals do not reflect commitment-like processes.

Classifiers specifically designed to pick up the clearly observable slow ramping activity (e.g. basic-slope LDA in Figure 5) or trained on a large search space of features (e.g. Haar-Adaboost) failed to produce strong or reliable early decoding.

Importantly, Haar-AdaBoost was not restricted to slow ramping activity or the RP. Its feature space included spatial, temporal, and time-frequency information, allowing it to detect other potentially informative EEG features beyond slow ramping activity, including changes in spectral power. Our key claim is not that no early neural signals exist. Rather, we found little reliable early predictive information across a broad range of EEG features. Our classifiers were sensitive to slow ramping activity (like the RP), spatial patterns, and changes in spectral power, yet robust discrimination emerged only close to movement onset. Thus, any earlier commitment signal would need either to differ from the RP and the other features accessible to our analyses, or to be largely invisible to scalp EEG and MEG. We have revised the abstract and selected sections to clarify that our findings concern early predictive information detectable by our classifiers in EEG more broadly, including but not limited to the RP. These revisions do not rule out an early neural commitment represented in signals inaccessible to EEG or MEG or to our classifiers.

Abstract:

“Although early ramping activity was present, our classifiers found little reliable predictive information in the EEG until about 100 ms before movement, when classification accuracy rose abruptly.”

“Our results substantially constrain the interpretation that early EEG antecedents, including the RP, reflect a commitment to act, instead indicating that EEG-detectable evidence consistent with such a commitment emerges closer to movement onset, in line with the timing of conscious intention and the capacity for late movement inhibition.”

Introduction:

“Here, we introduce a paradigm that tests whether early antecedents detectable in EEG—including the RP, but not limited to slow ramping activity—distinguish epochs culminating in self-initiated movement from well-matched epochs not culminating in movement, thereby assessing whether early EEG activity contains reliable outcome-specific information consistent with a commitment to act.”

“Thus, the analysis was not restricted to the RP nor to other slow potentials, but could exploit temporal, spatial, and spectral differences throughout the EEG signal.”

Results:

“Our results show that the slow buildup observed in movement-locked averages was not reliably predictive under more carefully controlled conditions (i.e. when classified against a control condition using the task-based approach). More broadly, classifiers sensitive to diverse temporal, spatial, and spectral features found little reliable early predictive information in the EEG.”

Discussion:

“Our results cannot rule out the possibility of an early neural commitment that is undetectable in EEG or MEG. Rather, neither the RP nor slow ramping nor other early EEG/MEG features accessible to our classifiers exhibited reliable commitment-like predictability. Any earlier commitment signal would therefore need to be distinct from the RP and either largely invisible to EEG/MEG or too weak, distributed, or otherwise inaccessible for the present decoding methods to recover.”

Further, we agree that our study may be underpowered to detect extremely small early effects. However, several observations suggest that any such effects are weak and unlikely to reflect a robust commitment. First, the slight early above-chance decoding visible in Figure 3A is primarily driven by a small number of participants and disappears following outlier exclusion (Figure S7). Second, even when present, early decoding remained extremely weak (approximately AUC ≈ 0.52–0.55), far below the level expected from a stable neural commitment to a specific action outcome. Third, when examining the earliest time points at which single trials became consistently and continuously classifiable until movement onset (EDT analysis), decoding emerged only very close to movement onset (~100 ms on average, with the most extreme participant remaining below ~200 ms). This pattern was additionally replicated in the now-expanded Libet dataset (N = 8; Figure S4), despite early onset of low-decoding performance at the group level, single trials EDTs or non-outlier group onsets were also very close to movement onset.

Importantly, these conclusions were robust across multiple classifier architectures, parameter choices, and preprocessing pipelines. We additionally reprocessed the EEG data using a 0.05–80 Hz band-pass filter, a 58–62 Hz notch filter, and down-sampling to 512 Hz, following the reviewer’s recommendations (below). Although this preprocessing marginally increased early above-chance decoding, the effect again depended primarily on a single outlier participant (participant 6), and single-trial EDTs remained below ~200 ms even for the outliers (Figure S14). Excluding this participant eliminated the apparent early decoding.

