Insight Editions · No. 287
Consciousness and the Brain
Deciphering How the Brain Codes Our Thoughts
Stanislas Dehaene
Dehaene argues that consciousness, once narrowed to conscious access, is a measurable brain process.
A study guide to Consciousness and the Brain by Stanislas Dehaene. Shared free for personal reading; not for sale and not endorsed by the authors or publishers. Please read and buy the original.
The Book on One Page
Consciousness and the Brain (2014) is Stanislas Dehaene's account of how fifteen years of experiments at his Paris lab turned consciousness into a measurable laboratory phenomenon. By narrowing the question to conscious access, making stimuli visible or invisible at will and treating subjective reports as data, he finds four signatures of consciousness and explains them with the global neuronal workspace, a brain-wide broadcasting network. Consciousness, on this account, is global information sharing: one selected representation ignites a long-distance prefrontal-parietal network and becomes available to every specialised processor, which is what lets us hold, route, report and reason about it.
Conscious access is a late, all-or-none ignition of a long-range cortical network that broadcasts one content to the whole brain; everything else, including meaning, value and attention, can run unconsciously but only briefly and locally.
Cognitive neuroscience of visual and auditory awareness (masking, attentional blink, rivalry), fMRI, EEG, MEG, intracranial and single-neuron recordings, computer simulations, the coma clinic, infants, monkeys, schizophrenia and machine consciousness, from Descartes to about 2013.
~270 to 300 ms: onset of conscious ignition and the P3. About 50 ms: the masking threshold. 70% vs 25% of intracranial sites reached by seen versus unseen words. 4 signatures. 5 kinds of unconscious. 3 to 4 times slower ignition in two-month-olds.
Readers who want the strongest empirical case for a functional, testable theory of consciousness, and anyone who must judge what a system knows, reports or can be trusted to decide, from clinicians to AI architects.
Three ingredients made a science. Dehaene separates vigilance, attention and conscious access, and studies only access: when information becomes reportable. Minimal contrasts (masking, the attentional blink, binocular rivalry, inattentional blindness) hold the stimulus fixed while awareness varies, and reports are primary data, though introspection is never trusted as a window on mechanism.
The unconscious is deep but shallow in time. Masked words and digits are recognised, bound, understood, valued, attended to and fed into motor preparation; virtually every brain region can work unconsciously. But subliminal effects are small, decay within about a second, stay local (meaning in the left temporal lobe, for example) and cannot be chained across steps. Mapping this draws a photographic negative of consciousness.
Consciousness has specific jobs. It collapses unconscious probability distributions into one sample fit for action, holds information over time (trace conditioning, working memory), chains operations as a slow serial "human Turing machine", and shares a compressed, confidence-tagged brief with other minds.
Four signatures mark the moment of access. A self-amplifying ignition of parietal and prefrontal cortex; the P3 wave from about 270 ms; a late, sustained gamma burst; and long-distance synchrony with a bidirectional "causal explosion". Early sensory waves are identical for seen and unseen stimuli; awareness corresponds to a nonlinear, all-or-none phase transition. Content-specific neurons (Quiroga's "Clinton cell") fire only when the person reports seeing, and stimulation and disruption studies show late, long-distance loops are causal.
The global neuronal workspace explains them. Long-axon pyramidal neurons in layers II and III, densest in prefrontal and parietal hubs, form a broadcasting network. Simulations with Jean-Pierre Changeux reproduce all four signatures, show a vigilance-controlled threshold, and produce spontaneous ignitions that block outside input. The theory sorts the unconscious into preconscious, subliminal, disconnected, diluted and latent knowledge.
The theory passes clinical tests. The local-global auditory paradigm, Massimini's TMS-EEG and resting-EEG information sharing detect consciousness in patients, sometimes before the clinician, complementing Owen's tennis-imagery fMRI. Central thalamic stimulation, dopaminergic drugs and zolpidem can switch a silenced cortex back on.
Beyond adults. Two-month-old infants show a slow but real workspace; monkeys share illusions, blindsight, the local-global response and metacognition; humans add a recursive language of thought. Schizophrenia and anti-NMDA-receptor encephalitis look like diseases of top-down long-range connections, and a machine with global broadcasting, learning and autonomy could, Dehaene argues, be conscious and freely willing; the hard problem will dissolve like vitalism.
| Ch | Title | One-line payload |
|---|---|---|
| 0 | Introduction: The Stuff of Thought | Descartes's two challenges; three ingredients (access, manipulation, reports) make consciousness a lab science; the workspace in outline |
| 1 | Consciousness Enters the Lab | Access versus vigilance, attention and self; minimal contrasts via rivalry, blink, inattention and masking; subjective reports as primary data |
| 2 | Fathoming Unconscious Depths | Unseen stimuli are recognised, understood, valued and attended to across the brain, but effects are small, brief and local |
| 3 | What Is Consciousness Good For? | Consciousness collapses probabilities into one sample, holds information, chains serial operations and shares confidence-weighted briefs |
| 4 | The Signatures of a Conscious Thought | Ignition, P3, late gamma and long-distance synchrony at ~300 ms; content-specific neurons; stimulation proves late loops are causal |
| 5 | Theorizing Consciousness | The global neuronal workspace: long-axon hubs, simulations reproducing the signatures, spontaneous activity, a taxonomy of the unconscious |
| 6 | The Ultimate Test | Disorders of consciousness; Owen's imagery test; local-global, TMS-EEG and EEG information-sharing detectors; thalamic and drug interventions |
| 7 | The Future of Consciousness | Slow workspace in infants; animal and metacognitive consciousness; human recursion; schizophrenia; conscious machines and free will |
Introduction: The Stuff of Thought
Consciousness, long the last refuge of dualism, has become a laboratory phenomenon. Dehaene names three ingredients that made the change possible: a narrower definition (conscious access), tools that make stimuli visible or invisible at will, and a new respect for subjective reports as raw data. With them his team found reproducible signatures of consciousness and built the global neuronal workspace theory, which he presents as the answer to the two challenges Descartes said no machine could meet: flexible reasoning and reporting one's thoughts.
The Dreamer at Lascaux
In the Apse of the Lascaux cave, at the bottom of a sixteen-foot pit beside a disembowelled bison and a rhinoceros, lies one of the rare human figures in Paleolithic art, painted about 18,000 years ago. The man lies flat on his back, arms out, with an erect penis; beside him stands a bird on a stick. The sleep researcher Michel Jouvet read the scene as a dreamer and his dream. Jouvet's team discovered "paradoxical" (REM) sleep, in which the brain is nearly as active as in waking, the eyes dart about, and men invariably have an erection, even in dreams without sexual content. Jouvet's point is that our ancestors could easily have noticed this, and the bird is the natural emblem of the mind flying free while the body lies inert.
The image recurs across cultures, which is what makes the reading plausible:
- Ancient Egypt. The Ba, the immaterial soul, was a human-headed bird, often shown with an erect phallus; at death it took flight to the afterworld. Osiris is shown supine and erect while Isis as a bird hovers over him to conceive Horus.
- The Upanishads. The soul is a dove that flies away at death and may return as a spirit.
- Christianity. Doves and white-winged birds came to stand for the soul, the Holy Spirit and angels.
- Elsewhere. The Egyptian phoenix (resurrection) and the Finnish Sielulintu, the soul bird that brings a psyche to newborns and takes it from the dying.
Behind the allegory sits the dualist intuition: thought seems made of different stuff from the body, since in dreams it roams freely while the body lies still.
One reading among several. Jouvet's dream interpretation of the Lascaux shaft scene is speculative. The figure is usually described as bird-headed, and archaeologists have proposed rival readings: a shamanic trance, a hunting accident in which the bison has gored the man, or a mythic narrative. Dehaene uses it as a hook, and it works as one, but nothing in the painting settles what it depicts.
Descartes's Challenge
The idea of a separate mind runs from Plato's Phaedo (fourth century BC) through Thomas Aquinas's Summa theologica (1265 to 1274), but René Descartes (1596 to 1650) stated it as dualism: the conscious mind is a nonmaterial substance outside the ordinary laws of physics. Dehaene pushes back against the fashion, popularised by Antonio Damasio's Descartes' Error (1994), of mocking Descartes. Descartes was a pioneering mechanist and reductionist; his model of the body was the first exercise in synthetic biology and theoretical modelling, and his dualism rested on a logical argument, not a whim. William James credited him with being first to conceive "a completely self-sufficing nervous mechanism".
In Description of the Human Body, Passions of the Soul and L'homme (published posthumously in 1664), Descartes treated humans as sophisticated automata built like church organs: bellows force "animal spirits" into reservoirs and pipes whose combinations produce behaviour.
- Vision and action. Light on the retina opens specific pipes; the pineal gland leans one way and sends spirits flowing to move the limb toward the object.
- Memory. Memory is the selective reinforcement of pathways, like punching holes in cloth, an anticipation of the modern rule that "neurons that fire together wire together".
- Sleep and waking. Wakefulness is high pressure of the spirits, which fill every nerve and leave the machine ready to respond; sleep is low pressure, moving only a few threads.
Descartes even closed with a lyrical materialist flourish: no "vegetative or sensitive soul" is needed to explain bodily functions, only blood and spirits heated by the fire in the heart. So why keep an immaterial soul? Because two capacities seemed beyond any machine:
- Reporting thoughts in language. A machine could be wired to cry out in response to a given input, but could it compose words or signs to declare its thoughts and answer a question "as even the dumbest person can"?
- Flexible reasoning. A machine acts rigidly "according to the disposition of its organs"; it seemed "morally impossible" that any machine could have enough organs to act appropriately in all the occurrences of life, as reason lets us do.
Dehaene treats both challenges as still standing and as the test any science of consciousness must pass: how a brain can express and reflect on its own states, and how it can decide flexibly and rationally.
The Last Problem
Physics (Euclid, Gauss, Einstein, Newton, Hubble) and biology (Darwin, Pasteur, Watson and Crick) have explained the cosmos and life as "just plain physics". Consciousness alone seemed to stay in medieval darkness; President Obama made the same point launching the BRAIN Initiative on April 2, 2013. Montaigne complained in 1580 that no two authorities agreed on the soul's seat: the brain's ventricles (Hippocrates, Herophilus), the whole body (Democritus, Aristotle), the stomach (Epicurus), the heart (the Stoics), the blood (Empedocles), every organ (Galen), between the eyebrows (Strato).
Through the nineteenth and twentieth centuries consciousness lay outside normal science. In 1962 George Miller, founder of cognitive psychology, proposed banning the word "for a decade or two". As a student in the late 1980s, Dehaene found the "C-word" forbidden in lab meetings; experimenters flashed stimuli at threshold without even recording whether subjects saw them. Then, from the late 1980s, everything changed: the field now has its own societies and journals.
Three Ingredients
- A better definition. Separate vigilance (the state of wakefulness, which changes as we fall asleep or wake), attention (focusing mental resources on a specific piece of information) and conscious access (some attended information entering awareness and becoming reportable to others). Only conscious access counts as consciousness proper; the other two are not sufficient.
- Experimental manipulation. Dozens of methods make a stimulus cross the border between seen and unseen, so its fate in the brain can be compared across the two cases.
- Subjective reports as data. What the participant says she saw defines the phenomenon; if the experimenter sees the image but the subject denies it, the trial is scored invisible.
Conscious access is the simple fact that, when awake, whatever we focus on may become conscious. Fully awake and attentive, we sometimes see an object and can describe it, and sometimes cannot because it was too faint or too brief; the brain may still process it unconsciously. Access is also the gateway to higher forms: the sense of self, and recursive consciousness in which the "I" comments on its own performance and knows when it does not know. Even these are now experimental; labs can induce an out-of-body experience inside an MRI scanner.
Dehaene rejects the philosophers' "phenomenal awareness" as a separate target: the claim that experiences have private qualia (the sharpness of tooth pain, the greenness of a leaf) that no neuronal description or verbal report can capture. A phenomenal consciousness distinct from access is, he says, "highly misleading" and a slippery slope to dualism. Start with access; once we explain how information becomes reportable, the "insurmountable" problem of ineffable experience will disappear.
Dehaene (access first). Explain how information becomes globally available and reportable; phenomenal consciousness distinct from access is an illusion of intuition, and the "hard problem" will dissolve once access is explained.
Block, Chalmers and the phenomenal camp. Ned Block's distinction between phenomenal and access consciousness (1995) holds that experience may overflow what we can access and report; David Chalmers's "hard problem" holds that explaining any function, access included, leaves open why it feels like anything.
To See or Not to See
Seeing an object feels instant, but behind it lies an avalanche of activity in billions of visual neurons that can take nearly half a second before consciousness kicks in. The second ingredient gives a handle on that chain. Experimenters can flash a word too briefly to be noticed, clutter a scene so one item always loses the competition, distract attention as magicians do, or show different images to the two eyes, so that perception alternates and never shows both. Seen and unseen images may differ minimally at input, yet the difference must be amplified inside the brain, because one can be reported and the other cannot. Finding where and when this amplification happens is the new science's goal.
The minimal-contrast strategy cracked the problem. Find well-matched conditions in which one leads to conscious perception and the other does not, then ask what differs in the brain. The daunting mystery becomes a tractable comparison between two sets of trials.
Turning Subjectivity into a Science
The third ingredient was the controversial one. Behaviourists led by John Broadus Watson (1878 to 1958) expelled introspection: psychology was to be a "purely objective experimental branch of natural science". Dehaene calls that dogma "dead wrong" because it conflates two things:
- Introspection as a method cannot be trusted: naive subjects cannot tell us how their minds work, and claims of flying to the ceiling or meeting a dead grandmother should not be taken literally.
- Introspection as raw data must be trusted: unless the subject is lying, the report reflects a genuine mental event that needs explaining. The person who reports an out-of-body experience really feels dragged to the ceiling.
Illusions, misperceived pictures, delusions and imaginings are therefore the field's best material, because only they separate the physical stimulus from the subjective percept. Researchers are happiest with a display that is sometimes seen and sometimes missed: record the report on every trial, sort trials into conscious and unconscious, and look for brain patterns that separate them.
Signatures and the Global Workspace
With these three ingredients, consciousness became normal experimental science. Two findings frame the book:
- Deep unconscious processing. A staggering amount of processing occurs beneath awareness, and imaging can track how far an unseen stimulus travels and where it stops.
- Signatures of consciousness. Using fMRI, EEG, MEG and intracranial electrodes over fifteen years, Dehaene's team found brain markers that appear if and only if a person is conscious of a picture, word, digit or sound. They change massively at the moment of awareness and are stable across vision, hearing, touch and cognition.
To explain the signatures, Dehaene and colleagues propose the global neuronal workspace (GNW): consciousness is global information broadcasting within the cortex, arising from a network whose raison d'être is the massive sharing of pertinent information throughout the brain. Daniel Dennett calls it "fame in the brain". Its circuit is a set of giant pyramidal cells with long axons that criss-cross the cortex and bind it into a whole. When enough regions agree that incoming information matters, they synchronise into global communication, an ignition that produces the signatures. Computer simulations of this architecture reproduce the main experimental findings.
Conscious access adds functions the unconscious lacks:
- An inner space for thought experiments, detached from the external world.
- Holding information for as long as we like.
- Passing it to any other process, the flexibility Descartes demanded.
- Chaining arbitrary operations, so information is pondered and redirected, not handled reflexively.
- Reporting to others through links to the language areas.
The workspace is also autonomous. The brain is constantly crossed by spontaneous global patterns generated by neurons' partly random self-activation, so it does not wait for input as Descartes's organ did. Even in darkness it broadcasts, producing William James's "stream of consciousness", shaped mostly by current goals. Dehaene calls the result a "free-willing" machine that meets Descartes's challenge.
