Can neural activity reliably predict free decisions?

Can neural activity reliably predict free decisions?

24th _{February 2016}

Alastair Haigh 10531998

Supervisor: Yair Pinto Co-assessor: Marte Otten

MSc Brain & Cognitive Sciences, Cognitive Neuroscience Track University of Amsterdam

Abstract

Libet’s experiments of the early 1980s brought the question of free will within the remit of neuroscience. Libet seemingly showed that during spontaneous movement, the readiness potential, (RP; movement-related brain activity), begins before a person becomes aware of wanting to move, thus challenging our intuitive sense of conscious control. This thesis examines three decades’ work since Libet. Notably, the association between the RP and spontaneous movement has loosened, calling Libet’s work’s significance into question. However, recent fMRI research has revealed apparently movement-related activity patterns occurring seconds before conscious awareness. As technology improves, our ability to predict behaviour may follow suit. Consciousness’s executive role therefore remains uncertain, which for some theoreticians is irrelevant to the question of free will whereas for others it is vital. These incommensurate definitions of free will, as well as the vagueness of ‘freedom’ as a scientific concept, prompt the conclusion that the term is best abandoned.

1. Introduction: Libet’s original experiment and its interpretation

What does it mean to act spontaneously? If you suddenly decide to look up from what you are doing, your experience is of a deliberate and instantaneous act. You did not sit there pondering whether to move, you simply moved. What are the neural underpinnings of such behaviour? Is your brain as calm as you are before you move? Answers to such questions became possible following the first

description, by Kornhuber and Deecke (1965) of a pattern of electrical activity, averaged over multiple electroencephalogram (EEG) recordings, which reliably preceded voluntary, spontaneous movements. They termed this slow, negative shift the Bereitschaftspotential, or readiness potential (RP, Figure 1. See also Box 1). The movements which this potential preceded were of the sort we might associate with the concept of free will: as unencumbered as possible by outside cues, influences or pressures, giving us a sensation of maximum conscious control. Philosophers have been tackling the question of free will since at least the time of Plato and Aristotle (Cary, 2007), but it was not until the late 20th

Century that the problem became not just one of philosophy but of neuroscience as well. Benjamin Libet (Libet, Gleason, Wright & Pearl, 1983) inaugurated this transition by asking a simple

question: Does the RP extend further back in time than our sensation of conscious control? If it does, then might this provide the first scientific evidence that in a mechanistic universe, the seductive feeling that we are in control, that our rich sensation of consciousness is the driving force behind our actions, is in fact an illusion; in short, that free will as it is commonly understood, does not exist?

The experiment was straightforward. Subjects sat facing a screen displaying a clock face. A spot of light moved around the clock and subjects were simply asked, after one full rotation, to make a quick, abrupt movement of the wrist. In some trials they were asked to remember the spot’s location at the moment they moved (reported as the variable M), and on others to remember its location at the moment they felt the urge to move (variable W). These variables were controlled for in a further set of trials in which subjects reported the time they felt an electrical stimulus on their skin (variable S). The difference between S and the actual time of stimulation was then subtracted from the other variables W and M, thereby, according to Libet, taking mental processing time into account (Libet et al., 1983).

Figure 1: Example RP averaged over 40 trials for one subject. Vertical line represents movement onset. Two traces taken from contralateral, precentrally located electrodes. (From Libet et al., 1982)

The results showed mean W to be around 200 ms before movement onset. However, RP onset came reliably earlier, around 1000 ms before movement (Figure 2). Brain activity was apparently preceding not just the movement, but the reported time of subjects’ conscious intention to move. Libet therefore concluded that the brain ‘decides’ it is going to make a movement before the subject becomes aware of the ‘decision’. Is consciousness then relegated to the role of a mere observer with delusions of control; a nominal leader with the real decisions being made behind the scenes? Libet felt his results made it clear that we do not consciously initiate decisions. However, he saw a way to reconcile this apparent fact with a kind of free will. He saw that there is enough time—around 200 ms—between the conscious feeling of wanting to move (W) and the

movement itself to allow the brain to ‘veto’ the decision in the full light of consciousness (Libet et al., 1983).

Figure 2: Timeline from onset of first and second stage of RP (RP1, RP2), time of urge to move (W), subjective time of movement (M), subjective time of skin stimulus (S) and actual movement onset (time zero). (From Hallet, 2007, using data from Libet et al., 1983)

Libet’s results have since become famous, and every aspect of his seminal experiments has been and continues to be the subject of wide-ranging debate. The results themselves are perhaps the only part of the work that are relatively uncontroversial, having been replicated many times since; but the methods, and in particular Libet’s interpretation of his results, remain a hot topic at the

interface of neuroscience and philosophy. Responses have ranged from specific methodological criticisms, to redefinitions of the concept of free will itself in light of the results. This thesis will examine in detail methodological and conceptual questions arising from Libet’s work and how researchers and theoreticians have addressed these questions; and how, in doing so, have advanced the field of free will research begun by Libet over three decades go.

Methodological issues discussed include: the effects of data processing on the results—does the averaging of EEG waveforms give a false impression of RP onset, as well as obscuring important individual differences? How reliable is the timing of subjective events? Does the clock, or do other aspects of the

experiment, introduce systematic biases? Was Libet right in claiming that it is possible to veto a movement without accompanying unconscious neural activity? Conceptual questions include the nature of the RP itself: how movement-specific is it, if it is specific to movement at all? Can it be dissociated from spontaneous movement? Does awareness of wanting to move appear instantaneously or gradually? Is activity preceding movement random? Advances in technology since Libet may have opened up new avenues in this area of research. Recent studies have shown it is possible to obtain information predictive of movement from patterns of brain activity several seconds before conscious awareness. It also now appears to be possible to predict movements in real time. The implications of these findings and possible future developments will be

discussed. The research covered will then be put into diverse theoretical context. For some, these results demonstrate that the sort of free will we may feel we possess must be an illusion. For others, even if decisions begin unconsciously, we are still the authors of our own destiny and free will, in that sense, remains unscathed. Still others insist that conscious control is as real as it feels, and that Libet’s work has been misinterpreted. Finally, in light of ongoing confusion surrounding the term ‘free will’, both academically and publicly, it will be suggested that the term be abandoned as unfit for scientific purpose.

Box 1: Glossary

Readiness Potential (RP): A slowly rising negative event-related potential thought to occur before a spontaneous movement, onset occurring around 1 s before movement onset. It is strongest at central EEG electrodes. Detection typically requires the averaging of waveforms from many trials, time-locked to the movement. Libet and colleagues (1982) classified the RP into two (or three) distinct types. Type I is a gradual, ramp-like increase in negativity associated with movements made at pre-set times. Types II and III are associated with spontaneous movement and seem to have two distinct components: a slow rise beginning around 1 s before movement, followed by a more rapid shift around 575 ms before movement (see Figure 1 in main text).

Laterlised Readiness Potential (LRP): A late component of the RP originating in the motor cortex contralateral to the hand which is about to move (Eimer 1998).

