The evidence from the brain

In 1983 Benjamin Libet reported that a brain signal preceding a voluntary movement begins about a third of a second before the person feels themselves deciding to move, and the result has been reported ever since as the experimental refutation of free will.

The experiments are real, well designed for their era, and they replicate. What they show is considerably narrower than the headline, and the most interesting development is that the central signal is now widely thought to be something other than what everyone assumed it was. This lesson works through the measurements, because the honest assessment depends on the numbers rather than on the summary.

What was measured

The background is a discovery by Hans Kornhuber and Lüder Deecke in 1965. Averaging electroencephalogram traces backwards from the moment of a self-initiated movement, they found a slow negative drift over the motor areas beginning up to a second or two before the movement itself. They called it the Bereitschaftspotential, the readiness potential, and it became a standard tool.

Libet's addition was to time the subjective side. His subjects sat with a modified oscilloscope in front of them, on which a spot of light revolved like a clock hand, taking 2.56 seconds per revolution, so that each degree of arc corresponds to about 7 ms. They were asked to flex the wrist whenever they felt like it, with no preplanning, and afterwards to report where the spot had been at the moment they first became aware of the wish or urge to move. That reported moment is called W. Muscle onset was recorded by electromyography and used as time zero.

The results, from Brain in 1983, are the numbers everyone quotes. For spontaneous movements reported as unplanned, the readiness potential began about 550 ms before the muscle activity. The reported moment of awareness, W, came about 200 ms before muscle activity. So the brain signal precedes the felt decision by roughly 350 ms.

Libet checked the timing method rather than assuming it. Subjects were also asked to report the moment of a small skin stimulus delivered at a random time, and their reports were biased by a few tens of milliseconds, which lets the W reports be corrected for the same bias. This is often overlooked by critics, and it matters: the method has a known error and the effect is much larger than the error.

Libet himself did not conclude that free will is an illusion. He argued that since W falls about 200 ms before the muscle activity, and the final motor command occupies about the last 50 ms, there remains a window of roughly 150 ms in which the conscious subject can abort the movement. Conscious will, on his picture, is not the initiator but the editor: not free will but, in his phrase, free won't.

Example. Work out the two intervals in Libet's design and say what each one is evidence about.

The first interval is from the readiness potential to W: $550-200=350$ ms. This is the celebrated result, and it is evidence that measurable preparatory activity precedes the moment a subject reports first noticing an urge. The second is from W to muscle onset, 200 ms, and subtracting the roughly 50 ms of final motor command leaves about 150 ms. This is what Libet's veto proposal rests on, and it is the weaker of the two claims: it depends on a subject being able to cancel in a window they cannot report on afterwards, and the direct evidence for a veto is much thinner than the evidence for the timing.

Now you. A critic says the whole design is worthless because people cannot accurately time their own mental events. Take the objection seriously and say what survives it.

Ten seconds, and sixty percent

The most dramatic follow-up came from Chun Siong Soon, John-Dylan Haynes and colleagues in 2008, using functional magnetic resonance imaging. Subjects chose freely between pressing a left or a right button while watching a stream of letters, and reported which letter was on screen when they decided. Applying pattern classification to activity in frontopolar and parietal cortex, the researchers could predict which button would be pressed up to about 10 seconds before the reported decision, some eighteen times further ahead than Libet's readiness potential.

The number that matters is the accuracy, and it is about 60 percent against a chance level of 50. That is a real effect, statistically solid, and small. Betting on the decoder would win 60 times in 100 instead of 50, an edge of 10 percentage points. It is the signature of a weak prior bias, of the kind you would expect if a subject who pressed left three times running is slightly disposed to press right next, and it is nothing like a readout of a decision already taken.

A third strand comes from single neurons. Itzhak Fried and colleagues in 2011 recorded from electrodes implanted in patients with epilepsy for clinical reasons, and found that populations of a few hundred neurons in the supplementary motor area changed their firing rates progressively before the reported moment of decision, allowing the impending choice to be predicted with better than 80 percent accuracy a few hundred milliseconds ahead. This is a much stronger signal than the fMRI result, from a much more direct measurement, and it is the best evidence that the preparation is real neural activity rather than an artefact of averaging.

The accumulator

The deepest challenge to the standard interpretation is not a criticism of the experiments but a rival explanation of the readiness potential, published by Aaron Schurger, Jacobo Sitt and Stanislas Dehaene in 2012.

Start from a fact about the task. The subject is told to move whenever they feel like it, with no reason to prefer any moment. There is nothing to decide, so something has to break the symmetry, and the obvious candidate is ongoing spontaneous fluctuation in motor cortex activity. Model it as a noisy accumulator drifting up and down, with a movement triggered when the accumulated activity crosses a threshold.

Now consider what happens when you average trials backwards from the moment of movement, which is what the readiness potential is. The trials selected are exactly those in which the noise happened to be drifting upward towards the threshold. Averaging them produces a slow rising negativity before the movement in every trial, even though on no individual trial was there any decision at the moment the average appears to start rising. The readiness potential, on this account, is a picture of the noise that got selected, not a picture of an unconscious decision.

The model makes a prediction that distinguishes it. If subjects are interrupted by an occasional cue demanding an immediate movement, their reaction times should depend on where the fluctuation happened to be when the cue arrived: fast if it was near threshold, slow if it was far. Schurger's group tested this and found the predicted pattern. Later work, including a 2021 review by Schurger and colleagues, has continued to support the reinterpretation, and the readiness potential is no longer safely described as the neural signature of a decision.