The workspace under adversarial test. The Cogitate Consortium's adversarial collaboration, pre-registered with input from both camps, pitted GNW against integrated information theory (IIT) using fMRI, MEG and intracranial recordings in more than 250 participants; results were presented in 2023 and published in Nature in 2025. Neither theory came through cleanly. For GNW, the predicted "ignition" when a stimulus disappears was not found, and prefrontal cortex carried only limited information about some conscious contents (category was decodable, other features less so). For IIT, the predicted sustained synchronisation within posterior cortex was absent. The 2023 announcement also prompted an open letter signed by over 100 researchers calling IIT "pseudoscience", which itself drew sharp replies. The upshot for this book: the signatures are robust findings, while how much of them is consciousness itself rather than what follows it is now openly contested (see the no-report debate in Chapter 1).
The Future of Consciousness
Dehaene ends with stakes. Thousands of patients in coma or vegetative states lie motionless after stroke, accident or anoxia; some may be conscious but "locked in". His lab is building tests that tell reliably whether a person is conscious, already used in coma clinics, and the same signatures may answer whether and when infants are conscious. "No science will ever turn an is into an ought," but objective measures of subjective states should improve ethical decisions. Finally, better theory should allow chips that mimic the operation of consciousness, raising the question of machines that know their own knowledge, have a sense of self, or experience free will.
Machines and the workspace since 2014. Dehaene, Hakwan Lau and Sid Kouider sharpened the machine question in Science (2017), splitting consciousness into C1, global availability of information, and C2, self-monitoring (metacognition), and arguing that then-current deep learning had neither. After the rise of large language models, a 2023 report by Patrick Butlin, Robert Long and colleagues (including Yoshua Bengio) derived "indicator properties" from leading theories, with global workspace broadcasting among them, and concluded that no current AI system was a strong candidate for consciousness but that nothing technical blocked building ones that satisfy the indicators.
For architects of AI systems. Dehaene's definition is operational: consciousness as a limited-capacity broadcast hub that selects one item, holds it and makes it available to every specialised module. That is the pattern of an orchestrator with a shared blackboard over many specialist agents. The analogy is useful for design (a single shared state, explicit selection, a confidence tag on what gets broadcast), but meeting a functional description is not the same as having experience, which is exactly the gap the phenomenal camp presses.
Consciousness Enters the Lab
Consciousness became a science by narrowing the question to conscious access (why some sensations become conscious perceptions and others do not), by building minimal contrasts in which the same stimulus is sometimes seen and sometimes not, and by trusting subjective reports as data. Dehaene tours the toolkit (binocular rivalry, the attentional blink, inattentional and change blindness, masking) and ends with a research programme: find the objective signatures of subjective states.
A Dot That Vanishes
Stare at the cross in the figure below: after a few seconds some of the twelve grey dots fade out and later pop back, sometimes all at once. This is Troxler fading. The stimulus is constant; the experience is not, so something in the brain must change, and that change can be tracked. In the 1990s Francis Crick and Christof Koch realised that such illusions let scientists follow the fate of conscious versus unconscious stimuli in the brain: record neurons while the dots are seen, compare with moments when they are not. They chose vision because its pathways from retina to cortex were well mapped and because it offers countless illusions that contrast seen with unseen.
Three features put this programme within reach. The illusions need no elaborate notion of consciousness, only seeing or not seeing. Dozens of techniques exist to make words, pictures, sounds "and even gorillas" disappear. And although the illusions are purely subjective, they are reproducible: everyone reports the same kind of experience.
Sorting the Meanings of "Consciousness"
Stuart Sutherland's 1996 dictionary entry declared consciousness undefinable and added that "nothing worth reading has been written about it". Dehaene's reply is that science advances by carving new distinctions, as nineteenth-century physics did by separating heat (energy transferred) from temperature (mean kinetic energy). Melting ice absorbs heat at a fixed 0 °C; a firework spark at thousands of degrees carries too little heat to burn. "Consciousness" in daily speech is like the layman's "heat".
Conscious access
At any moment a massive flow of sensation reaches us but only a sliver is accessed: Dehaene drives past the same houses daily without noticing their roofs. Conscious access is extraordinarily open (any colour, scent, memory, strategy or error can become conscious with a switch of attention) and inordinately selective: we are reduced to about one conscious thought at a time, though one thought can be a sizeable "chunk", like a sentence's meaning. Notice the pressure on your legs: a second ago that information was preconscious, accessible but not accessed, and still used unconsciously to adjust posture. Access made it available to language, memory, attention, intention and planning at once. That switch is the book's subject.
Attention is not access
William James (1890) defined attention as "the taking possession by the mind, in clear and vivid form, of one out of what seem several simultaneously possible objects or trains of thought". Dehaene splits this into two processes with distinct brain mechanisms:
- Access. The "taking possession by the mind" is conscious access: the object becomes something we keep in mind and can report.
- Selection. Isolating one of many candidates is selective attention, a sieve that discards the irrelevant on the basis of salience or goal relevance and amplifies the winner.
Selection must work mostly unconsciously (we could never think if we first had to consciously review every candidate thought). Attention is often the gateway to consciousness, but the two dissociate: a lab display with one item needs almost no selection, while attention can covertly amplify or squash inputs that never reach awareness.
Vigilance, wakefulness and the self
- Vigilance (intransitive consciousness). "The wounded soldier remained conscious" names a graded state lost in sleep, fainting or anaesthesia. Dehaene separates wakefulness (the sleep-wake cycle, driven by subcortical mechanisms) from vigilance (the level of excitation in cortical and thalamic networks). Both are enabling conditions, necessary but not sufficient for access. A patient made colour-blind by a small visual-cortex stroke is awake and attentive yet cannot access colour; vegetative-state patients (Chapter 6) wake and sleep but seem to access nothing.
- The self. Antonio Damasio defines consciousness as "the self in the act of knowing". Gordon Gallup's mirror test (a red mark placed on the forehead, found using a mirror) is passed by children at 18 to 24 months, and claimed for chimpanzees, gorillas, orangutans, dolphins, elephants and magpies; the Cambridge Declaration on Consciousness (July 7, 2012) concluded that humans are not unique in possessing the neural substrates of consciousness.
- Why the mirror test is not a litmus test. An unconscious predictor that compares expected and seen body movements could pass it, as when one shaves without thinking. Pigeons can be trained to pass after extensive conditioning that turns them into mirror-using automata. And self-reference is unnecessary for consciousness: a sunset or a concert can absorb us without any "I am enjoying this". For Dehaene, self-consciousness is just conscious access to a representation of "me" (body, behaviour, feelings, thoughts).
- Metacognition. Self-reflection contains a "strange loop": the "I" is both perceiver and perceived. Auguste Comte (1798 to 1857) thought this impossible, since one organ cannot observe itself. John Stuart Mill answered that observer and observed can be different systems or different moments. Tip-of-the-tongue states, noticing a reasoning error and brooding over a failed exam are everyday metacognition; prefrontal regions monitor plans, attach confidence to decisions and detect errors, working as a closed-loop simulator with memory and imagination.
| Concept | What it means | Example of dissociation from access |
|---|---|---|
| Wakefulness | Sleep-wake cycle, subcortical | Vegetative patients wake but access nothing |
| Vigilance | Excitation of cortex and thalamus | Graded in drowsiness and anaesthesia |
| Selective attention | Unconscious sieve that amplifies one input | Attention can boost a stimulus that stays invisible |
| Conscious access | Information enters awareness, becomes reportable | The book's target |
| Self-consciousness | Access to a representation of "me" | Absorbed in a sunset with no thought of self |
| Metacognition | One system monitoring another | Tip-of-the-tongue, error detection |
Ingredient one is to start from conscious access and keep self and recursion for later.
Minimal Contrasts
Bernard Baars, in A Cognitive Theory of Consciousness (1989), argued that dozens of experiments already bore on consciousness and that the best offer a "minimal contrast": two situations that differ minimally but only one of which is consciously perceived. Conscious perception becomes an experimental variable while the stimulus stays nearly constant, and every brain operation common to both drops out of the comparison. Delimiting the unconscious draws consciousness "as in a photographic negative".
Contrasts run in both directions:
- Conscious to unconscious. Learning to type begins slow and self-conscious, then becomes automatic; comparing the two phases identifies a large cortical network, notably prefrontal regions that activate whenever conscious access occurs.
- Unconscious to conscious. Visual illusions push a stimulus into awareness. Troxler fading probably occurs because the visual system treats a perfectly stable image as a defect of the eye, like the blood vessels in front of the photoreceptors, and fills it in from surrounding texture, the same filling-in that hides the blind spot. A tiny eye movement shifts the dots on the retina, signals that they are external, and they pop back.
Rival Images
Each eye receives a slightly different image; the brain fuses them and uses the disparity to see depth, as Charles Wheatstone showed in 1838. Wheatstone then asked what happens with completely different images (a face to one eye, a house to the other) and built the mirror stereoscope to find out, starting a Victorian craze for stereo pictures. Vision became unstable: face for a few seconds, then house, then face, with brief transitions. "It does not appear to be in the power of the will to determine the appearance." This is binocular rivalry: two incompatible images fight for consciousness.
Rivalry is "an experimenter's dream": the stimulus is constant, the percept changes, and the same image is alternately fully visible and fully suppressed. David Leopold and Nikos Logothetis trained monkeys to report their percept with a lever, showed that monkeys experience the same alternations, and recorded single neurons:
- Early cortex. In primary visual cortex, many cells followed the physical stimulus and did not change when the animal's percept switched.
- Higher areas. In V4 and inferotemporal cortex, more and more neurons agreed with the report, firing strongly when their preferred image was seen and little or not at all when it was suppressed. This was "the first glimpse of a neuronal correlate of conscious experience".
Hundreds of rivalry experiments have followed. Continuous flash suppression keeps one image invisible indefinitely by flashing bright colourful rectangles to the other eye. The lesson: an image can sit in the eye for a long time and enter the visual brain yet be wholly suppressed; what matters for consciousness is a later stage at which a single winner emerges, because consciousness cannot hold two objects at one location. Countless potential percepts compete; only one wins at a time.
Rivalry needs attention
Trying to favour the face lengthens its dominance only slightly: the fight starts at stages beyond voluntary control. But the existence of a single winner depends on attention; "the fighting arena itself is made up of the conscious mind". Testing this looks circular (asking a distracted person what she sees forces her to attend). Peng Zhang and colleagues at the University of Minnesota solved it with frequency tagging: each image flickers at its own rate, and both rates appear in the EEG. During attended rivalry the two tags alternate (one strong, the other weak); when attention is withdrawn, the alternation stops and the tags coexist. A second study found that the image seen on returning attention differs from what continued alternation would predict. Rivalry requires an attentive observer.
When Attention Blinks
Attention's limited capacity yields more contrasts. In the attentional blink, a stream of digits appears at one location at one item every 100 ms, with two letters to remember. The first letter (T1) is easy. A second (T2) arriving half a second or more later is also reported, but at short lags it is often missed entirely, and viewers are surprised to learn there were two. Attending to the first creates a "blink of the mind".
Imaging shows that all the letters enter the brain, reach early visual areas and may even be classified as targets, yet the missed one never reaches awareness. Registering a letter in awareness passes through a stage that handles one chunk at a time: binocular rivalry is competition in space, the blink is competition in time. "Conscious access imposes a narrow bottleneck."
The psychological refractory period
Sometimes both letters are seen (about a third of the time at the worst lags), and in daily life we hear a car horn while looking at a picture. Dual tasks do not refute the bottleneck; they reveal it. We never consciously process two unrelated items at exactly the same moment: one is accessed while the other waits in an unconscious buffer, producing a measurable delay, the psychological refractory period (PRP).
- The wait is invisible. Because consciousness is busy, we cannot notice the delay. Asked when the second item appeared, we postdate it to when it entered consciousness; two simultaneous inputs feel successive, the attended one first.
- The wait is risky. Noise, distracting thoughts or new inputs can erase the buffered item: that is the blink. In dual tasks both occur, and the longer the delay, the higher the chance of total blackout.
Refractoriness and blinking are one mechanism. While consciousness is occupied, other candidates queue in an unconscious buffer; delay is the cost when they survive, invisibility when they do not.
Inattentional and Change Blindness
A longer distracting task makes us oblivious to the world for longer, as absorbed readers, chess players and mathematicians know. In a classic lab demonstration of inattentional blindness, participants fixate the centre and attend to the top of the screen, where a letter will appear. After two practice trials, on the third an unexpected shape (a dark spot, digit or word) appears at the centre for nearly a second. Up to two-thirds fail to notice it.
Dan Simons and Christopher Chabris's "invisible gorilla": viewers count passes by the team in white (nearly everyone gets fifteen in a roughly thirty-second film), while an actor in a gorilla suit walks in, beats his chest and leaves. Many viewers miss it and are so sure that they accuse the experimenter of switching films.
- Change blindness (Kevin O'Regan and Ron Rensink). Two versions of a photo alternate about once a second with a brief blank; viewers swear they are identical even when a jet loses an engine or a road's centre line changes from broken to continuous.
- The live swap (Dan Simons). An actor asks a Harvard student for directions; workers pass between them; two seconds later a different actor, with different hair and clothes, continues the conversation. Most students miss the swap.
- Choice blindness (Peter Johansson). A man picks the more attractive of two female faces; while the card is face down the experimenter swaps it. Half the participants do not notice and justify the choice they never made.
- Whodunnit? A one-minute detective film for Transport for London contains twenty-one changes made by five assistants (a stuffed bear becomes a suit of armour, coats and props are swapped); it ends with the mayor's warning to look out for cyclists.
- Pilots. In flight simulators, pilots absorbed in talking to air traffic control can become so oblivious that they crash into a plane they never detected.
Some numbers in this tour are generous. In Simons and Chabris's 1999 study, about half of observers missed the gorilla overall, more in the harder condition; "a majority" holds only for some conditions. In Simons and Levin's 1998 live-swap study, roughly half of the people approached missed the swap, not "most". The phenomena are robust; the proportions depend heavily on task, instructions and viewer.
Masking: Total Invisibility
Inattention has two lab problems. It is labile: once viewers suspect a trick, they watch for it, so it cannot be repeated over hundreds of trials. And it is hard to prove the change was unseen: questioning after every trial alerts subjects, while questioning at the end lets forgetting contaminate the answer. Some argue change-blind viewers see the whole scene but fail to store the details for comparison. Dehaene doubts forgetting explains everything (a gorilla is memorable) but admits "a lingering doubt". What is needed is an image 100 percent invisible however informed or motivated the viewer: masking, or subliminal presentation (Latin limen, "threshold").
Making an image faint degrades it too much to evoke brain activity. Instead flash it briefly, for 33 ms (about one movie frame), sandwiched between masks at the same location. Brevity alone is not the cause (a microsecond flash in darkness is visible); the masks are. The word radio between geometric shapes simply is not there for the viewer, though a photocell confirms it was on screen.
- 33 ms: the typical subliminal flash, about one movie frame.
- Below about 30 ms: all participants deny seeing the word on every trial.
- 40 ms vs 60 ms: typically invisible at 40, seen on most trials at 60.
- About 50 ms: the threshold at which the same image is seen half the time; it varies a little across people.
- 37 words, up to 20 repetitions: masked words in a Dehaene experiment left no recognition memory.
The threshold is sharp, which justifies the words subliminal and supraliminal: a flashed image is metaphorically either in or out. At threshold the stimulus is constant and only perception varies, the ideal design.
- Masked emotions. A smiling or fearful face can be hidden between scrambled images, letting researchers probe unseen emotion.
- Substitution masking. In an array of shapes, the one cued by four long-lasting dots vanishes while the rest remain visible; the dots and the space they enclose replace the shape in consciousness.
- The race. With one target and one trailing mask, the target alone would always be seen; with the mask, the mask catches up and erases it. One account treats the brain as a statistician weighing evidence for one item or two: with a short target and strong mask, the evidence favours "only the mask".
The Primacy of the Subjective
Dehaene's lab now simply asks after each trial whether a word was seen. Critics call this "too subjective"; Dehaene replies that subjectivity is the subject matter. Three arguments back the reports:
- Agreement. Masking produces consistent reports across viewers; only the minimal duration needed varies.
- Objective consequences. When subjects report nothing, they usually cannot name the word; forced to guess, they do slightly better than chance (a sign of subliminal processing, Chapter 2). They cannot judge whether a masked digit is larger or smaller than 5 a few seconds later, nor recognise masked words seen up to 20 times.