Libet’s variables: Each of these variables were reported by subjects after they had noted the time of their occurrence using the clock face visible to them during the experiment.

W: The time a subject feels the urge to make a spontaneous

movement but has not yet moved.

M: The moment the subject first feels herself make a spontaneous

movement.

S: The moment the subject feels a skin stimulus. This was used to

control for processing time in an attempt to produce more accurate values for W and M for each subject. The difference between S and the actual time of the stimulus is subtracted from

2. Methodological issues arising from Libet’s experiments and their treatment in subsequent research

2.1 The effect of data processing, and other technical aspects of the experiments on results and their interpretation

In much work involving brain potentials, averaging is a necessary evil: the signal-to-noise ratio is improved at the cost of per-trial, or even per-subject

information. If events occur at different times in different trials then that information is lost when EEG waveforms are averaged. Trevena and Miller (2002) noted this drawback in studying the RP and its relationship with voluntary movement. They pointed out that when waveforms are averaged across trials and/or subjects, a smearing artefact may bias the result: not only would the variability of RP onset times across trials be obscured, but the

averaging process would push the mean onset time back to a misleadingly early point. In an attempt to mitigate this, they conducted a Libet-style experiment with the difference that subjects were cued to move their left or right hand but were free to choose when to do so, or not at all. Both hands were used so that the onset of the LRP could also be measured and compared with W times (the

moment subjects feel the urge to move). Crucially, they compared the earliest W times to RP and LRP onset to offset the bias caused by smearing which gives unrepresentatively early RP and LRP onset times. They found a mean RP onset time of -1300 ms and LRP onset time of -300 ms. Their earliest W times began around -400 ms, occurring after RP onset, as Libet found but before LRP onset. However, these early times only represented 20% of trials, the other 80% of W times coming after LRP onset. Nonetheless, they “tentatively” concluded with the speculation that in fact LRP occurs after W on every trial, but that this is being obscured by the unavoidable smearing artefact artificially pushing the LRP onset time back. If this speculation is true, it puts Libet’s conclusions in doubt, for although RP reliably precedes W, the authors attest that this represents only “anticipation” of movement whereas the LRP represents specific preparation, which might begin only after the time of the conscious decision.

How much difference do technical details of experimental setup make to results? Verbaarschot, Farquhar and Haselager (2015) addressed this question, asking how two factors affects RP onset time: the choice of which electrodes to record from, and the method of determining RP onset time. Their subjects performed a Libet-style task and for each subject three sets of data were processed:

recordings from only the Cz electrode; recordings from Cz, C3 and C4; and recordings from the three electrodes that produced the most pronounced RP for each subject. For each of these three sets of data, they calculated RP onset by three methods: visual inspection (the method used by Libet et al (1983)); a criteria-based form of visual inspection; and a statistical method using a t-test. They found large differences in RP onset time and W between subjects. Mean RP onset also varied strikingly, between -881 ms and -254 ms depending on the combination of electrodes recorded from and onset evaluation method used. They concluded firstly that technical details in experimental setup can make large differences to the outcome and therefore to its interpretation; and secondly that data averaged between subjects should not be used to infer causal

relationships between the RP and conscious time of intention: subjects should be assessed on a case-by-case basis. Moreover, they concluded that great care should be taken when comparing data between different studies. It should be noted, however, that not all details of the experimental setup have significant effects on the results. A battery of tests was carried out in which variables such as the clock’s rotation speed, whether subjects should fixate on the clock’s centre, and the exact instructions given to subjects, were manipulated, but results were not significantly affected (Pocket & Miller, 2007).

Taken together, these studies most of all demonstrate a need for methodological consistency in research of this type: differences in results in the range of

hundreds of milliseconds can lead to very different interpretations in terms of causality and hence conclusions about the nature of spontaneous decisions and of free will. However, the issue of timing brings with it a question quite different from how equipment is used and results analysed: that events being timed are subjective.

2.2 W: The problem of timing subjective events

One of the major difficulties with researching free will, and perhaps the reason why a complete picture of the subject remains elusive, is how to reliably measure something that remains, for the present at least, necessarily subjective; and how to do so with millisecond precision. Libet et al., 1983 attempted to control for any biases in their subjects by measuring the time difference between delivery of a skin stimulus and its subjective report (their variable S). However since then, other problems with the timing of subjective events have been noted.

Libet received immediate criticism on the subject of timing subjective events in the open peer commentary following his 1985 paper. Breitmeyer for example questioned the use of a clock as a timing method, and with using S, (times reported by subjects that they felt a skin stimulus), as a control: processing speeds of visual and tactile stimuli are not be the same, and internal and external attention are similarly not equivalent. Rollman also took issue with the validity of the clock, suggesting that Libet was mistaken in equating times reported as W with that actual time of awareness. He proposed a time N between light from the clock hitting the retina, and the brain becoming aware of the clock’s position. Libet’s W should therefore be N seconds longer than it is. Estimating a value for

N in the hundreds of milliseconds leads to the conclusion that conscious

awareness of clock position may take place after movement. The same argument could apply to M, which would push it from before to after movement as well. He also criticised the use of S as a control, arguing that the things measured by S and

M—external stimulus and internal state respectively—are not equivalent

because their latencies are different and therefore they should not be compared. Might the increased activity of brain areas involved with the task during Libet’s experiments produce biases in judgment of timing? Lau, Rogers, Haggard and Passingham (2004) had previously found an association between “attention to intention” (i.e. watching out for the urge to move, W) and enhanced activity in the pre-supplementary motor area (pre-SMA). They then looked for a similar

association between heightened activity in the cingulate motor area and estimation of movement onset (M) (Lau, Rogers & Passingham, 2006).

Specifically, they wanted to know whether such heightened activity correlated (positively or negatively) with accuracy of judgment of M, (that is, do we see greater accuracy with increased activity, or greater bias?) They found that degree of neuronal enhancement did correlate, negatively, with subjects’ accuracy—the greater the enhancement the earlier the perception (Lau et al., 2006). Assuming the same bias was affecting subjects’ judgments of onset of intention as had been affecting onset of movement, they concluded that average W times should be corrected from -228 ms to -120 ms; the difference being the bias associated with the act of attending to their movements. In a follow-up study, the same research group (Lau, Rogers & Passingham, 2007) used transcranial magnetic stimulation (TMS) to manipulate brain activity around the moment subjects chose to move in a Libet-style task. When TMS was applied to the SMA at the moment subjects moved, or 200 ms afterwards, their judgments of intention onset (W) were shifted back in time, and those of movement onset (M) were shifted forward (Figure 3). That experimental manipulation after an event could influence judgment of its timing was taken as strong evidence that perception is constructed and lags behind reality (Libet et al., 1964).