We therefore agree that early signals may exist. However, our findings suggest that the classical premovement ramping previously interpreted as evidence for an early decision does not behave like a commitment signal. At best, the early decoding observed in outliers could reflect weak biases or task compliance rather than early commitments to a specific course of action. Our conclusions do not depend on proving the complete absence of early information. Rather, they depend on showing that the early signals traditionally invoked in support of early-decision accounts lack the reliability and predictive strength that would be expected if a neural commitment to act were represented by a stable EEG signal.

In addition to the changes presented earlier, we have made the following changes to the manuscript to accommodate the above:

Introduction:

“Although very weak sustained above-chance classification was observed at early time points (Figure 3), these effects were driven by outliers (Figure S7). Crucially, this above-chance classification does not provide evidence for a commitment to initiate movement. Instead, it indicates that upcoming movements can be weakly discriminated from a matched control condition only in a small subset of participants and trials. Moreover, classifiers trained exclusively on slow ramping activity performed comparably—but not better—during the pre-movement period (Figure 5).”

Results:

“…we could decode only slightly better than chance…”

“…could only be classified reliably very close to movement onset…”

“Indeed, once outliers were excluded, classification remained at chance until very close to movement onset (Figure S7).”

“This single-trial EDT suggests that commitment, defined as sticking to a course of action, arises only late in time, close to movement onset.”

“This single-trial EDT suggests that commitment, defined as sticking to a course of action, arises only late in time, close to movement onset.”

“Additionally, while ramping activity is observable in multiple participants (Figure S3) it disappears when a high-pass filter with a 0.1 Hz cutoff is applied (Figure S14). This suggests that the effect may be driven by very slow fluctuations, slower than typical anticipatory signals, and may therefore reflect non-neural or artifactual processes.”

Discussion:

“, when using a classifier based solely on these slow ramping signals (basic slope LDA), performance also hovered around chance levels before movement, indicating that early ramping activity is not predictive of upcoming movements (Figure 5).”

“Some readers may note a slight above-chance performance in early time windows (Fig. 3). However, this effect was driven by outliers (Fig. S7). Even when present, such weak performance (e.g., AUC ≈ 0.53) falls far short of what would be expected from a reliable neural signature of commitment. Moreover, analyses of EDTs derived from correct classifications at the single-trial level show that correct predictions emerge and are sustained only very close to movement onset. This indicates that the underlying predictive signal itself arises only late. Together, these results suggest that the signal supporting early classification is not commitment-like.”

“In our primary analyses, we deliberately avoided filtering to preserve even the slowest fluctuations in the signal. However, this choice may have allowed low-frequency noise to remain, potentially obscuring smaller effects. To address this possibility, we re-analyzed the data using a 0.05–80 Hz band-pass filter, applied a band-stop filter to attenuate line noise, and down-sampled the data to 512 Hz (rather than 500 Hz, to minimize loss of frequency information). The filtered analysis (Fig. S14) revealed broadly similar patterns, with two notable differences: the slow drift in the passive ERP was no longer present, and group-level AUC onsets (both by-eye and 90\% criteria) appeared earlier. However, these early onsets were driven by a single outlier (participant 6; figure S14). When considering mean EDTs at the participant level, or when excluding this outlier, onset estimates remained late (figure S14) and consistent with our main findings (Figure 3).”

New figure in the appendix Figure S14

New figure with filtered data

Caption: “Decoding performance and MRCPs from filtered EEG data. EEG data were sampled at 512 Hz to avoid loss of frequency content during downsampling. Line noise was attenuated using a 58–62 Hz band-stop filter, and signals were further filtered using a 0.05–80 Hz one-pass zero-phase FIR filter (FieldTrip, firws) to reduce artifactual activity. (A) Grand-average movement-related cortical potentials (MRCPs) at electrode C3 for active (blue) and passive (red) trials, aligned to slide transition (t = 0; dotted gray line). Shaded areas indicate the standard error of the mean. MRCP onsets are marked by dashed gray lines. (B) Grand-average time course of validation AUC (10-fold cross-validation) for the task-based decoding approach (blue line). The x-axis reflects the leading edge of the sliding window relative to slide transition (t = 0; dotted gray line). Shaded areas indicate the standard error of the mean. AUC onsets are marked by dashed gray lines. Despite early apparent onsets at the group level (AUC onset-by-eye and 90% method > 1 s before movement), classification performance remains near chance and does not exceed ~0.55 until close to movement onset. At the participant level, both AUC onsets and single-trial EDTs average -0.09 s. Excluding participant 6 shifts group-level AUC onset estimates to -0.44 s (by eye) and -0.32 s (90% onset method), indicating that early onsets are driven by outliers.”