- A step change in information. Near threshold, trials labelled conscious come with a massive increase in available information: we can name, rate, judge and memorise what we see far better. Observers are neither random nor whimsical in their reports.
In 2001 Dehaene and Lionel Naccache wrote what became the field's manifesto: subjective reports are the key phenomena that a cognitive neuroscience of consciousness purports to study, primary data to be recorded alongside physiological measures. The aim is signatures of consciousness: neuronal events that unfold whenever a person reports a given conscious state.
Introspection as data, not as method
Introspection is not a window on mechanism. A patient who sees faces in the dark is not taken literally, but his experience is real and needs explaining, perhaps by spontaneous, possibly epileptic activity in temporal-lobe face circuits. Normal introspection errs too:
- Reading. People think they recognise words as wholes by shape, but the brain runs a series of letter-by-letter analyses they never notice.
- The stockings. Shoppers judging four identical pairs of nylon stockings strongly preferred the pair on the right and explained their choice by fabric quality, never mentioning position (the classic study by Richard Nisbett and Timothy Wilson, 1977).
So the behaviourists were right that introspection is a poor method, but as a measure it is "the perfect, indeed the only, platform": it supplies how the subject feels, and neuroscience must supply which neural events underlie it. Researchers actively seek "purely subjective" situations in which stimulation is constant or absent and perception varies, since the neural source is easier to isolate when no sensory confound comes with it.
Olaf Blanke's out-of-body experiences. Some surgical patients report floating at the ceiling looking down at their body. Hiding drawings on top of cupboards to "verify" this is pseudoscience; the right question is what brain dysfunction produces it. Across many neurological and surgical patients, the Swiss neurologist found that impairing or electrically stimulating the right temporoparietal junction repeatedly caused out-of-body sensations. That region merges visual, somatosensory and kinaesthetic, and vestibular signals into a representation of where the body is; when they conflict, self-location goes wrong, an exaggerated form of seasickness. Blanke then induced the illusion in healthy people with synchronised but displaced visual and touch signals, and with a robot inside an MRI scanner found activation in the temporoparietal junction, close to the patients' lesions. Out-of-body flight "really" happens, in the brain.
Reports may contaminate the signatures: the no-report turn. Since 2014 a major line of work has asked whether contrasting reported seen and unseen trials isolates consciousness or also picks up what the report requires (attention, decision, working memory, motor planning). Jaan Aru, Lucia Melloni and colleagues (2012) warned that the contrast mixes prerequisites, correlates and consequences. Frässle, Einhäuser and colleagues (2014) tracked binocular rivalry without any report, reading the percept from reflexive eye movements (optokinetic nystagmus) and pupil size, and found that frontal activation largely dropped out when subjects did not report. Michael Pitts and colleagues (2012, 2014) showed that a late P3b wave appeared for seen shapes only when they were task-relevant, whereas an earlier posterior negativity tracked seeing regardless. Naotsugu Tsuchiya, Victor Lamme and colleagues (2015) codified "no-report paradigms", and Koch, Tononi and colleagues argued for a posterior "hot zone" instead of prefrontal cortex. The workspace camp replied with report-free signatures of its own: Dehaene's colleagues Claire Sergent and colleagues (2021) reported a late, bifurcating brain response separating seen from unseen stimuli that did not depend on report, and monkey studies have decoded rivalry contents from prefrontal neurons without report. The debate is open, and it bears directly on the P3b and ignition signatures of Chapter 4.
The Programme
The tools are now in hand to make words, pictures and whole films vanish and reappear with minimal changes, or none. Two questions follow, the ones Descartes would have asked:
- What happens to an unseen image? Is it processed, for how long, how far into the cortex, and does the answer depend on how it was made invisible?
- What changes when it becomes conscious? Are there brain events unique to conscious access, signatures that can ground a theory?
| Paradigm | How awareness is removed | Competition type | Strength | Weakness |
|---|---|---|---|---|
| Troxler fading | Steady fixation; filling-in | Stimulus vs fill-in | Constant stimulus | Slow, uncontrolled timing |
| Binocular rivalry / CFS | Different images to the two eyes | Spatial, between images | Same input, changing percept; works in monkeys | Needs attention; slow switches |
| Attentional blink | Second target within ~200 to 500 ms of first | Temporal, between targets | Many trials, precise timing | Partial; T2 seen on ~a third of trials |
| PRP / dual task | Second task queued behind first | Temporal bottleneck | Measures the delay directly | Delay, not invisibility |
| Inattentional blindness | Attention on another task | Attention vs unexpected item | Dramatic, real-world | Works once; hard to verify |
| Change blindness | Blank or interruption between scenes | Memory comparison | Robust | Could be a failure of memory, not seeing |
| Masking | 33 ms flash between masks | Target vs mask race | Total invisibility, precise control | Very brief stimuli only |
For anyone evaluating models or people from self-reports. Dehaene's split between introspection as method (unreliable) and introspection as data (indispensable) is a good rule for explainability in AI: a model's or a user's stated reasons are evidence about the report, not about the mechanism. The stockings study is the human version of a post-hoc rationalisation; treat explanations as outputs to be validated against behaviour, the way Dehaene validates "seen" reports against naming and memory.
Fathoming Unconscious Depths
An invisible image travels much deeper into the brain than anyone expected: unseen words and digits are recognised, bound, understood, valued, attended to and even fed into decisions, and practically every brain region can work unconsciously. Mapping these powers draws a photographic negative of consciousness: whatever the unconscious can do is not what consciousness is for. Dehaene also marks the limits, since unconscious effects are small, short-lived and confined to local circuits, which sets up the next chapter.
Subliminal Images in Public Life
- RATS (2000). A George W. Bush campaign advert about Al Gore's prescription plan flashed RATS in huge capitals as the tail end of "bureaucrats". Not strictly subliminal, it went largely unnoticed and raised the question whether it reached voters' emotional centres.
- Mitterrand (1988). The face of François Mitterrand, president and candidate, was flashed daily inside the logo of France's main 8 p.m. news programme. In a nation of 55 million, even a tiny bias would mean thousands of votes.
- The Coca-Cola myth (1957). James Vicary's claim that a frame reading "Drink Coca-Cola" in a film boosted sales was a complete fabrication that he later admitted. The myth survives; so does the scientific question.
The question matters for science as much as politics. If early perception runs without awareness, it can be excluded from the search for consciousness; extending that elimination upward, delineating the contours of the unconscious progressively prints a negative photograph of the conscious mind.
Pioneers of the Unconscious
Crediting Sigmund Freud (1856 to 1939) with discovering the unconscious is a myth Freud largely made himself. The historian Marcel Gauchet calls the claim that the mind had been identified with consciousness before psychoanalysis "rigorously false".
- c. 460 to 377 BC Hippocrates, in The Sacred Disease (on epilepsy), notes the body acting against its owner's will and declares that from the brain alone arise our joys, sorrows and judgements.
- c. 129 to 270 Galen and Plotinus observe that walking and breathing run without attention.
- 965 to 1040 Alhazen (Ibn al-Haytham) treats the eye as a camera obscura that receives light and posits unconscious inference: the brain jumps to conclusions beyond the data, sometimes seeing what is not there.
- 354 to 1716 Augustine, Aquinas, Descartes, Spinoza and Leibniz describe motives hidden from introspection; Spinoza lists a baby's desire for milk, the injured man's wish for revenge, the drunkard's craving and the chatterbox's compulsion.
- 1790 to 1857 Marshall Hall defines the reflex arc from spinal cord to muscle.
- 1835 to 1911 John Hughlings Jackson describes a hierarchy from brainstem automatism to voluntary, conscious cortex.
- 1867 Hermann von Helmholtz's Physiological Optics revives the term unconscious inference.
- 1868 Henry Maudsley: "the most important part of mental action... is unconscious mental activity."
- Late 1800s In France, Théodule Ribot (action memory), Gabriel Tarde (unconscious imitation) and Pierre Janet (subconscious goals from childhood); Janet claims priority over Freud.
- 1890 William James: never take a person's sincere testimony that he felt nothing as proof that no feeling was there.
- 1899 Freud's Vienna colleague Sigmund Exner: say not "I think" but "it thinks in me", two decades before Freud's The Ego and the Id (1923).
Against this background Freud's contribution looks speculative: "the ideas that are solid are not his own, while those that are his own are not solid." He never tested his models although experimental psychology with reaction times and error rates was being born around him. Nabokov mocked those who believe mental woes are cured by "a daily application of old Greek myths to their private parts".
Older still. The "it thinks" formula predates Exner: the eighteenth-century physicist Georg Christoph Lichtenberg wrote that we should say "it thinks" as we say "it lightens". Freud took the term das Es from Georg Groddeck, noting that the usage went back to Nietzsche. This strengthens Dehaene's point. His verdict on Freud is deliberately one-sided: it judges psychoanalysis only as an empirical theory of unconscious processing, not as the clinical and cultural movement that put the unconscious into general discussion.
Where Does Unconscious Processing Happen?
In the 1990s three "simplistic dichotomies" competed:
- Cortex conscious, subcortex unconscious. Unconscious work would be confined to nuclei like the amygdala (fear) and superior colliculus (eye movements).
- Left hemisphere conscious, right unconscious, because the left hosts language and can report.
- Ventral stream conscious, dorsal stream unconscious. The ventral route (object and face identity) would be conscious; the dorsal route through parietal cortex (shape and location for action) would be the unconscious "dark side".
None survived. Virtually every brain region can take part in both conscious and unconscious processing.
Evidence from patients
- The amygdala's fast track. The almond-shaped amygdala flags emotionally charged events, especially fear; a snake can activate it via a fast route from the retina before the emotion registers consciously.
- Claparède's pin (early 1900s). Édouard Claparède pricked an amnesic patient with a pin hidden in his hand; next day she did not recognise him but refused to shake his hand, an unconscious emotional memory.
- Blindsight. Patients with lesions of primary visual cortex (V1) deny seeing anything in the affected field yet point accurately to objects and flashes there, "to see, but not to see". The first fMRI study of blindsight found strong activation in the superior colliculus, but also in the cortex; later work confirmed that invisible stimuli activate the thalamus and higher visual areas, bypassing V1. The inner zombie is not only subcortical.
- Patient D.F. (Melvyn Goodale and David Milner). At 34, carbon monoxide poisoning destroyed both lateral visual cortexes, leaving visual form agnosia: she could not tell a square from an elongated rectangle or report a line's orientation. Yet she posted a card through a slanted slot perfectly and scaled her grip to objects, though she could not show their size with finger and thumb on request. Her intact occipito-parietal (dorsal) pathway computed size, orientation and shape she could not see.
- Spatial neglect. After lesions near the right inferior parietal lobe, patients ignore the left side of space; one complained he had not been given enough food after eating only the right half of his plate. Their retinas and early visual cortex are intact.
The burning house. John Marshall and Peter Halligan (1988) showed neglect patient P.S. two drawings of a house, identical except that one had flames on the left. P.S. insisted they were the same, but asked which he would prefer to live in, he consistently avoided the burning one. His brain categorised the fire as danger without awareness. Later imaging showed that neglected stimuli still activate ventral visual regions for faces and houses, and that even the meaning of neglected words and numbers gets through.
What changed in these cases. The "low road" to the amygdala has been challenged: Luiz Pessoa and Ralph Adolphs (2010) argued that fast threat detection relies largely on cortical routes and that evidence for a functional subcortical shortcut in humans is weak. For blindsight, monkey work by Michael Schmid and colleagues (2010) showed that residual vision after V1 loss depends on the lateral geniculate nucleus of the thalamus, which projects directly to extrastriate areas such as motion area MT, supporting Dehaene's point that the bypass reaches cortex. The burning-house result was less consistent in later neglect patients tested by Edoardo Bisiach and Anna Rusconi (1990), some of whom preferred the burning house; the broader evidence that neglected stimuli are processed comes mainly from imaging and priming.
Subliminal Priming in the Healthy Brain
Lesioned brains might have altered the boundary between conscious and unconscious, so Dehaene's lab turned to normal brains and subliminal priming. A prime (say house) is masked to invisibility, then a visible target (radio) is classified, for instance as living thing or artefact. When prime and target are the same word, responses are faster and more accurate, as long as they are less than a second apart: subliminal repetition priming, replicated in dozens of experiments. The unseen word primes the pump.
- Abstract letter identity. Priming is just as strong across case (radio priming RADIO) as for identical words, although "a" and "A" share no shape. In expert readers the cultural convention linking them has become unconscious and compiled into the visual system.
- Where. fMRI showed unconscious activation over much of the ventral visual cortex, including the fusiform gyrus, which houses advanced shape recognition and the early stages of reading, and which responded to word identity regardless of case. This undid the idea that the ventral route is necessarily conscious: both streams can run below awareness.
Binding without consciousness
A widely held view (the "object file" or binding hypothesis) held that features such as colour and motion float free without attention and that binding, through neuronal synchrony or reentry, was the hallmark of consciousness. Some binding is in fact unconscious. range primes RANGE but anger does not, although they share all five letters: subliminal priming is sensitive to letter order, and anger primes no better than an unrelated word like tulip. Words that share 80 percent of their letters do not fool it.
Such results have been replicated hundreds of times with faces, pictures and drawings. The visual scene we experience is heavily processed: the retinal image is distorted, magnified at the centre, crossed by blood vessels, holed by the blind spot and constantly smeared by eye movements, yet we see a stable 3D world, corrected, filled in and reinterpreted in the light of experience, all unconsciously.
Chess, voices and two kinds of binding
- Chess. Grand masters parse a board at a glance into meaningful chunks. A simplified position flashed for 20 ms between masks still influences an expert's decision about whether the king is in check, but only for experts and only for a meaningful question. The visual system binds piece identities and positions into "black king under check" unconsciously.
- The McGurk effect (Harry McGurk and John MacDonald, 1976). Lips saying ga dubbed with the sound ba are heard as da. The brain fuses sight and sound unconsciously, weighting each by reliability, and we just hear the compromise. Imaging suggests the conscious outcome is represented in frontal cortex and then sent back to sensory areas. We hear neither sound waves nor see photons; we access an expert reconstruction.
- Routine versus novel binding. Wolf Singer proposed two kinds. Routine bindings are coded by dedicated neurons tuned to frequent combinations (letter strings like the, un, tion, familiar objects, chess patterns), learned from the statistics of experience, and they survive anaesthesia. Non-routine bindings create new combinations and may need a conscious state of synchrony: a sentence like "Time flies like an arrow; fruit flies like a banana" binds its words for the first time, and anaesthesia sharply reduces the brain's ability to integrate words into sentences.
A misattributed joke. Dehaene gives "fruit flies like a banana" to Groucho Marx, a common attribution with no good source; the sentence circulated as a linguistics and machine-translation example (often credited to Anthony Oettinger in the 1960s) of syntactic ambiguity.
Can Meaning Be Unconscious?
If consciousness is "the perception of what passes in a man's own mind" (John Locke, 1690), comprehension seems to require it; comprehension ("together-catching") and consciousness ("together-knowing") are near-synonyms. Yet when reading we never consciously compute each word's meaning; a glance places it in the gist. Two generations fought over this "like mad dogs".
- The cocktail party (1950s). Donald Broadbent proposed an early filter that blocks unattended voices before meaning. But your own name, spoken quietly behind you, grabs attention, so unattended speech was processed to the level of meaning, and unattended words can bias judgements of the attended conversation. Still not watertight: listeners might switch attention briefly.
- Anthony Marcel (1970s, Cambridge). Using masking, Marcel made words invisible on every trial (subjects could not even say whether a string was a word). An unseen colour word (blue, red) sped choice of the matching colour patch by about 1/20 of a second (50 ms).
- All meanings at once. Consciously, bank means money or river depending on context (river before bank primes water; save before bank primes money). When bank was masked, Marcel found both meanings primed even when context favoured one. "The unconscious mind proposes while the conscious mind selects."