Figure 3: Applying TMS to SMA at movement onset and 200 ms afterward shifts perception of movement forward in time (white bars) and perception of intention back in time (grey bars). (From Lau et al., 2007)

If TMS can shift W subjectively in time, is the feeling of intention determined by endogenous neural events, external cues or a combination of both? Banks and Isham (2009) asserted that W “is not uniquely determined by any generator of the RP”; that instead it is based on several cues, with the “apparent time of response” being the most important. To demonstrate this they carried out an experiment to test the hypothesis of Eagleman (2004) that judgment of one’s own intention depends on an action’s immediate feedback. If this hypothesis were correct then delayed feedback would systematically affect W judgments. Their subjects carried out a Libet-style task in which a beep played between 5

and 60 ms after a button was pressed while the subjects watched a clock and reported their time of intention to move. They found that longer delays were associated with later W times, although there was not a perfect correspondence. These results therefore provided evidence that judgment of intention is at least partially inferred from external feedback, and thus not entirely what it purports to be: a report of an individual’s thoughts as they happen.

Does the act of watching the clock affect results? Miller, Shepherdson and

Trevena (2011) addressed this point by carrying out an experiment to explicitly test whether the presence of the analogue clock contributed to, or indeed entirely caused, the RP. Their subjects performed a standard Libet-style task, reporting W and M from the clock. They also performed the task without the clock and without reporting anything. They found an RP-like negative slope in the clock condition but no slope in the no-clock condition. If it is indeed the act of clock-watching which produced this slope, the same result should be found if subjects do not press a button. They tested this in a second experiment in which subjects were instructed to listen out for a tone and report whether it was low or high pitched. In one condition subjects watched a clock and reported the time of the tone, in the other condition they did not. They found a negative slope in the clock condition and no slope when the clock was absent. They concluded that the clock seemed to be responsible for the rising negativity interpreted by Libet (Libet et al., 1983) as unconscious movement preparation, and so that

interpretation should be re-evaluated. Banks and Isham (2011) also tested the effect of the standard Libet-type clock, but rather than testing against nothing, did so against two digital alternatives, one which displayed numbers in sequence and another which did so in a random order. Subjects were asked to make a movement at a time of their choosing and to report either W or M. The Libet clock gave an M time of around 45 ms before the button press, in line with previous results. However, the randomised digital clock produced a value of M not significantly different from zero (i.e. the time of button press) and the sequential digital clock produced a value of M significantly after button press, around 60 ms (Figure 4). They conclude that the differences in M imply that the difference between M and button-press normally recorded cannot be due to the motor system (otherwise the three results here would not have been different from each other); rather that the difference is due to perception of the clock. Values of W also differed per clock, but not in the same way: in this instance the Libet clock gave a value of around -140 ms—in the normal range; but the

sequential digital clock was closer to the time of button-press at -30 ms, and the randomised digital clock gave the earliest value of W, at -385 ms (Figure 5). The researchers commented on these results saying that there is no way of knowing which of them is “correct”, and that “correct” may not in fact be a meaningful term.

Figure 4: Times of action reported by subjects using three different types of clock. Time 0 is the time the movement actually took place. (From Banks & Isham, 2011)

Figure 5: Times of decision reported by subjects using three different types of clock. Time 0 is the time the movement actually took place. (From Banks & Isham, 2011)

In a comprehensive critical assessment of Libet’s work, Gomes (1998) made a number of comments on the subject of timing subjective events. Firstly, he questioned whether W and M are as easily distinguishable as Libet originally claimed and as others replicating his work have apparently accepted. Remarking that he personally found the task of judging W to be difficult and unnatural; he suggested that the fact that Libet’s subjects said they found no difficulty in the task might have been at least partly down to their having understood the instructions and wanting to please the experimenter. He then posited an

it to a problem inherent in introspection, namely that expectation can affect results. When watching out for the urge to move (W) subjects may have expected it to come earlier than (M) and thus (consciously or not) adjusted their reports accordingly. Gomes also questioned Libet’s tactic of “training” subjects by telling them how close their reports of M and of S (timing of a stimulus to the skin) came to the actual onset times. This may have had the effect of subjects attempting to correct for their own biases to make their reported times as accurate as possible (even without being asked to). Given that the stated purpose of recording S times was to correct for systematic bias, allowing subjects the opportunity to correct this themselves would be counter-productive. Finally, Gomes pointed out that Libet’s idea of S as providing information to correct M and W times depends on the veracity of his hypothesis of the referral of conscious sensory experience subjectively backwards in time (Libet et al., 1979). Having provided evidence against this hypothesis, Gomes argued that latencies must now be taken into account (that is, the time between onset and conscious awareness). Because these may differ depending on stimulus modality (e.g. visual vs. cutaneous) and complexity, and because they are unknown, reported times of subjective

experience can no longer be considered reliable.

The task of timing subjective events is replete with methodological difficulties. That something as fundamental to Libet’s original experiments as the clock has been found to exert significant influence over results is worrying. Moreover, that such findings are relatively recent shows that, despite being over 30 years old and many times replicated, Libet’s method is far from established and settled. The surprising finding of the clock’s influence, together with the apparent susceptibility of W judgments to external manipulation, as well as questions about their possible inherent unreliability, underscores the difficulty of timing subjective events. Banks and Isham (2011) may indeed have been right in saying that there is no way of knowing whether a given value of W is “correct”. All may not be lost however. As we shall see later, another method—that of predicting behaviour from brain signals—is being brought to bear on the question of

conscious causation. Meanwhile, evidence has arisen to counter Libet’s assertion of a consciously controlled veto.

2.3 Does the veto allow for “free won’t”?

Although Libet (1985) interpreted his results as showing that the process of making a voluntary movement begins before we are aware of it, he did not, as previously mentioned, abandon the possibility of conscious control entirely. Because W—the time subjects felt the urge to move—occurred in advance of the actual movement, Libet believed this left room for actions to be consciously vetoed; so-called “free won’t”. The executive role of consciousness, then, was not to initiate acts, but rather to oversee them as they unconsciously arise, selecting which will or will not go ahead. New evidence has come to light, however, that challenges this conclusion.

In his explanation of the veto, Libet was making two claims. Firstly that the mechanism by which it occurred was not reducible to measurable brain activity; and secondly that the mechanism was driven by consciousness. Work by Brass

and Haggard (2007) calls the first of these claims into question. Noting that previous work had shown frontal areas to be involved in inhibiting stimulus responses, they asked whether similar areas might be involved in inhibiting endogenous action as well. They carried out an experiment with three conditions. The first was a standard Libet-type task, under fMRI, in which subjects watched a clock and reported the time they first felt the intention to press a button. The second condition was identical except subjects, if they wished, could cancel the button-press as late as possible. In the third, control, condition, subjects reported the clock position when a tone was played. A contrast of inhibition against action trials showed activation of frontal and temporal areas. Both inhibition and action trials showed activation of pre-SMA and SMA compared with tone trials, showing that there was indeed preparation for movement in inhibition trials. Further, increased frontal activity correlated with a greater probability of inhibition and well as decreased primary motor cortex activation. The results demonstrated that inhibition seems to involve different areas from action, and that these areas are located more anteriorly than those involved with action, supporting the idea of an anterior-posterior

functional gradient with higher-order intentions located in more anterior frontal regions (Taren et al., 2011). The researchers concluded that their results are evidence of a “neural correlate of the veto process”, rendering Libet’s (1985) hypothesis of a mysterious extra-neural veto mechanism redundant. Further work has recently strengthened this view. In an EEG study, cues told subjects to react either immediately or with a slight delay. These trials were contrasted with others in which subjects could choose whether to delay following a cue. There was greater pre-stimulus activity on choice trials with no delay than those with delay; a difference that did not appear when subjects were cued whether or not to delay. The researchers concluded that whether or not subjects chose to delay was dependent on pre-conscious activity, and that the choice therefore was not made consciously (Filevich, Kühn & Haggard, 2013). These two studies have provided evidence against Libet’s twofold claim about the veto: that it is directly caused by conscious will and that it operates by some undetectable non-neural mechanism. They have also shown that decisions to inhibit action involve different neural circuitry than decisions to execute actions; as far as the brain is concerned, initiation is not the same as inhibition. As well as the methodological issues discussed here, a number of conceptual and theoretical questions also naturally arise from research of this kind.