Although the authors acknowledge this limitation in the discussion section of their manuscript, their counter-argument is that the classifiers could reliably differentiate between conditions at time points very close to the motor response, and in the time-based analyses where substantive confounds are likely to be present, as demonstrated in a set of analyses. Based on this data, the authors imply that the study is sufficiently powered to detect effects across the range of time points used in the analyses. While it's commendable that these extra analyses were run, they do not provide convincing evidence that the study is necessarily sensitive to detecting more subtle effects that may occur at earlier time points. In other words, the ability of classifiers (or other analysis methods) to detect what are likely to be very prominent, large effects around the time of the motor response does not guarantee that such analyses will detect smaller magnitude effects at other time points.

In summary, the authors develop some very important lines of argument for why existing work may have misestimated the timing of neural signals that precede motor actions. This in itself is an important contribution to the field. However, their attempt to better estimate the timing of such signals is limited by a reliance on accepting the null hypothesis based on non-statistically significant results, and arguably a limited degree of sensitivity to detect subtle but meaningful effects.

Strengths:

This manuscript provides compelling reasons why existing studies may have misestimated the timing of the neural correlates of motor action preparation and execution. They provide additional analyses as evidence of the relevant confounds and provide simulations to back up their claims. This will be important to consider for many in the field. They also endeavoured to collect large numbers of trials per participant to also examine effects in individuals, which is commendable and arguably better aligned with contemporary theory (which pertains to how individuals make decisions to act, rather than groups of people).

The innovative control condition in their experiment may also be very useful for providing complementary evidence that can better characterise the neural correlates of motor action preparation and commitment. The method for matching image durations across active and passive conditions is particularly well thought-out and provides a nice control for a range of potential confounding factors.

Weaknesses:

There is a mismatch between the stated theoretical phenomenon of interest (commitment to making a motor action) and what is actually tested in the study (differences in neural responses when an action is prepared and made compared to when no action is required). The assumed link between these concepts could be made more explicit for readers, particularly because it is argued in the manuscript that neural correlates of motor action preparation are not necessarily correlates of motor action commitment.

We thank the reviewer for highlighting this important conceptual distinction. When proposing a matched control condition, the primary requirement was that the two conditions differ in the presence versus absence of a commitment to act. We argue that our two conditions (active vs. passive) satisfy this criterion. In the active condition, participants are instructed to execute a self-initiated motor action to advance from one slide to the next. In the passive condition, participants are not preparing or committing to any action during the slideshow.

We agree, however, that the two conditions are not matched in all other respects. Most notably, as pointed out by the reviewer, they also differ in the presence versus absence of movement itself as well as of accompanying movement preparation. As such, any differences observed between the conditions could in principle reflect motor-related processes, intention-related processes, or both. Importantly however, this ambiguity does not weaken our main conclusion. If early commitment-related signals were present and robust, they should still contribute to reliable discrimination between conditions. Yet despite the presence of visible premovement ramping activity in the averaged ERPs, classification remained weak and unreliable until very close to movement onset.

We have made the following changes to the manuscript accordingly:

Introduction:

“While the two conditions differ in the presence or absence of movement, they also differ in the presence or absence of an intention to move. As such, any early differences, if present, could reflect either motor-related or intention-related processes. However, this ambiguity is not a concern here, as we only observe differences at later time points.”

Discussion:

“A recent study (Derchi et al., 2023) contrasted volitional blinks with spontaneous blinks, thereby keeping the occurrence of movement constant while varying the presence of intention. Such an approach would avoid the movement-versus-no-movement contrast inherent to our paradigm. Importantly however, matching motor outcomes across conditions would likely make classification substantially more difficult, since motor-related information could no longer contribute to decoding performance.”

As mentioned in the summary, the main issue is the strong reliance on accepting the null hypothesis of no differences between motor action and computer initiation conditions based on a lack of statistically significant results from the modest (n=15) sample. Although a larger sample will increase measurement precision at the group level, there are some EEG data processing changes that could increase the signal-to-noise ratio of the analysed data and produce more precise estimates of effects, which may improve the ability to detect more subtle effects, or at least provide more confidence in the claims of null effects.