The great unconscious wars
Sceptics (the best known is Daniel Holender's 1986 critique) struck back. Effects were tiny, sometimes under 10 ms, and might come from a few trials on which the prime was glimpsed. A "didn't see anything" at the end of the session is not enough; prime visibility must be measured objectively, in a separate test under identical conditions, and only chance performance counts. Marcel's controls did not always meet this standard.
Greenwald undoes his own Science paper. In 1996 Anthony Greenwald (Seattle) reported in Science that masked emotional words primed positive-or-negative judgements of visible words: congruent pairs (happy then joy) beat incongruent ones (rape then joy) when subjects answered very fast. Subjects could not identify the primes, and priming did not depend on how well they guessed them. Then Greenwald and his student Richard Abrams tested a worry: with a small set of words repeated under a 400 ms deadline, subjects might link letters, not meanings, to responses. Indeed, scrambled hypap primed as well as happy. And tumor, built from the letters of the consciously classified positive words tulip and humor, primed a positive response. Priming was subliminal but shallow: letter fragments mapped to responses.
Unconscious Arithmetic
Greenwald's deadline may have been too short for meaning. In 1998 Dehaene and Lionel Naccache chose the simplest meaningful words: numbers below ten.
- Design. Digits 1, 4, 6, 9 were masked by random letter strings; a visible target number followed; subjects judged larger or smaller than 5. Separately, even when told of the hidden digit, they could not see it or classify it.
- Semantic priming. Congruent prime and target (both above 5) gave faster responses than incongruent ones: an unseen 9 sped responses to 9 and 6 and slowed them to 4 and 1.
- Motor cortex. EEG and fMRI showed a small activation of the motor cortex for the hand the invisible prime called for: unconscious votes travelled from perception to motor preparation.
The sceptics' remaining objections fell one by one:
- Notation-free. four primed 4 as much as 4 did; an invisible visual prime even primed a spoken target.
- Never-seen primes. Numbers never shown consciously in the experiment still primed, ruling out learned shape-to-response links (Greenwald's problem).
- Number-sense regions. fMRI showed repetition suppression (adaptation) in the left and right parietal "number sense" areas, which encode quantity, when the same number appeared twice (nine then 9), even though the first was unseen.
- The distance effect. Priming declined with numerical distance: strongest for four before 4, less for three, less again for two. Only a representation of quantity produces this.
Beyond numbers, and into the brain
- Other categories. Deciding that piano is an object is helped by an unseen chair and hindered by an unseen cat, even with primes never seen consciously.
- The amygdala reads. Electrodes implanted in the amygdala of epilepsy patients responded to masked threatening words (rape, danger, poison) but not neutral ones (fridge, sonata), slowly, half a second or more after the word, while the patient denied seeing anything.
- The N400. A word that does not fit its context ("coffee with cream and socks") evokes the N400, a negative wave peaking around 400 ms, larger the more absurd the word. It occurs for words made invisible by masking or by inattention.
- Unconscious phrases (Simon van Gaal and Dehaene). Two masked words formed "very happy", "not happy", "very sad", "not sad", then a visible war or love. The N400 to war was large after happy, modulated up by very and down by not: the brain unconsciously registered the incongruity of a "very happy war". Dehaene calls it as close as one gets to unconscious syntax and meaning in a phrase.
Unconscious can mean intense but local. The N400 can be the same size for conscious and invisible words, so consciousness is sometimes irrelevant to semantics. But source reconstruction places unconscious semantic activity within the left temporal lobe, whereas conscious words recruit much larger networks reaching the frontal lobes. Unconscious words are less influential because they stay local.
Attentive but Unconscious
Under competition, consciousness requires attention (the gorilla), but the converse fails: attention operates unconsciously. It must: selecting one of hundreds of candidates is better done by massively parallel, sub rosa workers, "armies of unconscious workers" that sift rubble until one strikes gold.
- Unseen cues capture attention. A flash in the corner of the eye, too brief to see, still makes you faster and more accurate at that location.
- Unseen distractors are worse. An irrelevant hidden picture slows you more than a visible one, because a conscious distractor can be voluntarily suppressed and an unconscious one cannot.
- A watchdog. Screams and your own name break through because the selective filter must keep monitoring everything while you do your taxes.
Psychologists had thought unconscious processing was passive "spreading activation", untouched by strategy or intention. Dehaene's lab "shattered this consensus":
- Attention gates subliminal priming. A prime at an unexpected time or place produces virtually no priming. Attention multiplies the brain response at the attended time and place, and unconscious stimuli benefit as much as conscious ones: attention can amplify a stimulus and still leave it below awareness.
- Intentions steer unconscious attention. Told to find squares, not circles, with both masked and the subject guessing at chance, the N2pc (a parietal marker of attentional orienting) still shifts toward the square.
- Targets get extra processing in the blink. During the attentional blink, the letter designated as target evokes more brain activity than distractors even when missed.
The value of an invisible coin
Attention depends on valuation, handled by the basal ganglia, and valuation too is unconscious.
Penny or pound (Mathias Pessiglione and colleagues, 2007). Each trial began with a coin image showing the stake, a penny or a pound, displayed for 17, 50 or 100 ms between masks; then subjects squeezed a handgrip to win money. Even when they could not tell which coin had appeared, they squeezed harder for a pound, their palms sweated in anticipation (skin conductance), and a reward region of the basal ganglia (the ventral pallidum) activated. Motivation was manipulated without their knowing why.
In a follow-up, values were learned unconsciously. A "signal" contained a masked shape: one shape meant go pays, another withhold pays, a third was 50/50. Within minutes subjects began earning money without seeing the shapes, and the ventral striatum attached values to each. Unseen symbols became attractive or repulsive, shaping attention and action. The brain's subliminal tags turn raw stimulation into a map of opportunities ranked by relevance.
Unconscious Mathematics
"We are conscious only of our conscious thoughts", so we overestimate consciousness. Julian Jaynes (Princeton): consciousness is a smaller part of mental life than we are conscious of, "because we cannot be conscious of what we are not conscious of". Dehaene adapts Hofstadter's Law into a rule of his own:
We constantly overestimate our awareness, even when we are aware of the glaring gaps in our awareness.
Henri Poincaré reported two flashes of insight: stepping onto an omnibus at Coutances he saw that the transformations defining Fuchsian functions were those of non-Euclidean geometry; walking a seaside bluff, that the arithmetic of indefinite ternary quadratic forms was too. Jacques Hadamard built these into a model of discovery:
- Initiation. Deliberate, conscious attack on the problem, often fruitless, but it launches the unconscious.
- Incubation. An invisible brewing period with no conscious sign of work.
- Illumination. After sleep or a walk, the solution bursts into consciousness, usually correct.
- Verification. Slow, effortful, conscious checking of every detail.
Hunches and distraction
The Iowa gambling task (Antoine Bechara, University of Iowa). Players get a $2,000 loan in fake money and draw from four decks; each card wins or costs money (say $100). Two decks are bad: large early gains, larger later losses, net loss. Two are good: modest swings, small steady gain. Players eventually report which decks are bad. But in the "pre-hunch" period, while still sampling all decks and claiming no clue, their hands sweat before they draw from a bad deck (a skin-conductance response). The alarm probably comes from the ventromedial prefrontal cortex, whose activation predicts performance; patients with lesions there show no anticipatory response, only one after the loss is revealed.
Deliberation without attention (Ap Dijksterhuis). Students chose among four cars described by twelve attributes. Half deliberated consciously for four minutes; half were distracted with anagrams. The distracted group chose the best car far more often: 60 percent versus 22 percent (chance 25 percent). In the field, IKEA shoppers who deliberated hard were less satisfied weeks later than impulsive buyers. Distraction does not guarantee no conscious thought, Dehaene notes, but the result is "very suggestive".
Both tasks require weighing many values, which overloads conscious working memory (deliberators over-weight one or two features) but suits unconscious circuits that sum and average.
- Monkeys sum evidence. Monkeys learn to decide from the summed weights of arbitrary shapes, and parietal neurons track the running total (Tianming Yang and Michael Shadlen, 2007).
- Masked arrows. With five arrows, subjects judged whether more pointed right or left. Masked, they felt they were guessing but did better than chance, and parietal signals showed unconscious approximate summation.
- Eight numerals. With four visible and four invisible digits and the task of judging whether the mean exceeded 5, responses were biased by the invisible ones: averaging extended to unseen inputs.
Two of these findings have not held up well. The unconscious-thought advantage has a poor replication record: meta-analyses found weak or inconsistent effects, and a large preregistered replication by Mark Nieuwenstein and colleagues (2015) found no advantage of distraction over deliberation. For the Iowa gambling task, Tiago Maia and James McClelland (2004) showed that with more detailed questioning, players already had conscious, reportable knowledge of the good decks at about the time their choices improved, which undercuts the "decide before knowing" reading.
Statistics During Sleep
Sleep inspires insight (Ullrich Wagner, Jan Born and colleagues, 2004). Subjects transformed seven-digit sequences by a laborious rule and reported only the last digit. A hidden shortcut existed: the last three digits mirrored the previous three (e.g. 4 1 4 9 9 4 1), so the final digit always equalled the second. Few found it at first; a night's sleep more than doubled the proportion who discovered the rule (about 59 percent versus 23 percent). Elapsed time alone did not; sleep did.
A candidate mechanism: in rats, hippocampal and cortical neurons replay waking sequences during sleep, compressed in time and sometimes in reverse, precisely enough to decode where in a maze the animal is mentally travelling. Compression might turn a sequence into a near-simultaneous pattern in which hidden regularities can be learned. Sleep is "a period of boiling unconscious activity" that consolidates memory and supports insight.
A Subliminal Bag of Tricks
Even the "central executive", thought to be the province of consciousness, partly runs unconsciously:
- Unconscious inhibition (Simon van Gaal). In a stop-signal task, a masked stop signal slowed responses and occasionally stopped them altogether; the wave reached the executive control networks. "Invisible is not synonymous with out of control."
- Unconscious error detection. When the eyes stray in a saccade task, anterior cingulate control centres respond even when subjects deny the error.
- Unconscious task switching. A masked task-switch cue slows subjects and triggers a partial switch in cortex.
Most such processes launch but do not run to completion.
| Domain | Experiment (researchers) | Key result |
|---|---|---|
| Word form | Masked repetition priming (Dehaene lab) | radio primes RADIO; fusiform activation, case-invariant |
| Letter binding | range vs anger | Order-sensitive; anagram does not prime |
| Chess | 20 ms masked boards | Experts detect check unconsciously |
| Multisensory | McGurk and MacDonald (1976) | ba + ga heard as da |
| Colour words | Marcel (1970s-1983) | ~50 ms speed-up; both meanings of bank primed |
| Emotional words | Greenwald (1996); Abrams and Greenwald | Priming real but letter-based (tumor primes positive) |
| Numbers | Dehaene and Naccache (1998 on) | Congruity, notation-free, distance effect, parietal adaptation, motor activation |
| Fear words | Amygdala electrodes | Response to unseen danger, ~500 ms+ |
| Phrases | van Gaal and Dehaene | N400 to war tuned by unseen very/not |
| Attention | Masked cues; N2pc | Unseen cues orient attention; goals steer it |
| Value | Pessiglione (2007, 2008) | Unseen pound boosts grip; unseen shapes acquire value |
| Hunches | Bechara (Iowa gambling) | Skin conductance before bad-deck choices |
| Averaging | Masked arrows; 4 + 4 digits | Unseen items enter the sum |
| Insight | Wagner and Born (2004) | Sleep more than doubles hidden-rule discovery |
| Control | van Gaal; error and switch studies | Masked stop signals, unnoticed errors, partial switches |
Replication concerns about subliminal priming. The core masked-priming effects in this chapter (repetition priming, number congruity) have replicated well. Bolder claims have fared worse. Mudrik and colleagues' (2011) finding that scene-object incongruity is processed under continuous flash suppression failed to replicate (Moors and colleagues, 2016); claims that people read multi-word expressions and solve subtraction problems under CFS (Sklar and colleagues, 2012) met failed or weaker replications. Methodologists such as David Shanks and Miguel Vadillo showed that awareness checks in this literature are often underpowered, that discarding "aware" participants or trials post hoc inflates unconscious effects, and that near-chance awareness scores in small samples cannot establish zero awareness. Meyen and colleagues (2022) found that many reported cases where indirect effects exceed direct awareness measures do not survive proper statistics. Separately, high-profile social priming results outside this book (e.g. elderly-word priming of walking speed) failed replication (Doyen and colleagues, 2012). The current consensus: unconscious processing is real but often shallower and smaller than the most striking 2000s claims, and good studies now use trial-by-trial graded visibility ratings and well-powered awareness tests.
Poincaré's Question, and Its Limits
Poincaré (Science and Hypothesis, 1902) asked whether "the subliminal self" is superior to the conscious self, since it "succeeds where that has failed". Dehaene answers yes in many respects: our visual system solves recognition problems that defeat the best software. But he rejects the opposite extreme, that consciousness is frosting on the cake, a powerless backseat driver, as in The Matrix (1999). Consciousness is an evolved function that must fill a niche the parallel unconscious cannot.
Poincaré himself saw the limits: the unconscious did not engage without a massive conscious initial attack, and only consciousness could verify the insight step by step. Henry Moore (The Sculptor Speaks, 1937) said the same of artists: the conscious mind "resolves conflicts, organizes memories, and prevents him from trying to walk in two directions at the same time."
For decision systems in banking. Dehaene's catalogue is a reminder that fast parallel scoring (the unconscious "army of statisticians") is powerful at weighing many features into one value, but its outputs are small, short-lived and local unless something selects and commits to them. That maps onto the design split between models that score and workflow layers that decide, record and chain steps. The Dijksterhuis replication failure is also a caution: "trust your gut over deliberation" has weaker evidence than its popularity suggests.
What Is Consciousness Good For?
Against the view that consciousness is a useless epiphenomenon, Dehaene argues that the minimal contrast between seen and unseen reveals specific operations only conscious information supports. Consciousness collapses the unconscious brain's probability distributions into a single sample fit for action, holds information over time, chains operations like a slow serial computer, and, through language, shares a compressed, confidence-tagged summary with other minds.
Function, Design and Evolution
Pre-Darwinian biology was teleological: for Georges Cuvier, claws were "for" catching prey and lungs "for" breathing. Darwin replaced design with natural selection: organs are not designed for anything; they merely give a reproductive advantage. Opponents cited apparently useless features, the peacock's tail and the antlers of the extinct Irish elk (Megaloceros), spanning up to twelve feet and once blamed for its extinction; Darwin answered with sexual selection. Modern evolutionary-developmental biology adds tools that explain design without a designer:
- Spontaneous pattern generation. Alan Turing showed how chemical reactions can generate zebra stripes or a melon's ribs; some cone shells grow elaborate pigment patterns hidden under an opaque layer, so they cannot be for anything.
- Allometric relations. A useful increase in body size can drag along a proportionate change in an organ that is not itself useful; the Irish elk's antlers probably arose this way.
- Spandrels. Stephen Jay Gould's term for necessary by-products of an architecture, sometimes later co-opted ("exapted") for a new role; the male nipple is a by-product of the body plan for female breasts.
So consciousness need not be useful: it might be a decorative pattern, a side effect of the growth of the Homo brain, or a spandrel. Alexandre Vialatte* quipped that consciousness, like the appendix, "serves no role but to make us sick"; the puppeteer in Being John Malkovich (1999)* calls it "a terrible curse".
- Max Velmans. Many cognitive functions are indifferent to awareness; they would run as well in a zombie.
- Tor Nørretranders. Our sense of control is a "user illusion"; every decision stems from unconscious sources.
- Others. Consciousness is the backseat driver, observing actions it cannot control.
Dehaene takes the functionalist road: conscious perception turns information into an internal code that permits unique operations, and so it was probably selected for over millions of years. Evolution cannot be rewound, but minimal contrasts can show which operations are possible only when we report awareness.