3. Conceptual issues arising from Libet’s results, their interpretation and subsequent work

3.1 What is the RP?

Central to the original experiments of Libet and their many subsequent replications, elaborations and modifications is the readiness potential (RP), a gradually rising negative shift that typically begins around 1 s before a voluntary movement is made (see Box 1). Interpretation of the RP has therefore been an important part of the ongoing debate over how to make sense of the results of these experiments. Research since Libet has begun to cast doubt on any close association between the RP and spontaneous movement, and therefore on Libet’s

conclusions. The movement-specificity of the RP has been questioned, as has its causal relationship with the conscious feelings of intention and movement. Some results also suggest that it can occur in the absence of movement and that

movement can occur without an RP.

It is important to highlight the difference between the RP and its late component the LRP (see box 1). As mentioned in section 2.1, Trevena and Miller (2002) found suggestive though not conclusive evidence that although the RP reliably precedes conscious decisions, the LRP may always occur after them. If the two were causally linked, this would suggest the conscious decision causes the LRP. Haggard and Eimer (1999) compared the RP with the LRP as possible causes of the feeling of conscious intention (W). Noting that a characteristic feature of causal relationships is covariance, they found a covariant relationship between the LRP and W but no such relation between the RP and W. The researchers however, took this as evidence that the LRP is the unconscious cause of W, and not the other way around. But in their study not all subjects’ W times came after LRP onset. Thus, although the relationship between the LRP and W remains rather unclear, particularly in terms of causation, it seems unlikely that the RP as a whole is causal. Moreover, further work has suggested the RP does not reflect preparation to move one hand over the other: the same preparatory activity was found using MEG prior to a cue to move either the left or right hand (Herrmann, Pauen, Min, Busch & Rieger, 2008). If the RP as a whole represents only vague, non-movement-specific preparation, this undermines Libet’s claim that activity preparatory to specific movements occurs before we are aware of it. Might the RP, in fact, be unrelated to, and therefore dissociable from, movement entirely? The possible dissociation between RP and movement was addressed by Trevena and Miller (2010) who reasoned that if RPs represented motor preparation, then they should have a higher amplitude when movements are made than when they are not. They used a cued version of Libet’s experimental paradigm with the added instruction that subjects should make a spontaneous movement in only about half of the trials and no movement in the rest. They found an RP-like signal whether or not subjects moved, and concluded that there was no evidence to support the hypothesis that greater negativity (i.e. a higher-amplitude RP) preceded movement than non-movement. This work has since been criticied however. Alexander and collegues (2016) noted that the task was cued, so the timing of the movements was not spontaneous, only the decision whether or not to move was. Furthermore, after closely examining the data they note the “RP”s reported are of very low amplitude (around -2 µV, compared with around -10 µV in the original experiments of Libet et al., 1983; see Figure 1). These researchers therefore attempted to address these shortcomings by using a modified Libet design in which subjects, at a time of their choosing, either made a hand

movement or chose one of four letters in the centre of the clock face, noting the time of their decision in both cases. Comparing these “movement+decision” and “decision-only” conditions, they found no significant difference in RP amplitude. They concluded that the RP is mostly, if not entirely, non-motor-related; rather that it reflects activity related to spontaneous decisions generally (Alexander et al., 2016).

If RPs can occur without movement, might the opposite also be true: movement without an RP? Indeed, in the commentary following Libet (1985), Rugg asked whether there might be brain-damaged patients for whom this is true; that they can initiate movement without producing an RP. Sirigu and colleagues (2004) were the first to test this proposition. They recruited patients with lesions in parietal cortex, in the cerebellum, and healthy controls. Using a standard Libet task and recording with EEG, they found the gradual rise in negativity

characteristic of the RP was present in cerebellar patients and control subjects, but was either absent or of much lower amplitude in parietal patients. Thus, whatever the RP indexes, it seems that if it is compromised or even eliminated through brain damage, spontaneous movement is still possible. This experiment produced another result: the W and M times of parietal patients were the same, whereas in cerebral and control patients they were separated by around 250 ms. In other words, parietal patients are apparently unable to distinguish between the feelings of wanting to move and of actually moving. The ability to report W is therefore seemingly unrelated to the ability to make spontaneous movements. This further weakens the validity of W as a reliable measure of a known subjective event (see section 2.1).

There is, then, mounting evidence casting doubt on the RP as a reliable correlate for a pre-conscious intention to move. This is not to say that the RP is completely unrelated to movement, nor that Libet’s results are meaningless and his

conclusions unwarranted. However, as we have seen, the RP as a whole does not seem to be tied to any specific movement or causal of movement in general. RPs have been elicited without movement, and brains that have been damaged such that they no longer produce RPs can still make voluntary movements. Might there be, though, other neural signatures associated with voluntary movements that might provide a separate, complementary avenue to advance research of this kind? As we shall see later, a separate line of research has addressed this

problem, and other drawbacks inherent in Libet’s approach. First, we turn to the subject of the, possibly counter-intuitive idea of the gradual emergence into consciousness of the feeling of will.

3.2 Does the feeling of intention appear abruptly or emerge gradually? When we make a decision, particularly one as simple as moving part of our body, it seems to happen in an instant and the movement is made. But is this what actually takes place in the brain, or could it be a more gradual process, as the rising negativity of the RP suggests? The evidence, taken from behavioural and neural data, as well as mathematical modelling, is suggestive of three things: that rather than appearing instantaneously, our awareness of wanting to move, and then moving, emerges gradually; that neural activity also builds gradually; and that this underlying activity consists of random fluctuations which may or may not cross a threshold of action.