See our responses above regarding accepting the null. Also, we included an additional 5 participants to the Libet task (See updated Figure S4) to slightly improve our sample size. Although as previously stated in our manuscript, the number of trials matter more than the number of participants for a decoding study and with 1400 trials representing about 4 hours of data collection, should such early signals exist yet remain undetectable even with extensive trial counts and multiple decoding approaches, our results would suggest that EEG and MEG may have limited utility for studying them.

We have emphasized this point in our revised version of the manuscript:

Discussion:

“Some of our participants contributed up to 1400 trials, totaling up to 4 hours of data collection over 2 to 4 sessions in the EEG or MEG. Should such early signals exist yet remain undetectable even with extensive trial counts and multiple decoding approaches, our results would suggest that EEG and MEG may have limited utility for studying them.”

“[…] in eight additional participants and we found more or less identical results […]”

Figure S4:

Figure was updated to reflect an additional 5 participants

Caption updated to “AUC and ERP of Libet-task participants. Machine learning and ERP results for eight participants that we ran with a spontaneous movement initiation task based on Libet et al. (1983). (A) Readiness potential at Cz for the eight participants. In blue are the averages of the press trials and in red the averages of the no-press trials. Time 0 is the time of the trial end (dotted black line). Shaded areas are the standard errors of the mean. RP onsets using the 3 methods on press trials are reported and accordingly labeled (dashed grey lines) (B) Validation AUC for the eight participants using the task-based approach. Time 0 is the time of slide transition (dotted gray line). The AUC is aligned to the leading edge of the sliding window. Shaded areas are the standard errors of the AUC. AUC onsets using the three onset detection methods are reported and labeled accordingly (dashed grey lines). Our slideshow paradigm, while providing a more ecologically valid task, is agnostic as to the spontaneity of the participants’ movement decisions. To ensure that our task would generalize to more standard paradigms in the field of self-initiated action, we performed our analyses on data from a spontaneous voluntary movement paradigm based on (Libet et al., 1983). Here we collected eight further participants (4 female, 4 males, age M=24, 1 left-handed, 7 right-handed). They performed 350 trials of a task similar to the task performed by participants at OC except for the fact that there were no pictures, simply a fixation cross. Furthermore, the instructions for the manual trials were for participants to wait for a minimum of 3 seconds then start monitoring inwards for an urge to move. Whenever they detected such urge they were instructed to press as abruptly and spontaneously as possible, ending the trial. In the automatic trials, participants were instructed to do the same (monitor introspectively for an urge), except that they should not act on the urge if/when they felt it, but rather to wait for the next urge passively, and repeat such process until the trial ended automatically. This paradigm is much closer to the seminal studies on self-initiated actions (Libet et al., 1983; Kornhuber & Deecke., 1965). Here movements are performed spontaneously, and the matched condition (automatic) is identical in most regards apart from it not containing or terminating in a movement. All preprocessing and data analysis applied were the same. We found no strong qualitative difference with our main result (Figures 2A and 3A), suggesting that our task would generalize to other types of self-initiated actions.”

First, it is stated in the EEG acquisition and preprocessing section that the 64-channel Biosemi EEG data were recorded with a common average reference applied. Unless some non-standard acquisition software was used (of which we are not aware exists), Biosemi systems do not actually apply this reference at recording (it is for display purposes only, but often mistaken to be the actual reference applied). As stated in the Biosemi online documentation, a reference should be subsequently applied offline; otherwise, there is a substantial decrease in the signal-to-noise ratio of the EEG data, and a large portion of ambient alternating current noise is retained in the recordings. This can be easily fixed by applying a referencing scheme (e.g., the common average reference) offline as one of the first steps of data processing. If this was, in fact, done offline, it should be clearly communicated in the manuscript.

We did perform common average reference during the data preprocessing (lines 200-204 Slideshow2025/preprocess.m at main · lucasjeaybizot/Slideshow2025 · GitHub). We however clarified this description in the methods to clarify that this was performed offline:

Methods:

“We recorded EEG using a 64-channel BioSemi system keeping electrode offsets below 10 mV.”

The CAR was moved down for clarity.

“Data was re-referenced to the common average offline.”

In addition, the data is downsampled using a non-integer divisor of the original sampling rate (a 2,048 Hz dataset is downsampled to 500 Hz rather than 512 Hz). Downsampling using a non-integer divisor is not recommended and can lead to substantial artefacts in raw data as a result, as personally observed by this Reviewer in Biosemi data.