There is a second constraint from a different direction. Uri Maoz and colleagues reported in 2019 that the readiness potential appears before arbitrary choices, of the pick-one-at-random kind Libet used, and is largely absent before deliberate choices with real stakes, such as which of two charities should receive a donation. If that holds, the entire literature has been measuring the neural correlate of picking rather than of deciding, and the extrapolation to meaningful choices was never licensed.

Example. A newspaper reports that scientists can predict your decisions ten seconds before you make them, so free will is dead. Rewrite the claim so that it is accurate.

Something like this: in a task where subjects press one of two buttons for no reason at all, a pattern classifier applied to fMRI data predicts which button will be pressed with about 60 percent accuracy, against 50 percent by chance, from activity several seconds before the subject reports deciding. That accuracy corresponds to a weak bias rather than a settled decision, the task involves no reasons and therefore no deliberation, and nothing in the result distinguishes a deterministic brain from an indeterministic one. The honest headline is that arbitrary picking is preceded by measurable brain states that partly bias it, which is what anyone on any side of this debate would have expected.

Now you. What would an experiment have to show in order to genuinely threaten the sort of free will philosophers argue about?

The veto, tested

Libet's own positive proposal, that consciousness retains a power of cancellation, went untested for thirty years and then was tested rather well.

Matthias Schultze-Kraft and colleagues reported in 2016 an experiment in which subjects played a game against a computer. Electroencephalogram signals were decoded in real time, and when the system detected the build-up preceding a movement it presented a stop signal. The question was whether a movement already under preparation could still be cancelled.

It could. Subjects successfully aborted movements after the preparatory activity had begun, which is a direct demonstration that the readiness potential does not commit anyone to anything. But the ability had a deadline: cancellation failed if the stop signal arrived later than roughly 200 ms before the movement, a point the authors called the point of no return, corresponding to the stage at which the final motor command is on its way.

Both halves of that result matter. The first vindicates Libet's veto in outline and further weakens the claim that the readiness potential is a decision. The second sets a real limit: there is a last moment after which nothing can be recalled, and it is a fraction of a second wide. What the experiment does not show is that the cancelling is done by something outside the brain's ordinary causal processes, and nobody involved suggested it was.

Example. How does the 2016 result bear on the standard interpretation of Libet's 1983 finding?

It undermines it directly. The standard interpretation is that the readiness potential is the brain deciding, several hundred milliseconds before the person believes they are deciding, so the conscious decision is a report on a decision already taken. If a movement preceded by that same build-up can still be cancelled, the build-up cannot have been the decision: a decision that can be reversed at will by the agent is a preparation, not a verdict. Combined with the accumulator model, which explains the signal as accumulated noise selected by the averaging procedure, very little is left of the original interpretation. The measurements were sound and the story attached to them has been substantially rewritten by the same experimental tradition, which is what a healthy field looks like.

Now you. Does the point of no return, roughly 200 ms before movement, restrict free will in any philosophically interesting way?

What people report about their own reasons

A quieter literature bears on responsibility more directly than any of the timing work, and it concerns not when decisions are made but how badly people know why they made them.

Richard Nisbett and Timothy Wilson reviewed the evidence in 1977 and reported experiments in which shoppers evaluating identical items in a row chose the rightmost far more often than the leftmost, and then denied firmly that position had anything to do with it, offering explanations in terms of quality instead. The reports were not lies. The subjects had no access to the process that produced the preference, and generated a plausible account in its place.

Later work has made the point sharper. Lars Hall and Petter Johansson demonstrated choice blindness in 2005: subjects who chose which of two faces they found more attractive were handed, by sleight of hand, the face they had rejected, and asked to explain their choice. A large majority failed to notice the swap and then explained, in detail, why they preferred the face they had in fact rejected. Split-brain patients studied by Michael Gazzaniga do something structurally identical: the verbal hemisphere, presented with an action initiated by information only the other hemisphere received, produces a confident reason for it.

The implication for this course is not that we never know why we act. It is that the faculty producing our accounts of our own reasons is a constructor rather than a reader, and that a theory resting on the agent's own report, of the kind the hierarchy of desires might seem to invite, is resting on something unreliable. This is one place where empirical work genuinely constrains the philosophy, and it favours accounts, like reasons-responsiveness, that test the mechanism by what it does rather than by what the agent says about it.

What the evidence establishes

Three conclusions, and a warning.

The first conclusion is that the folk picture of a conscious self standing outside the brain and starting the causal chain is not supported and probably false. Preparation for movement is underway before people report noticing an intention, and the report itself is partly reconstructed after the fact. That is a genuine finding about how the mind works, and it is not nothing.

The second is that this refutes almost nobody in the debate. Compatibilists have held since Hobbes that a free action is caused by processes in the agent, and are untroubled by the discovery that those processes begin before the agent notices them. Libertarians need the decision to be undetermined, and no timing experiment addresses determinism at all: a noisy accumulator can be deterministic or stochastic, and the data do not distinguish those cases.

The third is that the specific signal at the centre of the literature is now contested. If the readiness potential is a selection artefact, the most-cited experimental result in the philosophy of action turns out to be measuring the shape of neural noise.

The warning is against the opposite overreaction. It would be wrong to conclude that neuroscience has nothing to say here, and this lesson is not a defence of the will against science. The evidence that does bear on responsibility is quieter, better replicated and rarely reported as a free will story: how much of behaviour is heritable, how strongly circumstances move people who believe themselves to be acting on principle, and what a tumour can do to a character. That is the next lesson, and it is the one that should worry you.