In Libet’s experiment subjects reported their feeling of will by the active process of noting and remembering the clock time, a method that, as we have seen, is controversial and possibly unreliable. An alternative approach by Matsuhashi and Hallett (2008), probed subjects during a modified Libet-type task. Tones

were played during trials; subjects were told to ignore the tones, unless they happened to coincide with an urge to move, in which case they should veto the movement if possible. By comparing tone and movement onset time

distributions, the researchers calculated an estimated “time of thought to act”, or “movement genesis”, T. Importantly, this was not the same measure as Libet’s W which requires subjects’ after-the-fact recall of a time-point. Mean T time was 1.42 s before movement – over a second before typical W times. They separated these two times conceptually, calling the latter—a state of awareness that subjects are able to independently remember and report (W)—“meta-awareness”. They envisioned the process of endogenous decision

formation as a ramping up, from unconsciousness, via probeable consciousness, into meta-awareness just before the movement itself (Figure 6).

This behavioural result suggesting a gradual increase in activation is supported by data from single-unit intracranial recordings made by Fried, Mukamel and Krieman (2011) from epilepsy patients. Extracellular recordings were made from medial frontal regions including pre-SMA, SMA and anterior cingulate cortex, as well as temporal areas. The patients carried out a Libet-style task. The

researchers found that 17% of frontal units recorded showed a significant change in firing rate in the time leading up to W, as compared with 8% of temporal units. They also found that the number of neurons whose activity changed increased as W approached. This study was unusual in applying the high spatial and temporal precision of intracranial recording to the question of

spontaneous decision emergence. The authors interpreted the results as suggesting a process of gradually increasing premotor activity, describing an integrate-and-fire model in which a threshold is crossed resulting in the subjective feeling of will.

Figure 6: Illustration of the passage of a subject's intention to move from unawareness, via probeable awareness, to reportable meta-awareness. T is the earliest point that subjects were aware when probed, P is the a point of no return beyond which a movement cannot be vetoed. BP =

To explain the gradual buildup of activity prior to a spontaneous movement, Schurger, Sitt and Dehane (2012) constructed a stochastic accumulator model, based on the evidence-accumulation theory of environmentally-driven decisions (e.g. Gold & Shadlen, 2007), with evidence replaced in this case by random neural noise. If empirical data matched this model it would imply that activity preceding movement was, rather than being deliberately preparatory, in fact random. Spontaneous Libet-type decisions would therefore occur when this activity, by chance, produced sufficient negativity to cross a threshold. To test this, the researchers reasoned that if the model were correct, the closer the negativity was to the threshold at a given moment, the less effortful it would be to respond to a cue at that time, and therefore the reaction time (RT) to that cue would be shorter. They carried out a variant of the Libet procedure in which during trials subjects would hear a click—a cue for them to make a movement immediately. If they felt the urge to move before the click however, they should also move. Thus, trials consisted of a mixture of spontaneous and cued movements, with each trial halting following one or the other. Trials in which subjects moved following the click were grouped by RT as either “slow” or “fast”. EEG data showed greater negative shift before clicks on fast trials. The shape of the early RP was a good fit to the model, implying that this portion of the RP was indeed random and hence non-volitional. This view of the RP can explain why greater negativity preceded fast RTs on cued trials than slow RTs: the potential was closer to the threshold of action in those cases. The alternative view of the RP—as specific preparation (conscious or not)—cannot explain this result, the authors argue. This ties in with evidence from the previous section that the RP does not reflect specific preparation from movement.

So far we have seen that the RP may be only loosely linked to preparation for voluntary movement, that awareness seems to emerge somewhat gradually from unconsciousness, and that this emergence may, at least in part, come from

randomly fluctuating activity. This view stands in contrast to Libet’s

interpretation of the RP as representing unconscious but deliberate movement preparation. However, this alternative view poses a problem for defenders of free will as well, at least as applied to the sort of actions studied here: even if a

spontaneous movement is not ‘decided’ until a threshold (and, possibly,

consciousness) is arrived at, that ‘decision’ is nothing more than the culmination of random unconscious activity: how, then, can it be said to be either free or willed? Taking a different approach, if, as negativity fluctuates towards a

threshold, could we assign a likelihood of a movement being made, based on how close to threshold activity was at that time? In other words, is it possible, using this and other techniques, to make predictions about a person’s decisions before they become aware of them? This possibility raises more questions and may present more problems for proponents of conscious control, as we shall see in the following section.

3.3 Can we predict behaviour from brain imaging data?

Since Libet, data from EEG recordings has dominated this field of research. However recently a new, complementary, approach has got underway, in which the high temporal resolution of EEG has been exchanged for the high spatial

resolution of fMRI, and focus has shifted away from the RP. One rationale for this new method was previous evidence that non-movement-related frontal areas were also associated with free decisions (Deiber et al., 1991). Another was that since the RP begins fairly shortly before the associated action and may not reflect the earliest point in the decision-making process (on top of the fact that, as we have seen, large parts of the RP itself may not be directly movement-related or are simply random) (Haynes, 2011). These experiments were also better able to tackle a problem relevant to concepts of free will: whether behaviour can be predicted in advance from brain activity alone, a hitherto under-addressed question.

In the first of these experiments, Soon, Brass, Heinze and Haynes (2008) used fMRI to try to answer three questions: whether brain activity preceding movement begins in the SMA (the presumed origin of the RP signal); whether there is evidence of preparatory activity earlier than RP onset; and whether measured brain activity could be used to reliably predict actions. Watching a 2 Hz letter-stream, (in place of the standard analogue clock face), subjects were asked to make a movement with their choice of left or right hand the moment they felt the urge to do so, and to remember the letter on the screen at the time of the urge. Using a linear multivariate pattern classifier and searchlight to look for predictive patterns in voxel activation, the researchers found that activity in the frontopolar cortex (FPC) and an area between the precuneus and the posterior cingulate cortex was significantly predictive of hand choice up to 10 s prior to movement, albeit at less than 60% accuracy, (chance = 50%). Activity patterns from the pre-SMA were predictive of the movement’s timing up to 5 s before, with similar accuracy. These results were interpreted as reflecting unconscious decision-preparation activity stretching back much further in time than had been suggested by even the most liberal measures of RP onset. The FPC was taken to be the seat of spontaneous movement genesis; the top of a top-down hierarchy controlling spontaneous action, with the precuneus acting as a storage centre for the decision prior to its entry into consciousness and ultimate execution. As a side-note, although the FPC sits in some sense at the “top” of a hierarchy, its function appears to be limited to maintaining focus on a task in the face of environmental distraction, and switching to an alternative task if applicable; it does not appear to be involved in generating tasks itself (Koechlin & Hyafil, 2007). In a follow-up study, Bode and colleagues (2011) repeated the experiment using a more powerful scanner. Interestingly, they found that activity patterns became increasingly self-similar approaching movement execution. This evolutionary process of increasing stability may reflect neural activity building towards a threshold, such as in the accumulator model proposed by Schurger and colleagues (2012).

In a second follow-up study, the nature of the decision was changed. Up to this point only simple, meaningless motor decisions had been investigated. Instead, Soon, He, Bode and Haynes (2013), again using fMRI, asked whether the same processes underlie the making of an abstract choice as those associated with motor choices. Using a streaming series of screens showing numerals and letters, subjects were free to choose at any time whether to add or to subtract two of these numerals as they appeared. They found patterns of activity in the same

medial frontopolar area and precuneus/posterior cingulate cortex as before, which were choice-predictive with an accuracy of 59% up to 4 s before the decision. Information about decision-timing was encoded in pre-SMA, SMA and rostral cingulate cortex with accuracy significantly above chance.