We had initially downsampled to 500 Hz to keep our analyses consistent from PF to OC. However, this is a very valid point, we have therefore re-analyzed the data using a downsampling to 512 Hz. Results did not change qualitatively nor quantitatively. Descriptions can now be found in the new supplementary Figure S14 and as referenced in the main manuscript (see changes below).

Discussion:

“In our primary analyses, we deliberately avoided filtering to preserve even the slowest fluctuations in the signal. However, this choice may have allowed low-frequency noise to remain, potentially obscuring smaller effects. To address this possibility, we re-analyzed the data using a 0.05–80 Hz band-pass filter, applied a band-stop filter to attenuate line noise, and downsampled the data to 512 Hz (rather than 500 Hz, to minimize loss of frequency information). The filtered analysis (Fig. S14) revealed broadly similar patterns, with two notable differences: the slow drift in the passive ERP was no longer present, and group-level AUC onsets (both by-eye and 90% criteria) appeared earlier. However, these early onsets were driven by a single outlier (participant 6; figure S14). When considering mean EDTs at the participant level, or when excluding this outlier, onset estimates remained late (figure S14) and consistent with our main findings (Figure 3).”

Finally, although a 30 Hz low-pass filter is applied for visualisation purposes of ERPs, no such filter is applied prior to analyses, and no method is used to account for alternating current noise that is likely to be in the data. As noted above, much of the alternating current noise will be retained when an offline reference is not applied, and this is likely to further degrade the quality of the data and reduce one's ability to identify subtle patterns in EEG signals. Changes in data processing to address these issues would likely lead to more precise estimates of EEG signals (and by extension differences across conditions).

In order to assess whether line noise or the absence of filtering in our main analyses could explain the absence of strong early decoding, we re-processed our data with a low pass filter at 80 Hz, a high pass filter at 0.01 Hz and a notch filter from 58 to 62 Hz. We further downsampled our data from 2048 Hz to 512 Hz to maximally retain spectral features. We then ran this preprocessed data through our algorithm. These changes are all comprised in new figure S14 and described above (see response to previous comment).

With regard to possible effects extending hundreds of milliseconds before the response, it would be helpful for the authors to more precisely clarify the time windows associated with 'early' and 'late' theories in this case.

We agree this is a very important clarification for the readers and as it stands is only implied in the literature. We now formalized it as follows, an early signal is a signal that would precede the time participants become aware of their intentions to move. In the case of self-initiated action, a recent meta-analysis suggests the lower bound of the 95% predictability intervals lands at -257 ms (Braun et al., 2021).

We have made the following modifications to the manuscript to reflect this definition:

Introduction:

“According to early-decision accounts, decisions (i.e., commitments to a course of action) occur early at the neural level, before participants typically report becoming aware of them. The lower bound of meta-analytic estimates of awareness, at ~250 ms before movement onset (Braun et al., 2021), can be used as a pragmatic threshold to classify neural signals as “early” (more than 250 ms before movement onset) or “late” (less than 250 ms before movement onset).”

“[…] well in advance, prior to the typical range during which participants report becoming aware of having decided (Braun et al., 2021).”

“[…] in time to movement onset, during the typical range during which participants report becoming aware of having decided (Braun et al., 2021).”

Discussion:

“Early-decision accounts posit that there is an early neural commitment to a decision that occurs prior to the subjective awareness of the decision.”

The EEG data that would be required to support 'early' theories is also not made sufficiently clear. For example, even quite early neural correlates of motor actions in this task (e.g., around -500ms from the response, or earlier) could still be taken as evidence for the 'late' theories if these correlates simply reflect the accumulation of evidence toward making a decision and associated motor action, as implied by the Leaky Stochastic Accumulator model described by the authors. In other words, even observations of neural correlates of motor action preparation that occur much earlier than the response would not constitute clear evidence against the 'late' account if this neural activity represents an antecedent to a decision and action (rather than commitment to the action), as the authors point out in the introduction.

We thank the reviewers for pointing out a set of two very pivotal assumptions in our work: early models predict that there are early commitment signals and that those signals both, (1) precede the subjective experience of the decision and, (2) are highly specific to the decisions’ outcome.