The "backseat driver" and readiness potentials. The strongest empirical prop for the epiphenomenalist view has been Benjamin Libet's (1983) finding that the readiness potential begins several hundred milliseconds before people report deciding to move, extended by Chun Siong Soon, John-Dylan Haynes and colleagues (2008), who decoded upcoming choices from fMRI seconds early at modestly above chance. In 2012 Aaron Schurger, Jacobo Sitt and Dehaene proposed that the readiness potential largely reflects spontaneous neural noise accumulating toward a threshold, averaged backward from the movement, not an unconscious decision already taken. That reading has gained wide support and weakens the "decided before you know it" interpretation, though how conscious intentions arise remains debated.
Unconscious Statistics, Conscious Sampling
Dehaene's picture is a division of labour: in the basement, an army of unconscious workers sifts data; at the top, a small board of executives reads a brief and slowly decides. The unconscious processors act as near-optimal statisticians, like a weather bureau combining dozens of observations into a chance of rain, computing the probability that colours, shapes or people are present. Consciousness offers only a sample from that distribution: one best interpretation, handed to decision systems.
Any acting organism may need this, because no one can act on mere probabilities. "Alea jacta est", said Caesar crossing the Rubicon: action needs a point of no return. Consciousness may be the brain's scale-tipping device, collapsing probabilities into one sample so we can move on. Buridan's ass, starving midway between water and hay, shows the cost of not choosing.
Perception as abduction
Charles Sanders Peirce, after Helmholtz, saw that even "I see an azalea" is an abduction: what we perceive is an image made intelligible by a hypothesis, and "the whole fabric of our knowledge is one matted felt of pure hypothesis". Modern cognitive science calls this Bayesian inference (after Thomas Bayes, c. 1701 to 1761): reasoning backward from observations to hidden causes. Classical probability asks how likely three aces are from a fair deck; Bayesian reasoning asks, given three aces, how likely it is that the deck was rigged. A plate's rim projects as an ellipse compatible with countless potato-shaped objects; we see a circle because the visual brain picks it as the most probable cause.
The aperture problem
Neurons in motion area MT see through a small receptive field. Through such a peephole a moving bar's direction is ambiguous (it could move perpendicular to itself or in countless other directions): the aperture problem. Individual neurons suffer from it; consciousness never does. We see the bar moving perpendicular to itself, the minimal motion, and a moving rectangle as one rigid motion that satisfies the constraints from all its edges.
- Convergence takes time. For about a tenth of a second MT neurons encode only local motion; after 120 to 140 ms they switch to the global direction. We see only the end result.
- Convergence needs consciousness. Under anaesthesia, this convergence vanishes: neurons stop exchanging bottom-up and top-down signals until they agree, and inference stops short of a single interpretation.
Sampling in proportion to probability
Two superimposed gratings moving in different directions can be seen with either one in front. We never see a blend; perception alternates every few seconds. Alexandre Pouget and colleagues varied speed and spacing and found that the time spent in each interpretation matches its likelihood given the evidence. Our unconscious perception works out the probabilities, and then our consciousness samples from them at random.
Sampling requires attention. Binocular rivalry stops when attention is withdrawn (Chapter 1). Unconscious processes therefore approximate the true distribution, while consciousness "shamelessly" reduces it to all-or-none samples; in that sense the unconscious is more objective.
- A quantum analogy. As measurement collapses a wave function (no half-dead Schrödinger's cat), attending to an object collapses the distribution of its interpretations. Dehaene flags this as possibly superficial; the neural mechanism is almost certainly classical physics.
- Language. Both meanings of bank are primed unconsciously; we are aware of one at a time.
- Attention. Selection feels like one spotlight but is probabilistic, weighing several hypotheses.
- Memory. Guess the share of the world's airports in the United States, then discard it and guess again. Your second guess is not random, and the average of the two beats either (work by Edward Vul and Harold Pashler, 2008, the "crowd within"). Conscious retrieval draws samples from a hidden distribution.
Consciousness as the spokesperson. Like the FBI, which knew more than any agent could grasp yet failed to surface the hints before September 11, 2001, a brain of a hundred billion neurons needs a brief and a single spokesperson. The conscious brief must be stable, integrated and written in an abstract "language of thought" that can interface with intention and decision.
Lasting Thoughts
Daniel Dennett (Kinds of Minds, 1996): the thoughts "that persist the longest, acquiring influence as they persist, we call our conscious thoughts." Henri Bergson (1911): consciousness is "the hyphen which joins what has been to what will be". A compressed, unambiguous code is compact enough to carry forward in time into working memory. Organisms with active memory are freed from the present: a predator that disappears behind a rock must be remembered. A transient-memory mechanism exists in humans, monkeys, cats, rats and mice.
Working memory is a dominant function of the dorsolateral prefrontal cortex and its partners. These regions activate whenever we hold a phone number, a colour or a flashed shape; prefrontal neurons keep firing long after the stimulus is gone, sometimes for dozens of seconds, and when prefrontal cortex is impaired or distracted, the memory falls into oblivion.
- General picture. Lack of foresight, stubborn adherence to the present, failure to inhibit (utilisation behaviour: seizing and using tools; imitation behaviour: compulsively copying others), poor planning, and in severe cases apathy.
- Hemineglect. Disturbed awareness of one half of space, usually the left.
- Abulia. Inability to generate voluntary actions.
- Akinetic mutism. No spontaneous verbal reports, though repetition may be intact.
- Anosognosia. Unawareness of a major deficit, including paralysis.
- Impaired autonoetic memory. Inability to recall and analyse one's own thoughts.
- Perception. Tampering with prefrontal cortex can disrupt even perceiving and reflecting on a brief visual display.
Trace conditioning (Robert Clark and Larry Squire). An air puff to the eye causes a blink; a tone that predicts it comes to trigger the blink (Pavlovian conditioning). In delay ("coincidence-based") conditioning the tone lasts until the puff, so learning is coincidence detection. In trace ("memory-trace") conditioning a brief tone is separated from the puff by an empty gap, so the organism must hold the tone in memory. Results: coincidence-based conditioning is unconscious, and works even in a decerebrate rabbit with no cortex, basal ganglia, limbic system, thalamus or hypothalamus. Memory-trace conditioning needs the hippocampus and connected structures, including prefrontal cortex, and in humans occurs if and only if the person becomes aware of the tone-puff relation. Elderly people, amnesics and distracted subjects fail at trace but not delay conditioning, and those who gain awareness are those who activate prefrontal cortex and hippocampus.
The paradigm suggests consciousness evolved for learning across time rather than living in the instant, giving us what Gerald Edelman called "the remembered present". Because it is simple, the test works across species: mice performing trace conditioning activate anterior regions homologous to human prefrontal cortex, hinting that this function may be widespread.
The unconscious decays
The influence of a subliminal stimulus can be timed by waiting until it decays to zero. A visible image has long-lasting effects; a masked one activates visual, orthographic, lexical or semantic codes for about a second, then fades. Lionel Naccache, reversing Jacques Lacan, put it: "the unconscious is not structured as a language but as a decaying exponential." Effortful maintenance stretches it slightly, but recall after a few seconds barely beats chance.
Two claims that later work complicates. First, working memory without awareness: David Soto, Teemu Mäntylä and Juha Silvanto (2011) found above-chance memory for unseen gratings over several seconds, and Dehaene's own group (Darinka Trübutschek and colleagues, 2017) showed that unseen stimuli can be held for several seconds in an "activity-silent" form (short-term synaptic changes, not sustained firing), and proposed a theory of working memory without consciousness or sustained activity. That softens this chapter's sharp line, though conscious maintenance remains more robust and flexible. Second, trace conditioning as a test of awareness: Tristan Bekinschtein, Mariano Sigman and colleagues (2009) found trace conditioning in some vegetative and minimally conscious patients and read it as residual awareness. Using a behaviour both as evidence of consciousness and as the thing consciousness explains risks circularity, and the human awareness-conditioning link has itself been debated.
The Human Turing Machine
Compute 12 × 13 in your head. You can usually report every step (Dehaene recalls 12² = 144 and adds 12), and the report is accurate, cross-validated by response times and eye movements. Such accurate introspection is rare: we have no insight into recognising faces or naming words. Dehaene's explanation: algorithms, strings of elementary steps, are a uniquely conscious function. Flashed subliminally, 12 × 13 could never be computed.
What is needed is a router that stores intermediate results and dispatches them to the next step. Consciousness collects information from processors, synthesises it and broadcasts a conscious symbol to arbitrarily chosen processors, which apply their unconscious skills; the cycle repeats. The result is a hybrid serial-parallel machine: parallel unconscious computation interleaved with serial conscious decisions and routing.
With Mariano Sigman and Ariel Zylberberg, Dehaene likened this to a production system, a 1960s AI architecture:
- Working memory. A database holding the current state.
- If-then rules. A large set of productions (e.g. if A is in working memory, replace it with BC).
- Matching. At each step, all rules check whether they match the current state.
- Stochastic competition. If several match, they compete under a prioritising system (unconscious competition).
- Ignition. The winner fires, changes working memory (conscious ignition and broadcasting), and the cycle resumes.
Production systems can implement any effective procedure; they are Turing-equivalent. Alan Turing (1936) answered David Hilbert's 1928 challenge (can a mechanical procedure decide whether a statement follows from axioms?) by modelling "a man in the process of computing a real number". Relying on introspection, Turing captured only the mathematician's consciously accessible operations, which is why, Dehaene contends, the serial Turing machine models conscious thought reasonably well.
- Not the brain-as-computer cliché. The brain is massively parallel, self-modifying and computes over distributions, not discrete symbols; only its behaviour during long calculations looks serial.
- Additive timing. Time to compute 235 + 457 is the sum of the steps (5 + 7; carry; 3 + 5 + 1; 2 + 4).
- Leaky stages. Steps overlap and cross-talk. In Jérôme Sackur's task (take n, add 2, decide whether n + 2 > 5), subjects began comparing n itself with 5 before finishing the addition. A computer's master clock prevents such errors; a brain evolved for a probabilistic world "recycles" its networks for arithmetic and pays in errors.
One step yes, two steps no
If consciousness is the brain's lingua franca for routing, then a single routine operation may run unconsciously but a chain cannot: 3 + 2 perhaps, but not (3 + 2)², (3 + 2) − 1 or 1/(3 + 2).
Sackur and Dehaene's two-step test. A digit n was masked so it was seen about half the time. In separate blocks subjects named it, computed n + 2, compared n with 5, or did the two-step n + 2 > 5. On unseen trials the single operations were above chance: naming was correct on nearly half of trials against 25 percent chance with four digits, and n + 2 and n > 5 also beat chance. But the two-step task fell to chance, although merely naming the digit to themselves and using the name would have given high accuracy. Subliminal information was present but could not be passed through two stages.
This does not contradict unconscious summation of arrows, digits or car features (Chapter 2): accumulating evidence is a single operation. An open accumulator can be biased by conscious or unconscious inputs; what the unconscious never seems to do is reach the threshold, commit, and pass the decision on. Hence no unconscious strategy: with five conscious arrows, people stop once three point the same way; with masked arrows they just add them all.
The mighty unconscious generates sophisticated hunches, but only a conscious mind can follow a rational strategy step after step. By routing information through any sequence of processes, consciousness gives the brain a whole new mode of operation.
A Social Sharing Device
Nietzsche (The Gay Science, 1882) called consciousness "a connecting network between man and man". Other animals may have used consciousness as a nonverbal buffer and router for millions of years; in Homo language, pointing and gesture let a conscious synthesis jump between minds, enabling "multicore" social algorithms that pool viewpoints and expertise.
- Reportability as criterion. We do not usually credit someone with awareness unless they can at least partly put it into words (barring paralysis, aphasia or infancy); the "verbal formulator" works only on conscious content.
- Overflow. "Consciousness overflows language": a Caravaggio, a Grand Canyon sunset, a baby's face exceed description. Still, whatever we are aware of can be at least partly framed linguistically.
- Compression. Speech carries only 40 to 60 bits per second, about 300 times slower than a 14,400-baud fax. So the brain compresses experience into short strings of symbols, and what others need is not my viewpoint but a multisensory, viewer-invariant, durable synthesis.
- Gossip. Robin Dunbar estimates that about two-thirds of conversation concerns social topics and proposed the "grooming and gossip" theory of language as a bonding device. Dehaene grants the point but argues conversation also carries what collective decisions need.
Two heads are better than one (Bahador Bahrami and colleagues, 2010). Pairs judged which of two displays contained a near-threshold target, first independently; when they disagreed, they discussed and gave one joint answer, which could be scored like a single observer. If the two partners' sensitivities were reasonably similar, the pair beat its best member. The optimal rule weights each person's signal inversely by their noise, the same law that governs multisensory integration in one brain, and it can be approximated by sharing a categorical answer plus a confidence level. Successful pairs did exactly that, using words like "certain", "very unsure", "just guessing", some even inventing numeric scales, and approached the theoretical optimum.
Hence confidence sits at the centre of conscious thought: every conscious content carries a certainty label, and we track who said what and whether they were right, which is what makes gossip useful. Bayesian decision theory applies the same rule to our own thoughts and to others' reports: weight each source by its reliability. The primate prefrontal workspace already combined past and present evidence weighted by reliability; a perhaps uniquely human step opened it to inputs from other minds.
| Function | What consciousness adds | Evidence in the chapter |
|---|---|---|
| Sampling / collapse | One unambiguous interpretation, fit for action | Aperture problem; MT convergence lost under anaesthesia; plaid alternations track likelihood (Pouget); airport guesses |
| Lasting thought | Information held as long as needed | Prefrontal persistent firing; trace conditioning (Clark and Squire); subliminal exponential decay (Naccache) |
| Serial chaining | Results routed step to step; strategies | 12 × 13; additive timing; n + 2 > 5 fails unconsciously (Sackur and Dehaene); arrow strategy |
| Social sharing | Compressed, confidence-tagged brief | 40 to 60 bits/s speech; Bahrami's dyads; shared self/other networks |
Self as another
Social cognition activates the frontal pole and ventromedial prefrontal cortex, with the temporoparietal junction and precuneus, a network tied together by long fibre tracts with prefrontal cortex as hub. It is part of the "default mode" system that turns on at rest, when we spontaneously return to social tracking. Many of these regions activate both when we judge our own confidence and when we think about others' thoughts; frontal pole and ventromedial PFC respond so similarly that thinking about one primes the other. Each of us represents "oneself as another" (Paul Ricoeur).
The self, then, is a statistical inference: we watch our own behaviour as we watch others' and "make up our mind". We enjoy privileged access to conscious motives and strategies, but remain ignorant of our unconscious determinants and cannot predict ourselves outside familiar situations. "Know thyself" stays an inaccessible ideal; the self is a database in the same format as our model of others, with the same gaps and delusions. David Lodge's novel Thinks... has an AI entrepreneur tell a novelist, in a hot tub, that selves exist but "we make them up all the time".
Partial self-delusion may be the price of sharing confidence-weighted knowledge. That ability built alphabets, cathedrals, jet planes and lobster Thermidor, and also let us fake, forge and lie. Nabokov: literature was born not when a boy ran from the Neanderthal valley crying "wolf" with a wolf behind him, but when he cried "wolf" with no wolf there. Consciousness is the mind's virtual-reality simulator.
Since 2014. The sampling account has been developed as a general theory (Adam Sanborn and Nick Chater, 2016, "Bayesian brains without probabilities"), which explains biases as the result of drawing few samples. Later large-scale studies confirmed the "crowd within" benefit of averaging one's own repeated guesses, though it is much smaller than averaging with another person. Estimates of speech information rate have converged near Dehaene's figure: Christophe Coupé and colleagues (2019) measured about 39 bits per second across 17 languages, and Jieyu Zheng and Markus Meister (2024) argued that human behavioural throughput in general is only about 10 bits per second, against vastly higher sensory input, sharpening the "compressed brief" argument. Meanwhile the Cogitate adversarial collaboration (published in Nature, 2025) found that prefrontal cortex carried less information about some features of conscious content than the workspace view predicted, which bears on how much of the "conscious brief" is written in prefrontal cortex.