Even in the absence of brain imaging data, humans are not completely unpredictable. The meaningfulness of predictions made using multivariate pattern analysis on fMRI data has therefore been called into question, because of possible contamination from trial-to-trial sequential dependences, or carry-over effects. In other words, choices made on one trial may have influenced the supposedly independent choice made on the next. To investigate this, Lages and Jaworska (2012) replicated only the behavioural component of the experiment of Soon and colleagues (2008); that is, they took no fMRI data. They trained a linear classifier to predict subsequent choices from previous ones. They found that prediction accuracy using this method to be comparable to that gleaned from fMRI data, around 60% (chance = 50%). The researchers did not take this finding as a dismissal of the fMRI findings, but more as a reminder of the dangers of reading too much into results without considering other possibilities. However, it may be simply a coincidence that the predictive accuracies found using these two quite separate techniques turned out on this occasions to be about the same. It remains to be seen whether improvements in the resolution of brain imaging technology will be able to push predictive accuracy substantially higher than that found here using only behavioural data. If this happens, and presumably no comparable improvements in behavioural data analysis techniques take place, then surely concerns raised here,

about reading too much into imaging data’s predicate power, would no longer be warranted. Moreover, even if there are

unavoidable dependencies in this type of experiment, if predictions can be made in advance of a

subject’s conscious awareness,

then the conclusion that at least some specific unconscious preparation occurs, and the implication this holds for

conscious control, remains valid (Bode et al. 2014).

fMRI scanning is not the only method that has been used to predict the occurrence and

timing of movements. Mentioned in the previous section, Fried and colleagues (2011) recorded from intracranial electrodes placed in epilepsy patients’ heads, while they made spontaneous movements. Recording the patterns of activity leading up to a movement from neurons in the medial frontal area, they used a support vector machine classifier to predict the occurrence of W from preceding data taken from 37 units in an individual session. Performance rose above chance

Figure 7: Predictive performance of pattern classifier leading up to W (dashed vertical line). Dashed

horizontal line represents chance performance. (From Fried et al., 2011)

(50%) around 1000 ms prior to W, gradually rising to around 90% just before W (Figure 7). Combining multi-patient data to form a 512-recording

pseudopopulation gave 90% accuracy in predicting W 500 ms before its reported time, as well as predicting W’s timing an average of 152 ms prior.

A study by Maoz, Ye, Ross Mamelak and Koch (2012) also used intracranial recording, but this time combined with a system able to make “on-line real-time” (ORT) predictions from the data. Recording from a combination of intracranial depth microelectrodes and subdural grid electrodes, their system was able to predict, with “far above chance accuracy”, which of two actions an epilepsy patient would perform, then pass that information to an experimenter before the

patient moved. This allowed experimenters to play a game with patients: a

two-choice version of “rock, paper, scissors”, in which patients stood to win money— an important difference from previous Libet-type experiments in which choices are meaningless. Patient and experimenter both held a button down under each hand, then on a cue each player lifted either their left or right hand. The rule was simple: to win, the patient had to lift a different hand from the experimenter’s. Half a second before the cue, a screen visible only to the experimenter would show a left or right-facing arrow to indicate the hand the system had predicted the patient would lift; the experimenter would then lift that hand. This game was conducted with two patients, using low-frequency LFP data from ten electrodes giving an accuracy of 68±3% half a second in advance of movement. Using off-line data from 30 electrodes (i.e. not ORT), accuracy rose to 83%. The

researchers speculated that with sufficient processing power, this level of accuracy could be achieved in real time, as well as predicting decisions “several seconds before movement onset”. It should be stressed that the design of this experiment precludes direct comparison with Libet-type studies—movements were cued and not spontaneous so no W time could have been recorded and nothing can be said about whether data about unconscious decisions was gathered. However, these results, together with those of Fried and colleagues (2011), strongly suggest that predictions of spontaneous decisions are possible with potential accuracy much greater than is currently possible using fMRI. A major drawback of the research just outlined is its invasive nature. Not only is the opportunity for such experimentation rare, but the experimenter also has no control over the location of any implanted electrodes, with clinical requirement being the only criterion. Recent research by Bai and colleagues (2011) however, has shown that non-invasive ORT movement prediction is possible. This study used EEG data but rather than trying to pick out RP onset from the noise of single-trial recordings, their predictive model was based largely on event-related desynchronisation in the beta frequency band (15-30 Hz) and deliberately designed to give a low false positive rate, at the expense of higher failures to predict. The experiment simply required subjects to make self-paced wrist movements – there was no clock. They found that for predictions made less than 1.5 s before movement the mean prediction time was 0.62±0.25 s before

movement. For the seven subjects, the mean prediction accuracy was 75%, though only 40% of movements were preceded by a prediction. In a follow-up study the same research group used their predictive model to investigate what subjects were thinking about when they made spontaneous movements

(Schneider et al. 2013). When the model predicted an imminent movement subjects were immediately asked whether they had been about to move; that is, whether they had felt the urge to do so, and if not, what they had been thinking about. They found that in 14% of cases subjects were thinking about something other than movement during movement preparation.

The ability to predict behaviour from brain data is a very recent achievement in neuroscience and the results so far are modest and sometimes equivocal. None of the studies described above can be said to deliver a definitive answer to the question of whether, or to what extent, the details of a person’s upcoming behaviour can be decoded from patterns of neural activity before the person becomes aware of it himself. Moreover, as the findings of Lages and Jaworska (2012) remind us, no act exists completely in isolation and without context; and however hard a person tries to act spontaneously, producing a truly

unpredictable behavioural sequence appears to be extremely difficult if not impossible (Wagenaar, 1972; Schulz, Schmalbach, Brugger & Witt, 2012): humans are at least somewhat predictable even in the absence of imaging technology. Nevertheless, these recent results have been made possible largely through advances in imaging and computer processing technology. There is no reason to suppose that this technology will not continue to improve, thus

providing an increasingly clear picture of neural activity over the seconds leading up to an act of volition. There are surely limits to how accurate predictions can be (in principle or in practice) and how far in advance they can be made, but there seems no reason to assume those limits have already been reached. The question of behavioural predictability and its implications for free will therefore remains open.

4. Applicable theoretical work

The studies reviewed above may be used to argue for or against various theoretical conceptions of how spontaneous decisions are made, the role of consciousness in making decisions, the implications of these things for “free will”, and finally what precisely we mean by that term. We will briefly look at three kinds of response to the question of free will: one in which conscious control is deemed an illusion; a second in which consciousness is not an important component of free will; and a third in which consciousness does perform an executive role.