Assumption (1) rests on a set of findings that participants report being first aware of their decision up until very close to movement onset (Libet et al., 1983) which, according to a recent meta-analysis occurs up to around ~250 ms before movement onset (Braun et al., 2021). It is additionally supported by a large corpus of literature on movement inhibition: participants can inhibit their movement up until within ~225 ms from executing them (Vergbruggen et al, 2019). A similar time is also reported specifically for self-paced arbitrary movements (Schultze-Kraft et al., 2016). As such we take it that early-decision accounts, to satisfy (1), predict the early signal to be present more than ~250 ms before movement onset.

Assumption (2) relates to the conceptualization of decisions. Early-decision accounts posit that decisions are made early. A decision can be viewed as an act of settling on a plan of action. That is, unless new information interferes, the plan of action that was settled on by the act of deciding will be carried out. Maintaining this settled plan of action until it is carried out implies that once a decision has occurred, its accompanying plan of action must be maintained somewhere in the brain. In other words, there should be a brain signal that could in theory be used to predict with high accuracy an upcoming action. While ‘high accuracy’ is a vague concept, another prediction is that it should be equally useful over time to predict the decision, in other words, once a trial begins being correctly classified, it should remain so until movement onset.

In our paradigm, evidence for early theories would therefore be EDTs (Figure 3B) that occur more than 250 ms before the slide transition.

We have added the following section to clarify this in our manuscript:

Discussion:

“Evidence for early accounts would require EDTs that occur earlier than the ~250 ms lower boundary of meta-analytic estimates of the subjective awareness of the decision (Braun et al., 2021). In our data, no participants' EDTs (Figure 3B) occurred earlier than 250 ms before slide transition. While AUC onsets computed on the group data did occur earlier than 250 ms (Figure 3A), this was driven by outliers (Figure S7) and not present at the individual level.”

In addition, there is some discrepancy regarding the data that is used by the classifiers to differentiate between the conditions in the EEG data and the claims about the timing of neural responses that differentiate between conditions. Unless we reviewers are mistaken, the Sliding Window section of the methods states that the AUC scores in Figure 3 are based on windows of EEG data that extend from the plotted time point until 0.5 seconds into the past. In other words, an AUC value at -100ms from the response is based on classifiers applied to data ranging from -600 to -100 milliseconds relative to the response. In this case, the range of data used by the classifiers extends much earlier than the time points indicated by Figure 3, and it is difficult to know whether the data at these earlier time points may have contributed (even in subtle ways) to the success of the classifiers. This may undermine the claim that neural responses only become differentiable from around -100ms from response onset. The spans of these windows used for classification could be made more explicit in Figure 3, and classification windows that are narrower could be included in a subset of analyses to ensure that classifiers only using data in a narrow window around the response show the high degree of classification performance in the dataset. If we are mistaken, then perhaps these details could be clarified in the method and results sections.

The reviewer correctly describes the plotted data. However, it is important to note that because of two reasons (1 and 2 below), the concern of the onset of the AUC of the leading edge of the sliding window not reflecting the onset of the outcome-specific brain activity, can be substantially reduced. (1) The analysis we conducted is a sliding window, which means that an onset detected at -100 ms is above chance performance when classifying data taken from -600 to -100 ms but no such classification for the previous window from -620 to -120 ms. While this could mean that the entire window of activity was necessary for classification, it also suggests that only when incorporating data from the -120 ms to -100 ms window can the classifier perform above chance. (2) We used Haar features of different sub-window length combined with AdaBoost. AdaBoost selects the most informative features given the data, under this circumstance an increase in performance occurring only when including data from -120 ms to 100 ms indicates that data from that window was necessary for this increase in performance.

Additionally, to ensure more precision we now tested window of 250 ms and 100 ms (updated figure S11), the results of which align with what we previously report.

Discussion:

“Another potential concern is that predictive performance observed in wide sliding windows (e.g., -600 to -100 ms) could reflect information present as early as the beginning of the window. However, this interpretation is inconsistent with the temporal profile of decoding performance. For example, if the -620 to -120 ms window is non-predictive whereas the -600 to -100 ms window becomes predictive, the most parsimonious interpretation is that predictive information emerges near the newly included time points close to -100 ms, rather than being present throughout the entire window. This interpretation is further supported by the comparable performance observed across different window widths (figure S11), suggesting that the classifier was able to localize informative features within broader windows.”