Chain-of-thought as a conscious bottleneck. Dehaene's "human Turing machine" anticipates a lesson from large language models: a parallel network that answers in one pass handles single-step pattern problems well but stumbles on multi-step ones, and performance on those improves when the model writes intermediate results into a serial scratchpad and feeds them back (chain-of-thought and reasoning models). The analogy is functional, not a claim about machine experience. For enterprise agent design the transferable rules are Dehaene's: commit at each step (a decision, not a running score), pass explicit intermediate state, and attach calibrated confidence to every handoff, as Bahrami's dyads did, so that combining human and model judgements approaches the optimum instead of following the louder voice.
The Signatures of a Conscious Thought
Comparing seen and unseen versions of the same stimulus reveals four reliable signatures of conscious access: a late, self-amplifying ignition of parietal and prefrontal cortex; a slow P3 wave that begins about a third of a second after the stimulus; a late, sustained burst of high-frequency (gamma) activity; and long-distance synchrony that binds distant regions into a "brain web". Single neurons in the anterior temporal lobe fire for what the person reports seeing, not for what hits the retina, and stimulation studies show that late, long-distance loops, not early local activity, cause and can abolish conscious experience.
From Fortress to Glass Skull
Marcel Proust called a person "a shadow which we can never penetrate", and Julian Jaynes pictured consciousness as "a secret theater of speechless monologue". Dehaene's opening point is that in about twenty years the fortress was breached by three techniques, combined with the minimal contrasts of Chapter 1 so that consciousness is the only manipulated variable.
- Functional MRI (1990). Invented by Seiji Ogawa and colleagues. Active neurons draw a blood-flow surge two or three seconds later; deoxygenated haemoglobin acts as a small magnet while oxygenated haemoglobin does not, and the scanner reads these distortions. It localises activity at millimetre resolution but cannot time neuronal firing.
- Electroencephalography (EEG). The 1930s recording of brain waves, now with up to 256 electrodes and millisecond resolution over the whole head.
- Magnetoencephalography (MEG). From the 1960s: ultrasensitive detectors pick up the minuscule magnetic fields of cortical currents, again with millisecond timing.
- Intracranial electrodes. Placed in epilepsy patients for clinical monitoring, they give a direct view of cortical activity and, with fine wires, of single neurons.
With masking or inattention a visible image can be made to vanish, or placed at threshold so that it is seen half the time. Sorting identical trials by report asks a clean question: which circuits, waves or oscillations appear only on conscious trials? A marker present whenever perception is conscious, like a signature on a document, would be a signature of consciousness.
Signature 1: The Conscious Avalanche
In 2000 Kalanit Grill-Spector, then at the Weizmann Institute, flashed pictures for one-fiftieth to one-eighth of a second followed by a scrambled mask. Images shown for under 50 ms were very hard to see; those shown for 100 ms or more were visible. Primary visual cortex and its neighbours responded to all images regardless of masking, but in higher visual centres (fusiform gyrus, lateral occipitotemporal region), which build invariant representations of faces, objects, words and places, activation tracked conscious report.
Dehaene's own whole-brain scans of masked words complicated the picture. Even the fusiform gyrus activated without consciousness, and abstract operations in temporal and parietal cortex ran subliminally (recognising that piano and PIANO are the same word, or that 3 and three mean the same quantity). Yet crossing the threshold changed everything:
- Amplification. In the visual word form area, activation was multiplied by twelve on conscious trials.
- Invasion. A whole set of regions absent on unconscious trials appeared in the parietal and frontal lobes, down to the anterior cingulate on the midline.
- Progressive gain. The masking-induced loss grew stage by stage. Subliminal perception is a surf wave that looms on the horizon but licks your feet at the shore; conscious perception is an avalanche, a snowball that gathers mass until it triggers a landslide. Words flashed for only 43 ms could still end in a major multi-region activation.
- Correlation. On conscious trials distant regions peaked and receded together, suggesting mutually reinforcing message exchange; synchrony was much stronger for conscious targets.
The pattern generalises beyond vision. In a noisy scanner, sound pulses set so that only half are detected produce activity near primary auditory cortex when missed; when heard, the same avalanche breaks into inferior parietal and prefrontal areas. In a go/no-go task with the no-go cue made subliminal on half the trials, responses still slow slightly: unconscious inhibition recruits the pre-supplementary motor area (pre-SMA) and anterior insula. When the cue is visible, activation in those two regions nearly doubles and a much larger parietal and prefrontal network joins in.
Signature 2: The P3 Wave
Timing the avalanche
fMRI says where, not when. With Claire Sergent, Dehaene used EEG and the attentional blink: participants memorised a first letter string (T1), which briefly occupied them, and then had to detect a target word (T2). After each trial they reported visibility with a continuous cursor (nothing, a few letters, most of the word, all of it). Parameters were tuned until the same words were seen on half the trials. Reports were all-or-none: people saw the word perfectly or claimed there was no word at all, rarely anything in between.
Sergent and Dehaene's attentional-blink EEG. The early waves were identical for seen and unseen words: the P1 (positive, about 100 ms) and the N1 (negative, about 170 ms) reflected normal entry into a hierarchy of visual areas, just as strong when the word went unreported. Between 200 and 300 ms the traces split: activity faded on unseen trials and advanced toward the front on seen trials. By about 400 ms only conscious words drove intense activity in both frontal lobes, the anterior cingulate and parietal cortex. After more than half a second, activity returned to the back of the brain, including primary visual cortex (a backward wave of unknown meaning, perhaps a sustained memory of the percept). Unseen words did not vanish at once: they reverberated for about half a second in the left temporal lobe, at sites tied to word meaning, which is how blinked words can still prime their meanings (Chapter 2).
In less than a tenth of a second, between 200 and 300 ms, the recordings went from no difference to a massive all-or-none effect. The wave either broke through the dike of the frontal and parietal networks or stayed contained in posterior systems.
The P300
That breakthrough produces a scalp marker. On conscious trials only, a large positive voltage sweeps across the top of the head, starting around 270 ms and peaking between 350 and 500 ms. It is the P3, or P300: the third large positive peak, often starting near 300 ms. It is only a few microvolts, "a million times smaller than an AA battery", but easily measured. The P3 is the second signature, recorded across many paradigms whenever access suddenly occurs.
The recordings also explained why targets were missed. There were two P3 waves, one for the distractor string (always seen) and one for the target when seen, and they traded off: when the first P3 was large and long, the second was likely to be absent, and those were the missed trials. When the second target was seen despite this, its P3 was sharply delayed, the refractory period of Chapter 1. Descartes had noted that "we cannot be very attentive to several things at once", blaming a pineal gland that could lean only one way. The localisation was wrong; the observation was right: the conscious brain cannot host two ignitions at once.
Consciousness lags behind the world
A direct consequence: we are aware of unexpected events at least a third of a second late, and more than half a second late when a faint input needs slow evidence accumulation (an astronomer's long exposure for faint stars), with further delay when the mind is busy elsewhere. Even hitting the brakes at a car's taillights slows when you are on the phone. We rarely notice: a sunset or a symphony does not care about half a second. Only acting "in real time" exposes the lag: a pianist cannot consciously steer each finger in an allegro, and by the time you press the shutter, the lizard's tongue is gone.
- Unconscious autopilot. Fast sensorimotor loops act before awareness: a burnt finger withdraws before pain is felt (Descartes again), and a reaching hand corrects for a target jump before any change is consciously detected. Conscious intentions set up these loops; they then run unconsciously.
- Anticipation. Sensory and motor areas learn temporal regularities and predict events, so predictable stimuli are perceived closer to real time. The cost appears in the flash-lag effect: a flash aligned with a moving dot seems to lag behind it, because the predictable dot is perceived sooner than the unpredictable flash.
The lag becomes vivid only when anticipation fails, as when you drop a glass of milk and feel consciousness grope after the event.
Errors become conscious the same way
In an antisaccade task (look away from a flash), the eyes are often first pulled toward it, sometimes without the person noticing. For the first fifth of a second, the cortex responds identically to noticed and unnoticed errors: an autopilot in the cingulate gyrus fires vigorously whenever the motor plan departs from instructions (the error-related negativity), even when the error remains unconscious. Only when the error is consciously detected does a late positivity follow, the error positivity (Pe), which is virtually indistinguishable from the P3. Actions and sensations reach awareness by the same two-step route: unconscious appraisal, then late conscious ignition.
Isolating the Conscious Moment
A sceptic can object that seeing something triggers report, memory, evaluation and action, so ignition might reflect those consequences rather than access. Dehaene walks through a sequence of controls, each answering the devil's advocate's next move.
- Matched tasks. In the blink study, participants had to decide whether a word appeared, and saying "no word" is at least as hard as saying "word"; both answers used the same kind of key press.
- Temporal dilution. A "seen" decision has a sharp onset, while "unseen" can only be decided at the end of the trial. Hakwan Lau and Richard Passingham removed this by exploiting blindsight: they equalised objective performance on conscious and unconscious trials, and higher visibility still meant stronger left prefrontal activation. The blindsight patient G.Y. showed a full parietal and prefrontal pattern on aware trials.
- Same stimulus and same response. Moti Salti, with Dominique Lamy, sorted trials by two responses (where was the flash among four locations, and did you see it or guess?). Blindsight produced many "unaware, correct" trials to set against "aware, correct" ones. Early activity up to about 250 ms was strictly identical; after 270 ms the P3 became massively larger on aware trials and spread into both frontal lobes, whereas unaware trials produced only a small posterior parietal positivity.
Salti's control is the cleanest form of the contrast: identical stimulus, identical correct response, different awareness, and the only difference is a late, frontal, bilateral P3. He also rejected the idea that unaware trials were a mixture of random guesses and normal P3s: the unaware P3 was too small, too short and too posterior, an avalanche that started and fizzled.
Signature 1 Revisited: Ignition as a Phase Transition
Dehaene calls the sudden burst "global ignition", after Donald Hebb (The Organization of Behavior, 1949), who showed how mutually exciting neurons fall into a self-sustaining "cell assembly", like an audience going from a few handclaps to broad applause. Large pyramidal neurons in the upper cortical layers broadcast excitation widely; when it exceeds a threshold and becomes self-reinforcing, strongly interconnected neurons jump into a reverberating high-activity state. Physicists call such a change a phase transition (water freezing), mathematicians a bifurcation: a threshold at which minute noise decides the outcome.
Del Cul and Dehaene's phase-transition test. Antoine Del Cul flashed a digit for one video frame (16 ms), then a blank, then a letter mask (250 ms), varying the blank in 16 ms steps (16 to 100 ms), like slowly cooling a vial of water. Visibility did not improve smoothly; it showed a step, with a threshold near 50 ms. In EEG, the P1 and N1 did not change with delay. The next visual waves grew in proportion to the delay (linear accumulation, cut off by the mask), and were still strong on trials where the digit was denied. From about 270 ms, activity diverged in an all-or-none way: a P3-like avalanche whose size was out of proportion to the small change in delay, arising from occipital, parietal and prefrontal nodes in both hemispheres, although the digit had been flashed to one side.
Early visual waves (P1, N1, later visual components). Identical or linear in the stimulus; strong even when the person sees nothing; their shape does not match the step in subjective report.
Late ignition (from ~270 ms). Nonlinear and all-or-none; bilateral, frontoparietal; its size tracks subjective visibility, not the physical delay. Only this matches what people say they saw.
Deep Inside the Brain: Signatures 3 and 4
Intracranial avalanches
Intracranial recordings in epilepsy patients brought the same picture closer to neurons. Pooling ten patients, Dehaene's team (the study led by Raphaël Gaillard) tracked a masked word along the visual pathway. Seen and unseen traces began alike, diverged quickly and, after about 300 ms, differed massively: on unseen trials frontal activation was virtually absent; on seen trials it was hugely amplified. Because electrodes were placed only for clinical reasons, their sampling was effectively random, which makes one number telling: nearly 70 percent of sites responded to conscious words, against 25 percent for unconscious words. Unconscious information stays confined to a narrow circuit, while consciously perceived information is broadcast to most of the cortex for an extended time.
Signature 3: late gamma
Brain rhythms come in bands: alpha (8 to 13 Hz), beta (13 to 30 Hz) and gamma (30 Hz and above). Any word, seen or not, evoked a burst of gamma activity in the first 200 ms; gamma typically reflects local neuronal discharges. On unseen trials it died out; on seen trials it was sustained, and by 300 ms the difference was all-or-none. Rafi Malach's group at the Weizmann Institute found the same with masked pictures. A late, sustained increase in gamma power from about 300 ms is the third signature.
This revises an old idea. In the 1990s Francis Crick and Christof Koch speculated that consciousness rides on 40 Hz oscillations circulating between cortex and thalamus. That was too strong: unconscious stimuli also induce high-frequency activity across the whole gamma band, as any active cortical circuit sculpted by inhibition does. What marks consciousness is late amplification of gamma, not its presence.
Signature 4: the brain web
Synchrony probably exists because it eases transmission: in a forest of randomly firing cells, a small assembly that shouts in unison is more likely to be heard, and neurons that oscillate together share windows of opportunity to receive each other's signals. Several teams found that massive long-distance synchronisation is a fourth signature. About 300 ms after an image, distant electrodes lock together, only if the image is seen. Invisible images produce brief synchrony restricted to the back of the brain. The long-range coupling usually appears at lower frequencies, beta (13 to 30 Hz) or theta (3 to 8 Hz), presumably because slow carriers bridge the conduction delays over several centimetres.
Frequency analysis is not the whole answer: brain activity rarely oscillates at a pure frequency, spans broad bands, and shows nesting (high-frequency bursts locked to phases of slower waves). Dehaene therefore used Granger causality, devised in 1969 by the British economist Clive Granger for economic time series.
- Definition. Signal A "Granger-causes" signal B if A's past improves the prediction of B's present beyond what B's own past provides.
- Bidirectional. Nothing stops A from causing B while B causes A.
- Why it matters here. Most long-range cortical connections are bidirectional, and top-down projections often vastly outnumber bottom-up ones, so the direction of flow during access is an open question.
Applied to intracranial data, the analysis showed a "causal explosion" on conscious trials at about 300 ms: most recording sites joined a tangled web of relations, running mostly forward (visual cortex to frontal lobe) but also top-down. The descending wave can be read as attention amplifying the input or as confirmation that it fits the higher-level interpretation. The best summary is a "distributed attractor": a large-scale pattern that briefly sustains itself by reverberation. On unconscious trials the web never ignited; there was only a short period of causal links in ventral visual cortex, ending near 300 ms and dominated by top-down signals, as if anterior regions were "desperately interrogating" sensory areas that failed to answer consistently.
| Signature | What it is | When | Best evidence in the chapter |
|---|---|---|---|
| 1. Ignition / amplification | Self-amplifying spread into parietal and prefrontal cortex | Late; all-or-none | fMRI of masked words (×12 in VWFA), sounds, no-go cues; Del Cul's step |
| 2. P3 (P300, P3b) | Large slow positive wave on top of the head | Starts ~270 ms, peaks 350 to 500 ms | Sergent (blink), Del Cul (masking), Salti (same stimulus and response), Pe for errors |
| 3. Late gamma | Sustained high-frequency power | From ~300 ms | Intracranial words (Gaillard), pictures (Malach) |
| 4. Long-distance synchrony | Phase locking and bidirectional Granger causality across distant regions | From ~300 ms, beta/theta carriers | Face synchrony; Granger "causal explosion" |
The P3b and the timing of awareness since 2014. Chapter 1's update box covers the no-report challenge: Michael Pitts and colleagues showed the P3b can vanish when seen stimuli are task-irrelevant, leaving an earlier posterior negativity around 200 ms (the visual awareness negativity, studied by Mika Koivisto and Antti Revonsuo; Dembski, Koch and Pitts, 2021, generalised it as a "perceptual awareness negativity") as the candidate marker of experience. Many researchers now treat the P3b as a marker of post-perceptual processing (report, decision, task relevance) rather than of awareness itself, while the workspace camp answers with report-independent late ignition. Supporting evidence for ignition came from monkeys: Timo van Vugt, Pieter Roelfsema, Dehaene and colleagues (2018, Science) recorded V1, V4 and dorsolateral prefrontal cortex during threshold detection and found that weak signals were often lost on the way to frontal cortex, where seen stimuli evoked an all-or-none, ignition-like response. The Cogitate results (Chapter 0 box) added the twist that the predicted ignition at stimulus offset did not appear.