4.1 Wegner’s theory of apparent mental causation

Introduced by Wegner and Wheatley (1999), this theory states that the

experience of direct conscious control of our actions is an illusion and that in fact our actions and our feeling of will both arise from unconscious precursors. It is also a theory in a second sense: the individual, provided with the evidence of the co-occurrence of thought and action automatically (mis)attributes (or

Figure 8: Schematic diagram contrasting the actual causal paths (green) that occur when an action is executed, with the causal path that is assumed by the subject (purple), according to the theory of apparent mental causation (From Wegner, 2003).

Three conditions must be met for this illusion of conscious will to occur: “priority”, in which the action immediately follows the experience of will; “consistency”, in which the thought and the action always ‘match’—that each exercise of will ‘successfully’ generates the desired action; and “exclusivity”, no other plausible causes of an action are present.

If the experience of conscious will and its associated action are not causally linked, they should be dissociable—it should be possible for one to exist without the other. Wegner and Wheatley (1999) offer an experiment of Brasil-Neto and colleagues (1992) as an example of manipulation behaviour without changing a person’s sense of volition. In this experiment focal TMS was applied unilaterally to the motor cortex. In a forced choice task, subjects were asked to move their left or right fingers each time they heard a click. Subjects moved the hand

contralateral to the stimulus more often, but they were unaware that the TMS was affecting their behaviour; that they were acting under their own volition. This experiment seems to show that behaviour can be manipulated, leaving the sense of volition unchanged. Wegner and Wheatley (1999) also cite examples such as Ouija boards and automatic writing, as well as disorders like alien hand syndrome as examples of apparent voluntary action taking place in the absence of conscious will. Wegner and Wheatley (1999) also carried out an experiment to demonstrate the dissociation of will and action in the opposite direction, that is, to show that the experience of will can be induced without an accompanying voluntary action. A subject and a confederate both placed their hands on a computer mouse. The confederate moved the mouse, moving a pointer on a screen. When subjects were primed with the name of an on-screen object under

the pointer, and the confederate, under instruction, stopped moving the mouse either one or five afterwards, subjects reported that they felt they themselves had caused the pointer to stop. Conversely, if the object was mentioned 30 seconds before, or one second after the pointer had stopped, there was no such induced sense of agency. Thus, in this experiment, the condition of priority was

manipulated, giving the illusion of agency. The study of Banks and Isham (2009), discussed earlier, also supports this theory, Delayed feedback was used to

manipulate subjects’ perception of the time they intended to move, suggesting that the feeling of intention is at least partly inferred from external cues.

Libet’s (1985) interpretation of his results is of course consistent with Wegner’s (Wegner & Wheatley, 1999) picture of an independent origin of action and will; indeed, Wegner (2002) suggested that although we do not yet know exactly what is indexed by the RP, it could reflect these two parallel processes. Wegner

contrasted the results of Libet with various experiments concerning rapid reactions to stimuli, in which the reaction (e.g. hitting a tennis ball) occurs several hundreds of milliseconds before consciousness has the chance to “catch up”. Although ‘automatic’ actions such as reacting to a moving ball cannot be considered voluntary and spontaneous in the way those studied by Libet would be, Wegner saw the temporal mismatch—onset of the RP occurring before the urge to move—as evidence that consciousness was no more an initiator of action in voluntary movement as it is in automatic or reflexive movements. He also argued that mental rehearsal of action may run at a slower speed than the action itself, thus they cannot run concurrently (Wegner, 2002).

4.2 Bonn’s model of volitional action

Bonn (2013) has proposed a model of volitional decision-making as a complete process, thus it is broader in scope than that of Wegner, incorporating diverse areas of brain function (Figure 9). It also lays out a conception of free will in which consciousness plays a minimal role at best. It is based around two

feedback loops that are, roughly, inward and outward-looking from the point of view of the subject. The first of these loops is inward-looking because it is based around the default mode network (DMN) and therefore operates at times when interaction with the environment is minimal. The DMN integrates information drawn from conscious experience of the environment, via memory, to create a constantly-updating picture of the world. Interestingly, patterns of activation associated with spontaneous decision preparation have been found to overlap with those associated with the DMN (Soon et al., 2013). The second loop interacts directly with the environment and is anti-correlated with the first— when one is active the other is not. Goals are formed by the executive control network and executed via the motor system. Importantly, this model gives only a limited causal role to conscious experience. It allows for the possibility of direct conscious intervention via Libet’s (1985) veto, (although, as we have seen, there is reason to believe veto-decisions arise via unconscious processing first, see Brass & Haggard, 2007 and Filevich et al., 2013), but otherwise, conscious experience feeds into memory, passing through the DMN before arriving at the control network; a highly indirect route.

Figure 9: Model of volitional action featuring two feedback loops centred around the default mode network (top) and the executive control network (bottom) (From Bonn, 2013).

Despite consciousness’s limited executive role, this model is in accord with Bonn’s (2013) argument that humans possess free will, although the description of the concept put forward here is quite carefully defined. Most notably, it does not require conscious awareness of everything one is doing—if you walk somewhere you might not be aware of every step you take, but you still feel in control and moving of your own volition. This view of free will is centred on independently generated ideas and actions, rather than conscious control. It could be thought of, perhaps, in opposition to the out-dated behaviourist view of action as being entirely stimulus-driven (Skinner, 1971). Consciousness,

meanwhile, is relegated to a supervisory role, providing feedback from, but not directly controlling, behaviour. Indeed, Bonn describes the “folk” notion of free will, in which conscious awareness is the originator of thought and action and sits in the driving seat of our behaviour, as “dead in the water”. The experiments of Libet (1985) and those following cannot count as evidence against this kind of free will; they merely show that preparatory brain activity occurs before we are aware of having made a decision. Nor can the results of Soon and colleagues (2008) with their limited but nonzero powers of prediction. Indeed, even if their predictive capacity were much greater, the actions being predicted would be no less a creation of the individual making them and thus count as freely willed according to the definition given here (Bonn, 2013).

4.3 Is conscious causation possible?

Bonn (2013), in his conception of free will, is content to assign to consciousness executive powers that are limited at best, therefore the implications of the results of Libet (1985) and claims of pre-conscious behavioural prediction (Soon et al., 2008), pose no threat to free will so defined. For others however, a libertarian concept of free will that admits consciousness as a causal force unencumbered by

antecedent neural activity—precisely the scenario Libet claimed to have (mostly) disproven—is possible, provided one is willing to entertain “non-reductionist theories of agency” (Batthyany, 2009). These theories assert that the

phenomenon of agency cannot be “reduced to” (that is, fully explained by) the brain’s physiology, which itself is bound by the laws of physics. This allows for consciousness powers of causation that are independent of the physical brain; although neural activity can account for a large part of mental processing and behaviour, it cannot account for the full picture, according to this view. If one is prepared to accept non-reductionist theories, what impact do the results of Libet (1985) and others previously discussed have on this consciousness-first view of free will? Batthyany (2009) contends that they do not affect it in the way that is usually claimed; that is, by showing decision-related activity taking place pre-consciously. He argued that this is because actions can be divided into those that are active (voluntary) and passive (involuntary). Passive actions are responses to urges or desires, the appearance of which we do not control—this is not free will. It is in resisting these urges that we exercise the kind of freedom allowed in a non-reductionist worldview. Because the experiments in question involve responding to urges (i.e. to move one’s finger, press a button, etc.) subjects are only acting passively. An active decision would involve resisting this urge; Libet’s veto in other words.