Figure S11

Figure was updated to include analysis run with 250 ms window and 100 ms window Figure caption: “Evaluation of the sliding window widths. (A) Average validation AUC of Haar-Adaboost of the three participants of the PF EEG dataset using a sliding window (light blue) or growing window approach (pink) with the task-based approach. Time 0 is the time of slide transition (dotted gray line). The AUC is aligned to the leading edge of the sliding window. Shaded area is the standard error of the mean. A growing window analysis is similar to the sliding window analysis except that at each iteration, instead of sliding the window, the window width is increased by 0.02 s. This means that for an epoch beginning 3 s before movement, the window of analysis for the leading edge aligned to the slide transition would be 3 s wide from -3 s to 0 s (while for the sliding window method it would be 0.5 s wide from -0.5 s to 0 s). While it enables the capture of more patterns, it is a computation heavy analysis. To ensure that we were not missing out on some substantial longer-term patterns with sliding window, we ran a growing window on our PF EEG data. We found no improvement of using growing window over sliding window. (B) Average validation AUC of Haar-AdaBoost on the main OC EEG dataset using sliding windows of 100 ms (light blue), 250 ms (medium blue), and 500 ms (dark blue). Shaded areas represent the standard error of the mean. While clear classification differences emerged following movement onset, no substantial differences were observed prior to movement onset. This indicates that the 500 ms window was not too large to obscure finer temporal patterns. Furthermore, these results suggest that Haar-AdaBoost can identify localized informative features within larger windows, supporting the interpretation that the leading edge of the sliding window approximates the earliest time point at which predictive information becomes available.”

Reviewer #2:

Summary:

The authors set out to investigate how well the onset of a self-initiated movement could be predicted at different times prior to action onset. To do so, they collected EEG and MEG data across 15 human participants who watched natural landscape images on a screen. These participants performed active self-initiated movements or observed passive actions to have a new image appear. By comparing the neural activity prior to active and time-matched passive actions, the authors found that even though a build-up of neural activity is visible close to 1s prior to action, action onset could only be reliably predicted around 100ms prior to action. These results confirm what was already suggested in previous literature: the commitment to action is only clear from the late stages in the visible neural ramp-up to action onset.

Strengths:

(1) The paper presents a well-thought-out methodology to assess the predictive value of neural activity prior to a self-initiated movement and passively observed action, while keeping all other experimental factors identical. This methodology can be applied outside the specific scope of this paper as well, in efforts to assess the correspondence of a neural signature with an observed behavior.

(2) The results are a strong confirmation of what was suggested less clearly in previous research (Trevena & Miller, 2010, Consciousness & Cognition; Schmidt et al., 2016, Neuroscience & Biobehavioral Reviews; Travers et al., 2020, NeuroImage).

Weaknesses:

(1) Although the authors conducted a solid confirmatory study, the importance of this confirmation is less clear to me. How do the current results change our interpretation of the relation between conscious intention and neural preparation for action? Do these results affect our interpretation of free will? Why does it matter at all whether we see neural preparatory activity prior to the report of a conscious intention to act, or prior to action observation? This study does not clarify the relationship between the observed neural phenomenon, the action or the experienced intention. It does not explain whether this relation is causal, correlational or something else.

We thank the reviewer for their review. While it was previously suggested that the commitment to action comes late (Schultze-Kraft et al. 2016) and mechanisms were proposed to explain the early ramp-up in light of this late commitment (Schurger et al., 2012; Schmidt et al., 2016). Our main contribution is not merely confirmatory. Rather, our study introduces a matched control condition together with a classifier-based framework specifically designed to test whether early premovement activity is sufficiently predictive to qualify as a commitment-like signal.

Trevena & Miller 2010 differ largely from our approach, their control condition being matched for ‘decision’ but differing only in motor output. Their results are therefore still consistent with the RP being a decision-following signal. In contrast, our paradigm directly tests whether the early ramping activity reliably distinguishes epochs culminating in self-initiated movement from well-matched epochs not culminating in movement. Travers et al. 2020, while more directly testing the question of the specificity of the RP to self-initiated actions, provided ambiguous and contradictory interpretations, stating both that “[RP-like events] occur to the same degree in control data” and that “these events do not support the hypothesis that Readiness Potentials happen in the absence of action” with results allowing for multiple interpretations. Our study addresses this ambiguity directly by testing whether these early signals contain reliable outcome-specific information predictive of an upcoming action.