Precursors Are Not the Avalanche
Conscious perception, in summary, is a wave that tips the cortex past an ignition threshold: from about 300 ms, frontal regions receive sensory input bottom-up and send massive top-down projections back, forming the brain web whose facets are the four signatures.
The avalanche metaphor also settles a dispute over when awareness begins. Some labs find seen/unseen differences as early as 100 ms. Dehaene answers that small pre-ignition differences exist but are not consciousness, for three reasons:
- Inconsistency. Many excellent experiments with identical stimulation find only the late ignition.
- Wrong shape. Under masking, early events grow linearly with stimulus duration, whereas perception is nonlinear.
- Wrong size. Early effects are a small increment on a large subliminal response, so a big activation remains on trials the person denies seeing.
Early differences arise because random fluctuations in ascending activity raise the odds of later ignition, like the snowball or the butterfly of chaos theory. Averaging thousands of trials makes these small biases significant. Some studies even find predictors before the stimulus appears; that is not precognition but preconditions, ongoing fluctuations that make the cortex more or less ready. "Concluding that, at that stage, the brain is already conscious would be just as wrong as saying that the first snowball is already the avalanche."
Correlates versus signatures. "Neural correlate of consciousness" is the wrong target, because too many brain events correlate with perception, including fluctuations that precede the stimulus and cannot code for it. Dehaene wants a signature: present whenever conscious perception occurs, absent whenever it does not, and encoding the full content the person reports.
Decoding a Conscious Thought
The devil's advocate tries again: perhaps ignition is a nonspecific alert, a siren that sounds whenever anything becomes conscious. Brainstem and thalamic nuclei do label important moments. The locus coeruleus releases norepinephrine across the cortex during stressful, attention-demanding events, and some researchers proposed that the P3 reflects exactly this. If so, ignition would be like "the thump of the Sunday newspaper on our doorstep" mistaken for the news.
- Complete. It holds a full record of experience at exactly the level of detail perceived.
- Subjective. It ignores features the person misses even if physically present, and includes illusions and hallucinations.
- Similarity-preserving. If a diamond and a square look like two shapes rather than one rotated, so must the code.
- Invariant and stable. It stays put when the world seems stable and changes when it seems to move.
The invariance criterion rules out early sensory areas. Walking down a corridor sweeps images across the retina, and the eyes jiggle three or four times a second, yet the room looks stable; vestibular signals and motor predictions subtract self-motion (poke your eye with a finger and the world jumps). TV flicker at 50 to 60 Hz drives neurons in primary visual cortex but is never seen, and very fine gratings are encoded but invisible. Conversely, consciousness adds: the central visual field is massively over-represented early on, yet we see no magnifying lens; peripheral retina has few colour-sensitive cones, yet the periphery looks coloured; and the blind spot leaves no hole. Assembling a stable world from this jigsaw takes time, which may be why the signatures arrive a third of a second late.
Quiroga's neurons
Rodrigo Quian Quiroga, with Itzhak Fried and Rafi Malach, provided the decisive answer. Building on earlier work, neurosurgeon Fried developed fine electrodes that record individual neurons ("spikes") for hours or days while the patient lives normally. Excitatory neurons typically fire a few spikes per second.
- Extreme selectivity. Shown hundreds of pictures, an anterior temporal neuron typically responds to one or two: one fired to Bill Clinton and no one else; others to family members, the Sydney Opera House, the White House, Jennifer Aniston, Homer Simpson. The written name often suffices (Sydney Opera and its picture drive the same cell).
- A population code. Finding a Clinton cell by blind insertion implies millions of such cells active at any moment. Each picture yields a pattern of active and silent cells from which a computer can decode what the person sees with high accuracy.
- Tracking awareness. With masked pictures, most cells fired only when the patient reported recognising the image, though the display was identical.
- Independent of method. In binocular rivalry, a Clinton cell fired while Clinton's face was seen and stopped when a checkerboard in the other eye suppressed it, although the face remained on the retina.
- Late and sustained. Averaged over trials, the cells fired vigorously from about a third of a second, for a sustained period: the ignition pattern again, but content-specific, so it cannot be mere arousal.
- Imagery and memory. Most such neurons show the same selectivity for imagined pictures; a cell that fired to a clip of The Simpsons fired again whenever the patient, in darkness, recalled the clip.
The World Trade Center cell. At a fixed, threshold exposure the patient reported the picture about half the time, and the cell's spikes followed just those trials. The firing was so reproducible that a line drawn on spike count separated seen from unseen trials: a subjective state decoded from an objective brain state.
A single cell does not make a thought; content is spread across millions of cells in associative cortex. Their synchronous firing is strong enough that scalp potentials reveal whether someone sees a face or a building, slow parietal waves reveal the location and number of items held in short-term memory, and even fMRI of the anterior ventral temporal lobe can tell whether a face or a house was seen, with a stable pattern on conscious trials and none on unconscious ones.
A shrunken explorer looking for conscious spikes should seek three properties: stability over time, reproducibility across trials, and invariance over superficial change. The posterior cingulate cortex qualifies: its responses stay stable while the object stays put, even as the eyes move, and its cells code object location in the world. It connects to the parahippocampal gyrus and hippocampal "place cells", which fire for a location even in darkness and encode where the animal thinks it is. A rat "teleported" by switching room colours shows place cells flickering between two maps before settling on the illusory room, and position can be decoded from cell ensembles even during sleep, when trajectories are only imagined.
Causing and Destroying Consciousness
Inducing a hallucination
A signature must also be causal: recreating the neural state should recreate the experience. Dreams show it happens nightly. In rats, hippocampal and cortical patterns from the day are replayed in sleep; in humans, activity seconds before waking predicts dream content (activity in a face region predicts that the dream contained people). Two methods allow safe, reversible causal tests.
- Transcranial magnetic stimulation (TMS). Pioneered by S. P. Thompson (1910) and C. E. Magnusson and H. C. Stevens (1911), revived by modern technology: a current pulse through a coil creates a magnetic field that induces a discharge at a "sweet spot" of cortex, with only a click and occasional twitch. Over visual cortex in darkness it creates a phosphene; over MT/V5, a sense of motion; elsewhere, colour.
- Intracranial stimulation. In patients with epilepsy, Parkinson's or movement disorders, or during awake surgery (the brain has no pain receptors), small currents reveal function. Stimulating the face region of ventral visual cortex evokes a face; anterior temporal sites evoke memories: burnt toast, a full orchestra, giving birth, a horror film, a Proustian childhood scene. Wilder Penfield, who pioneered this, concluded that cortex holds a dormant record of life's events.
- Other sites. The insula evokes suffocation, burning, stinging, tingling, warmth, nausea or falling; the subthalamic nucleus can trigger instant depression (crying, monotone voice, glum thoughts); parts of the parietal lobe cause vertigo or an out-of-body experience.
Stimulation can produce "virtually any experience, from orgasm to déjà vu", but spread from the stimulation site blurs causality, and evidence suggests the initial induced activity is unconscious unless it reaches distant parietal and prefrontal cortex.
Desmurget's dissociation (Michel Desmurget). During surgery, low-threshold stimulation of premotor cortex moved the patient's arm, but she denied anything had happened (she could not see her limbs). Stimulating inferior parietal cortex produced a conscious urge to move and, at higher current, the conviction that she had moved her hand, though her body stayed still. Peripheral motor circuits can act without awareness; higher temporal, parietal and prefrontal regions generate reportable experience, even purely hallucinated.
Paul Taylor, Vincent Walsh and Martin Eimer set TMS over primary visual cortex at an intensity that produced a phosphene half the time, and recorded EEG. For 160 ms activity was identical on seen and unseen trials; then a P3 appeared, much larger when the phosphene was seen, with an earlier onset (about 200 ms) because the pulse skipped the first stages of vision. Causation runs from cortical activity to experience, but indirectly: local activity must be dispatched to distant sites, and ignition often returns to sensory areas, tying the fragments together. Only then do we experience "seeing".
Destroying a percept
If the late web causes experience, disrupting it should erase a visible stimulus.
- Dual-pulse TMS. A first pulse to MT/V5 produces a conscious feeling of motion; a second pulse to primary visual cortex shortly after erases it. The MT activity must loop back to V1 to be experienced: "consciousness lives in the loops."
- Replay. Stimulating visual cortex about 200 ms after a brief picture can make the person see it a second time, especially when told to remember it: a held image lingers as subthreshold activity in visual cortex, ready to be re-enacted.
- Parietal disruption. TMS to the left or right parietal region enhances masking and inattentional blindness; a faint but visible colour patch vanishes.
Lau's virtual frontal lesion. Hakwan Lau's team at Oxford delivered theta-burst TMS (pulses patterned to disrupt the theta rhythm, about 5 Hz) to each dorsolateral prefrontal cortex: 600 pulses in 20-second bouts, left then right, a "virtual lobotomy" lasting about twenty minutes. Objective discrimination of a near-threshold diamond versus square was unchanged. But participants lost confidence, could not rate their visibility, and felt their vision was unreliable. Before stimulation, visibility ratings tracked accuracy; afterwards the link was broken, which is the definition of blindsight, here induced in healthy brains.
Local or global loops?
Victor Lamme argues that any local recurrent loop (area A talks to B and B talks back) suffices for a form of consciousness: "we could even define consciousness as recurrent processing". Dehaene disagrees. Cortex is full of loops at every scale, and it would be odd if each tiny one held a fragment of awareness. His view: reverberation is necessary but not sufficient; only long-distance loops through prefrontal and parietal cortex create a conscious code.
Lamme: local recurrence. Any reciprocal loop between areas yields awareness, even without access or report. Supporting point: local loops disappear under anaesthesia.
Dehaene: global recurrence. Local loops serve unconscious assembly of a scene (small receptive fields must interact to detect global properties such as a large shadow). Loss under anaesthesia may be a consequence, not the cause. TMS to V1 at about 60 ms impairs conscious vision and blindsight together, so early local loops serve unconscious processing too; disrupting parietal and prefrontal regions, by contrast, impairs awareness while leaving performance.
Points to read with care. (1) The book dates the Sergent attentional-blink study to 1995; it was published by Sergent, Baillet and Dehaene in 2005. (2) The Weizmann Institute is in Rehovot, not Tel Aviv. (3) The theta-burst metacognition result has a replication problem: Daniel Bor, Anil Seth and colleagues (2017) failed to reproduce the loss of metacognitive accuracy after bilateral prefrontal theta-burst TMS, so this causal pillar is weaker than the chapter suggests. (4) Granger causality measures predictability, not physical causation, and on scalp or intracranial signals it is sensitive to volume conduction, common inputs and noise, so the "causal explosion" is best read as a large increase in directed statistical dependence. (5) The locus coeruleus account of the P3 is not refuted by content-specific neurons: both a nonspecific arousal burst and a content-specific code can coexist during ignition, and the chapter shows the second without excluding the first.
A Thing Which Thinks
Dehaene's conclusion is reductionist: every conscious experience, "from the sound of an orchestra to the smell of burnt toast", arises from massive cerebral circuits with reproducible signatures. Coordinated firing starts in local specialised regions, then invades the prefrontal and parietal lobes while staying synchronised with sensory regions; when that coherent web ignites, awareness is established.
He concedes that any signature might still be an epiphenomenon, like a locomotive's steam whistle that accompanies but does nothing. Yet disrupting high-level circuits removes subjective perception while sparing unconscious processing, and stimulation creates hallucinations, so electrical activity can both cause and destroy a state of mind. In principle neuroscientists accept the "brain in a vat" of The Matrix: "neural avalanches should cause mental symphonies". Technology lags, but silicon chips with thousands of electrodes and optogenetics (light-sensitive opsins from algae and bacteria, delivered by virus to chosen neurons, firing with millisecond precision at a laser's touch) are closing the gap; optogenetic stimulation of the hypothalamus has already woken a sleeping mouse.
Writing percepts into the brain. Dehaene's prediction that specific percepts would be induced de novo has partly come true in animals. James Marshel, Karl Deisseroth and colleagues (2019, Science) used holographic optogenetics to activate small ensembles of a few dozen neurons in mouse visual cortex and drove behaviour as if the mice had seen a specific oriented grating, and related work from Rafael Yuste's lab showed that holographically activating a learned ensemble could trigger the visually guided behaviour it was linked to. In humans, intracortical visual prostheses (for example Eduardo Fernández's team, 2021, implanting a microelectrode array in the visual cortex of a blind woman) have evoked phosphene patterns that could be recognised as simple shapes and letters. Human concept cells have been studied extensively since, and recordings with high-density probes now track hundreds of human neurons at once. Whether such writing produces experience in mice is, of course, inferred from behaviour.
The perfectly replayed sunset. Suppose a device could re-create, in your visual cortex alone, the exact spike pattern of the last sunset you saw, while silencing every long-range connection to parietal and prefrontal cortex. Lamme predicts you would experience the sunset without being able to report it; Dehaene predicts no experience at all, only an unconscious trace like a blinked word. Now reconnect the long-range fibres: Dehaene predicts the sunset would ignite into awareness after a few hundred milliseconds. The two theories make opposite predictions about the same physical state, which is why the local-versus-global question is empirical, and why stimulation studies matter more than any number of correlations.
Correlates, probes and interventions in AI systems. Dehaene's ladder of evidence (correlation, then content-specific decoding, then causal stimulation and disruption) is the same ladder now used to study what large models represent. A linear probe that decodes a concept from activations is a correlate; it may read a by-product, like the pre-stimulus fluctuations Dehaene refuses to count. Activation patching and ablation are the analogue of TMS: change the internal state and see whether the output changes. For anyone validating model explanations in a regulated setting, the transferable rule is Dehaene's: a feature that merely correlates with a decision is not evidence of how the decision was made until intervening on it changes the result.
Theorizing Consciousness
Dehaene's theory is that consciousness is brain-wide information sharing: a global neuronal workspace, built from long-axon pyramidal neurons concentrated in prefrontal, parietal, temporal and cingulate hubs, selects one representation, holds it in a self-sustaining ignited state, and broadcasts it to every specialised processor. Computer simulations with Jean-Pierre Changeux reproduce all four signatures, show spontaneous ignitions that block outside input, and yield a taxonomy of the unconscious (preconscious, subliminal, disconnected, diluted, latent) that explains why most of what the brain knows never reaches awareness.
The Need for Bridging Laws
The signatures of Chapter 4 say when consciousness is present, not why late firing, ignition and synchrony should produce a subjective state, or why firing in V4 yields colour and in V5 motion. Without a theory, searching for neural correlates is as empty as Descartes's pineal gland, because it keeps the very split (neural versus mental) that a theory should dissolve. What is needed is a set of bridging laws.
Physics faced a similar gap. How does a solid table arise from mostly empty space sparsely populated by atoms? Dissection alone was not enough; a synthetic mathematical theory was. The kinetic theory of gases of James Clerk Maxwell and Ludwig Boltzmann derived pressure and temperature from atomic motion, the first link in a reductionist chain that now covers glues, soap bubbles and stellar plasma. Neuroscience needs the equivalent, though its "condensed matter" is far harder: nested Russian dolls of processors, distributed circuits, dozens of cell types, and single neurons with tens of thousands of synapses. No experiment will record all hundred billion neurons, so only mathematical theory can show how the mental reduces to the neural. Over fifteen years, Dehaene, Jean-Pierre Changeux and Lionel Naccache have sketched one, condensing sixty years of psychological modelling.
Consciousness Is Global Information Sharing
The proposal in one sentence: To be aware of a piece of information means that it has entered a specific storage space that makes it available to the rest of the brain. Among millions of unconscious representations, one is selected for relevance to current goals and made globally available to all high-level decision systems. We have an evolved mental router. Bernard Baars called it a "global workspace": a system detached from the outside world in which private images can be entertained and spread across the specialised processors.