In another libertarian critique of Libet-type studies, von Wachter (2013) took the results of Libet (Libet et al., 1983; Libet, 1985), specifically those related to the veto, not as irrelevant to the question of conscious agency, but in its favour. He claimed that “vetoing is not the result of preceding processes, because there is before the vetoing the same RP than in cases without veto.” (Wachter, 2013). Going further, he takes this apparent evidence that vetoes have no neural precursor activity as suggestive that conscious causation may be much more widespread in human actions, occurring every time we bring about a “choice event” (Wachter, 2013). Acceptance of this interpretation of the results of Libet-type experiments requires holding what are surely, for most scientists at least, quite eccentric metaphysical positions. What Batthyany (2009) calls “non-reductionist theories of agency” are more vividly described by von Wachter (2013) as “not governed by laws of nature.” Assessment of metaphysical

positions such as this is usually regarded as beyond the purview of science, but there is a difference here that matters. If one wishes to defend conscious

causation of the sort advocated here, one must accept “a radically different kind of causation than that which is usually observed in nature.” (Batthyany, 2009). If not, one must accept that, because this kind of causation apparently cannot exist, our impression of conscious will is, as Wegner (2002) insists, illusory. Thus a judgment must be made about whether we live in a radically different universe than conventional science would suggest, or whether our intuition can

5. Conclusion

As we have seen, Libet’s seminal work (Libet et al., 1985) has ignited a field of research that over the following three decades has remained active and has diversified considerably. Libet’s experimental paradigm has been repeated many times, often with changes and additions to explore and overcome the various criticisms that have been levelled at it. These have included various technical aspects of the experiment and subsequent data processing, as well as the

problems inherent in reliably timing subjective events. The issue of the veto has also been explored: initially claimed by Libet as occurring without accompanying brain activity, there is now evidence that although inhibitory decisions are not the same as spontaneous actions, they both have neural correlates. The precise nature of the RP, the lynchpin of Libet’s results, continues to be extensively studied, producing confusing, even contradictory results. Thus, although Libet’s results have been replicated many times and are in that sense robust, subsequent work has cast significant doubt on the close connection between the RP and spontaneous action. The process by which unconscious neural activity becomes conscious has also been explored, revealing a somewhat counter-intuitive picture of a gradual emergence rather than a sudden appearance of the feeling of

conscious will. It may also be that this feeling and subsequent action come about following random neural fluctuations which, being random, cannot be freely willed. fMRI has also been used to look for pre-conscious decision-related activity and patterns have been found that are at least somewhat predictive of an action’s content and timing. Behavioural prediction has likewise been demonstrated using intracranial electrodes and scalp EEG, and has been achieved in real time. Although the accuracy of these predictions is at present modest, they may improve as technology advances. Results such as these can be contextualised in light of several theoretical views of free will. To some they demonstrate that the kind of free will most people think they have, namely the visceral sense of conscious control, is illusory. Others do not wish to abandon the concept of free will and define it therefore as an individual’s powers of independent agency, a view compatible with the work discussed here. Still others maintain the

libertarian view that consciousness and its apparent power to control our actions implies a kind of causation that is outside of conventional physicalist notions. Matters are therefore still far from settled over thirty years after Libet published his original results (Libet et al., 1983).

There are, then, multiple definitions of free will, and work discussed here affects them in different ways. Advocates of the libertarian definition may contend that the results of Libet do not challenge their view because what is being tested is merely a reaction and hence not true volition. Their view seems to rely on a conception of the world and what is possible that is so at odds with conventional science, that very strong evidence is needed to support it. As we have seen, the example of the ‘uncaused’ veto, cited as evidence of this position, has been challenged, and it seems that vetoes are no more ‘free’ than the movements they inhibit (Brass & Haggard, 2007; Filevich et al., 2013). What sort of evidence might be required to support the libertarian view? A Libet-type experiment, employing the sort of ‘active’ decisions that, according to Batthyany (2009) employ libertarian free will, might be conducted. If it were reliably found that

subjects’ W times preceded any associated neural activity, this may point toward conscious causation. Similarly, if fMRI data showed patterns of neural activity corresponding to a volitional act appearing fully-formed at or after the moment the subject decides to act, this too would be suggestive that such powers of volition may be attributed to consciousness after all. However, it currently seems unlikely that such results will be found. If they were, and were sufficiently robust, it would surely herald a revolution not just in neuroscience but in our very conception of the way the universe works: it would show that the process of biological evolution has brought forth a form of causation qualitatively different from the familiar chance and necessity (Verbaarschot et al., 2015).

In the absence of such compelling evidence however, most scientists continue to view the world in mechanistic terms. Some, such as Bonn (2013), as outlined above, view the concept of free will rather differently from the libertarians. Instead of requiring conscious causation, this view sees free will as independent agency. This conception of freedom, designed to be compatible with a

deterministic or mechanistic world, is known as ‘compatibilism’(McKenna & Coates, 2015). Advocates such as Dennett (1984, 2003) argue that it is not necessary to insist that the brain somehow has the power to overcome or to transcend the laws of physics in order to retain the sort of free will “worth wanting” (Dennett, 1984). At the heart of this view is the ability of humans and other animals, to choose from multiple courses of action in a way that is

independent of external influence. The essence of this ability is shared with relatively simple animals such as flies and leeches, whose brains are capable of producing varied behaviours under identical experimental conditions (Briggman, Abarbanel & Kristan, 2005; Friesen & Kristan, 2007; Quinn, Harris & Benzer, 1974). This has led Brembs (2010) to argue that the concept of free will should be formally defined as a biological trait that is widespread in the animal

kingdom. For Brembs, the results of Libet are irrelevant to this concept because it does not require conscious causation (whatever that would mean for a fly or leech). So there is a problem here. Even if the brain of a fly is equipped with circuitry that lets it make an unpredictable self-generated ‘decision’, this surely amounts to little more than a neural roll of the dice which, while ruling out a strictly behaviourist view of animals a deterministic automata surely does not comport with any commonly held view of free will.

The question of predictability is perhaps now at the vanguard of free will studies, and improvements in technology may produce more accurate predictions under a wider range of circumstances. At present it is impossible to say how accurate these may become, but two possibilities seem particularly unlikely. First, that predictions will approach 100% several seconds before decisions are made. Several studies previously discussed imply a gradual increase in predictive accuracy (e.g. Figure 7), which probably reflects the brain’s gradual arrival at a decision (as in the increasingly similar patterns found by Bode et al., 2011), so we should expect to see a similar gradual increase in predictive accuracy. A second unlikely result would be negligible predictability until the moment of decision, implying the sudden appearance of complex, decision-related neural activity. We have already seen what looks like unconscious buildup of decision-related activity in the context of simple, meaningless decisions. A crucial question