More broadly, while our findings are relevant to debates surrounding free will, our primary goal is neuroscientific rather than philosophical. Specifically, we address whether the signals traditionally interpreted as evidence for early decisions actually possess the properties expected of an early commitment signal. Our findings suggest that they do not.

Our data therefore invites a reinterpretation of the signals previously taken to challenge views of free will that require decisions to be made consciously. Specifically, the slow ramping signals traditionally interpreted as markers of decisions occurring very early in the brain do not exhibit the characteristics expected of a commitment-like signal. Rather than reflecting a stable and outcome-specific commitment to act, these early ramping signals may instead reflect broader preparatory dynamics that bias action likelihood without determining action outcome.

Our findings do not establish whether the relationship between premovement neural activity, intention, and action is causal or merely correlational. Addressing that question would require causal interventions beyond the scope of the present study. However, if early premovement activity were causally related to a settled commitment to act, one would expect it to contain strong and reliable predictive information regarding the upcoming action outcome. Our results do not support this account.

To clarify these points further, we have revised the manuscript accordingly and added the following text to the Introduction and Discussion sections.

Intoduction:

“Previous work showed that these early ramping signals were not specific to the outcome of a motor decision (Trevena & Miller, 2010). However, because their analyses remained time-locked to decision outcomes, they could not directly test whether the observed slow buildup specifically reflected a commitment-like process. Related work further suggested that RP-like patterns can emerge elsewhere in ongoing neural activity as features of spontaneous low-frequency fluctuations (Travers et al., 2020). Here, we introduce a paradigm that explicitly tests whether these early ramping signals distinguish epochs culminating in self-initiated movement, where a motor intention is present, from well-matched epochs not culminating in movement, where no motor intentions were present, thereby assessing whether they reliably reflect a commitment to act.”

Discussion:

“Our data therefore invites a reinterpretation of the signals previously taken to challenge views of free will that require decisions to be made consciously. Specifically, the slow ramping signals traditionally interpreted as markers of decisions occurring very early in the brain do not exhibit the characteristics expected of a commitment-like signal. Rather than reflecting a stable and outcome-specific commitment to act, this early antecedent ramp-like buildup may instead reflect broader preparatory dynamics that bias action likelihood without determining action outcome.”

(2) Whereas Derchi et al. (2023, Scientific Reports) were able to keep the entire experimental context similar across intended and unintended conditions, Jeay-Bizot et al. have one big difference between their passive and active conditions: the presence of a movement. Therefore, the present results explain the presence or absence of a movement rather than the presence or absence of an intention to act.

We thank the reviewer for pointing us to this very relevant and elegant paradigm introduced in Derchi et al., 2023. We agree that our active and passive conditions are not matched for movement itself, and therefore that any differences observed between these conditions could in principle reflect motor-related processes, intention-related processes, or both.

However, our primary goal was not to isolate a pure neural correlate of intention independent of movement, but rather to test whether the early premovement signals traditionally interpreted as markers of commitment contain reliable outcome-specific information predictive of an upcoming action. In that respect, the presence of an additional movement-related difference between conditions would be expected to increase, rather than decrease, decodability. Consequently, if reliable early commitment-like signals were present, they should still have contributed to reliable early classification performance in our paradigm.

More closely matching the motor output across conditions, as in Derchi et al. (2023), would likely make classification substantially more difficult by removing movement-related information from the decoding problem altogether. Importantly, despite movement differing between our two conditions, reliable decoding still emerged only very close to movement onset.

In addition to our modifications made with respect to a similar comment by reviewer #1 (see above), we have added the following paragraph:

Discussion:

“Our goal was not to isolate intention, but to test whether the signals traditionally interpreted as commitment signals actually possess commitment-like predictive properties. A recent study (Derchi et al., 2023) contrasted volitional eyeblinks with spontaneous eyeblinks, thereby keeping the occurrence of movement constant while varying the presence of intention. Such an approach would avoid the movement-versus-no-movement contrast inherent to our paradigm. However, matching motor outcomes across conditions would likely make classification substantially more difficult, since motor-related information could no longer contribute to decoding performance. More importantly, our paradigm is focused on the neural antecedents that are specific to whether or not a movement is initiated. This is a fundamentally different question from Derchi et al, which is focused on whether or not an intention is present. The two questions are complementary, however, and taken together, our study and the Derchi study contribute significantly to our understanding of spontaneous voluntary action.”