- Persistence. What enters the workspace can be held long after the stimulus is gone, independent of when and where it was perceived.
- Flexible use. It can be sent to language (hence reportability as the hallmark of consciousness), long-term memory, or planning.
- Flexible dissemination is, for Dehaene, the defining property of the conscious state.
Three metaphors the workspace synthesises
- The theatre (Hippolyte Taine, 1870). The mind is a stage "very narrow at the footlights", with room for one actor, while countless obscure figures wait in the wings. Decades before Freud, this implied one conscious item backed by a vast unconscious staff.
- The limited-capacity channel (Donald Broadbent). Pilots could not follow two streams of speech, one per ear; conscious perception must be a slow bottleneck handling one item at a time. The attentional blink and psychological refractory period later confirmed it; others speak of a "central bottleneck" or "second processing stage", a VIP lounge for the happy few.
- The supervisory system (1960s and 1970s). Consciousness as a central executive. William James (1890) called it "an organ added for the sake of steering a nervous system grown too complex to regulate itself". Dehaene removes the dualist flavour: the brain regulates itself hierarchically, with evolutionarily recent prefrontal centres leading (often inhibiting) posterior and subcortical systems.
Daniel Dennett warns that the theatre invites the homunculus fallacy: if consciousness is a stage, who watches, and who watches the watcher? The fix is to remove the audience's intelligence: "One discharges fancy homunculi from one's scheme by organizing armies of idiots to do the work." In Baars's version the audience is other unconscious processors, each acting on the broadcast according to its competence. The idea goes back to early AI "blackboard" architectures in which subprograms share a common data structure. "The conscious workspace is the clipboard of the mind."
Michael Posner and Tim Shallice proposed that information is conscious when represented in the high-level executive. Dehaene corrects this: subliminal stimuli can partially trigger executive control (Chapter 2). The relation runs the other way: executive attention is one of many receivers of workspace content. Whatever is conscious becomes available to language, long-term memory, attention and intention, which creates the feeling that our actions are "under control".
Beyond Modularity
Dehaene follows Baars: consciousness reduces to what the workspace does. Nothing in principle prevents silicon from doing it, but it is not trivial. Software is rigidly modular: a word processor holds text but cannot judge its global relevance or share it, so computers remain "despairingly narrow-minded"; only the clipboard lets programs share, and only under a human deus ex machina.
The cortex combines modular processors (patches of face-selective neurons; parietal and motor sectors for specific acts and body parts; abstract sectors for numbers, animals, objects, verbs) with a flexible routing system. The workspace evolved to mitigate modularity, and global availability is exactly what we experience as being conscious.
The thirsty elephant. Alone in the savannah, an elephant must walk to a distant, invisible waterhole. The decision needs a mental map, visual recognition of landmarks, trees and paths, and memory of past successes and failures. A long, exhausting journey is a decision that must pool every relevant source of knowledge; consciousness may have evolved "aeons ago" to tap all of them at once.
An Evolved Communication Network
If consciousness is sharing, it requires connectivity. Santiago Ramón y Cajal noticed in the late nineteenth century that neurons are enormously elongated (a motor-cortex neuron can project down the spinal cord) and that long-distance projection cells are dense in the cortex. Pyramidal cells send axons to the back of the brain or to the other hemisphere in bundles several millimetres thick and centimetres long, now visible with MRI tractography.
- Sensory areas are choosy. V1 connects mainly to neighbours in a coarse hierarchy (V1 to V2 to V3 and V4); early vision is functionally encapsulated, each neuron seeing a fragment "without any 'awareness' of the overall picture".
- Association areas break modularity. Long-axon neurons are most abundant in prefrontal cortex, which links to the inferior parietal lobe, middle and anterior temporal lobe, and anterior and posterior cingulate. These are the brain's major hubs.
- Reciprocity. If area A projects to B, B almost always projects back to A.
- Triangles. If A projects to B and C, then B and C are likely to be interconnected.
- Subcortical partners. The hubs connect to the central lateral and intralaminar thalamic nuclei (attention, vigilance, synchronisation), the basal ganglia (decision and action) and the hippocampus (episodic memory). Almost every pair of interconnected cortical areas also communicates through a parallel thalamic relay, and thalamic input keeps cortex in an "up" state; thalamic failure figures in coma and vegetative states (Chapter 6).
There is no single meeting point: the workspace is a decentralised network, "an elitist board of executives" in distant territories that stays in sync by messaging. This network coincides with the regions that ignite in Chapter 4, which explains signature 1: these areas have the long-range wiring needed to broadcast.
Neurons built for broadcasting
The relevant pyramidal neurons have giant cell bodies to sustain axons whose receptor molecules must travel centimetres from the nucleus. They concentrate in cortical layers II and III, which also carry the callosal connections between hemispheres.
- Thick upper layers (Constantin von Economo, 1920s). Layers II and III are much thicker in prefrontal and cingulate cortex and in parietal and temporal association areas: exactly the regions that activate in conscious processing.
- Huge dendrites and many spines (Guy Elston in Queensland; Javier DeFelipe in Spain). Dendrites (from the Greek for "tree") collect inputs at spines, mushroom-shaped protuberances at each synapse. Dendrites are much larger and spines far more numerous in prefrontal cortex than in posterior regions.
- Human exaggeration. Human prefrontal neurons are more branched and spinier than other primates'. Each may carry fifteen thousand spines or more, so it may hear from about as many neurons, most of them far away in cortex and thalamus.
- Genetic control. A family of genes mutated in humans shapes these trees, including FOXP2, with two mutations specific to the Homo lineage, which modulates language networks and whose disruption severely impairs articulation and speech. Mice carrying the two human mutations grew humanlike, larger dendrites and learned more easily ("although they still didn't speak").
Two genetic and anatomical points to correct or qualify. First, the two human FOXP2 amino-acid substitutions are shared with Neanderthals (Krause and colleagues, 2007), so they are specific to the Homo lineage but not to modern humans, and Elise Atkinson and colleagues (2018) found no evidence of a recent selective sweep at FOXP2 in modern humans, undermining the once-popular story of a language gene swept to fixation. Second, the humanised-Foxp2 mice (Wolfgang Enard and colleagues, 2009) showed longer dendrites and altered plasticity chiefly in striatal medium spiny neurons of the basal ganglia, not in cortical pyramidal cells as the chapter implies; later work linked them to faster switching from declarative to procedural learning. FOXP2 is a transcription factor that regulates many genes; it does not head a "family" of dendrite genes in a simple sense.
Von Economo the anatomist versus "von Economo neurons". The chapter credits Constantin von Economo with mapping layer thickness. He is also known for describing, in the 1920s, the von Economo neurons (VENs): large, spindle-shaped projection neurons in layer V of the anterior cingulate and fronto-insular cortex, found in humans, great apes, elephants and cetaceans, and selectively depleted in behavioural-variant frontotemporal dementia (William Seeley and colleagues, 2006), a disease that erodes social self-awareness. John Allman and Bud Craig proposed that VENs support fast long-range signalling for social and interoceptive awareness. They are an attractive candidate for workspace-style broadcasting from the salience network, but they are not part of Dehaene's argument, which rests on the far more numerous layer II and III long-axon pyramidal cells; their specific role in consciousness remains unproven.
Tracing a conscious face
Follow the wiring as a face is recognised, like the FBI tracing a call through telecom hubs. Short connections inside the retina clean up the image; the optic nerve carries it to the visual thalamus, then primary visual cortex; local U-shaped fibres pass it to face patches in the right fusiform gyrus. All this is unconscious. Then, as Swiss anatomist Stéphanie Clarke found, long-distance axons from the right inferior temporal lobe project in a single synaptic step to distant association cortex, including the other hemisphere, concentrating in Broca's area (inferior frontal) and Wernicke's area (temporal association cortex). At this stage words attach to the image, and because these nodes belong to the wider workspace, the content can reverberate through the whole inner circle. Access to this network is, on the theory, all that is needed for awareness.
Sculpting a Conscious Thought
The repertoire of possible conscious thoughts is effectively endless (every face, shade of emotion from anger to schadenfreude, fact, rumour and word in any language). Giulio Tononi points out that this constrains the code: it needs enormous differentiation, billions of distinct activity patterns, each sharply bounded ("either it's a bird, or it's a plane, or it's Superman"). A clear mind with myriad possible thoughts needs a brain with myriad possible states. Dehaene assembles his code from five earlier ideas.
- Cell assemblies (Donald Hebb, 1949). Neurons linked by excitatory synapses stay active after the stimulus ends, "capable of acting briefly as a closed system"; many local assemblies combine into a combinatorial code (a fire extinguisher coded by active patches of a few hundred neurons each for handle, body, hose).
- Pandemonium (Selfridge, 1959). A hierarchy of "daemons" shout interpretations in proportion to the evidence; eyes, nose and hair daemons together excite a face daemon, and a decision system listens to the loudest. Single-cell findings (line, colour, face, even president and film-star cells) support it.
- Recurrence and attractors. Unlike the feed-forward original, real neurons talk back and sideways (if α projects to β, β probably projects to α). Recurrent networks settle into attractor states, stable and reproducible patterns that reach consensus, restore missing parts, remove noise and suppress alternatives: the winning "neural coalition" of Crick and Koch.
- Convergence zones (Antonio Damasio). Hubs in prefrontal cortex, anterior temporal lobe, inferior parietal lobe and precuneus gather inputs across space and time, so the Mona Lisa's smile, hands, eyes, name and memories bind into one percept rather than a "disembowelled Picasso".
- Reentry (Gerald Edelman). Constant back-and-forth between levels lets groups of neurons act as expert statisticians who jointly explain the input: a shadow expert accounts for a dark zone if light comes from the top left, a lighting expert agrees, and a third concludes the rest is a face.
Dehaene's synthesis: a conscious state is the stable activation, for a few tenths of a second, of a subset of workspace neurons distributed across many areas, each coding a facet of the same representation. Through long axons they exchange messages until they converge on a coherent, synchronous interpretation, kept in step top-down by long-distance neurons, like fireflies harmonising their flashes. Without prefrontal access, moderate assemblies can still synchronise locally (a word's meaning in the left temporal lobe), but the content is not shared and stays unconscious.
Of the cortex's roughly sixteen billion neurons, each conveys only a few bits; at any moment one object of thought is selected and its neurons fire in partial synchrony "under the aegis of a subset of prefrontal cortex neurons". Silent neurons carry information too: their muteness says their feature is absent or irrelevant. Perception is like sculpting a statue: from hundreds of millions of uncommitted workspace neurons, the brain silences most and the remaining active set "delineates, quite literally, the contours of a conscious thought".
Why the P3 is positive
This picture explains signature 2. Active workspace neurons broadcast, but mostly onto inhibitory interneurons that hush irrelevant groups. In the active cells, synaptic currents flow from superficial dendrites toward the cell body, and because these cells are parallel their currents add to a focal negative scalp voltage over the regions coding the content. The inhibited cells, far more numerous, sum to a broad positive voltage: the P3. The P3 is so large, generic and reproducible because it mostly signals what the current thought is not about; the focal negativities carry the content.
Edward Vogel's working-memory negativities (University of Oregon). When people memorise an array of objects, slow negative voltages over parietal cortex indicate how many items were seen and where. They last as long as the items are held, grow as items are added, saturate at capacity, collapse when items are forgotten, and track the number remembered. Negative voltages directly delineate a conscious representation, as the theory predicts.
Simulating a Conscious Ignition
With Changeux, Dehaene built simulations to test whether workspace wiring produces the signatures. Gaston Bachelard (1938): science seeks "the mathematical how", not just the phenomenological how.
- Neurons. Integrate-and-fire units with realistic synapses modelling major neurotransmitter receptor types.
- Columns. Neurons wired into cortical columns (cells stacked perpendicular to the surface, tightly connected, sharing responses and descending from the same founder cell), organised in layers.
- Thalamus. A small multinucleus thalamus with realistic connection strengths and axonal delays, forming the thalamocortical column. The model fired spontaneously and produced an EEG-like signal.
- Hierarchy. Four areas (A to D), each with two columns coding two objects, a sound and a light (a drastic simplification for tractability).
- Competition. At the periphery both objects could be active in parallel; at higher levels they inhibited each other, allowing one "thought".
- Feedback. Feed-forward projections from area to area plus long-distance feedback from higher areas to the sensory areas that excited them.
- Vigilance. Every pyramidal neuron received a small vigilance current standing for acetylcholine, noradrenaline and serotonin from brainstem, basal forebrain and hypothalamus.
Injecting current into the visual thalamus (the lateral geniculate body) produced:
- Global ignition. Activity climbed A, B, C, D in order, then the columns for the perceived object ignited together through mutual excitation while the rival object was actively inhibited. The ignited state lasted hundreds of milliseconds, largely independent of stimulus duration: a brief pulse could start a lasting reverberation, which is how a flashed picture is held online.
- All four signatures. Most simulated neurons showed a late, sudden increase in synaptic input (signature 1, and with it the P3); excitation returned to sensory areas (late amplification); reverberation through nested loops (within columns, cortex to thalamus, across areas) raised broadband oscillations peaking in gamma (signature 3); and spikes became strongly synchronised among neurons coding the conscious content (signature 4).
- A phase transition. Because higher units excite the units that excited them, the system has two stable states separated by an unstable ridge. Below a critical input, activity dies; above it, it snowballs. Intermediate stimuli have an unpredictable fate, matching the 150-year-old psychological idea of a threshold between subliminal and supraliminal.
- A second, nested transition. The ignition threshold is itself controlled by vigilance. At low vigilance, spontaneous activity collapses and ignition disappears: even strong inputs activate primary and secondary areas and fizzle, like a sleeping or anaesthetised brain. Raising vigilance brings back a structured EEG and ignition, and the threshold falls as vigilance rises, so alertness makes faint inputs detectable. The model thus captures states of consciousness as well as access.
The Restless Brain
The simulation also showed spontaneous activity. With no input, neurons fired from random synaptic events, and at high vigilance this chaos self-organised into occasional global ignitions: a set of columns coding one object would activate briefly and fade, to be replaced a fraction of a second later by another. The only difference from perception was direction: spontaneous ignitions tended to start in workspace areas and propagate down to sensory regions.
The real brain does this constantly. The EEG is dominated by massive spontaneous waves; stimulus-evoked activity is so small that it needs heavy averaging and probably accounts for less than 5 percent of the brain's energy use. The nervous system is primarily an autonomous device that generates its own thought patterns.
- Animals. Amiram Grinvald's team at the Weizmann Institute used voltage-sensitive dyes on anaesthetised animals and saw, in darkness, whole assemblies activate spontaneously.
- Humans at rest. Resting brains show constantly changing, bilateral networks that recur across people; part of the language network that responds to stories also fires spontaneously in darkness, consistent with inner speech.
- Two components. Some resting correlation simply follows anatomy and persists in sleep, anaesthesia and unconscious patients. Another part betrays ongoing thought: the default-mode network turns on during reflection on oneself, autobiographical memory and comparing one's thoughts with others'. People interrupted during high default-mode activity report more mind-wandering.
- Competition with the world. During high default-mode activity, an unexpected picture no longer evokes a large P3. Absorbing internal thought occupies the workspace and blocks access, a version of inattentional blindness.
The simulation matched this: during spontaneous ignition, even a supra-threshold external input was blocked, mimicking inattentional blindness and the attentional blink. Ongoing fluctuations also explain why the same stimulus sometimes ignites and sometimes fizzles: it depends on whether pre-stimulus noise is aligned or at odds with the incoming spikes (the "precursors" of Chapter 4).
Darwin in the Brain
Dehaene calls spontaneous activity the most overlooked and most original feature of the model. Too many neuroscientists keep the reflex arc of Descartes, Sherrington and Pavlov, the brain as an input-output device. Autonomy is the primary property of the nervous system. The brain generates stochastic patterns; in development the useful ones are kept and the rest weeded out, a Darwinian selection of thoughts most visible in young children.
William James asked why the hemispheres are not just a spinal cord with more reflexes, and answered that the brain is "an organ whose natural state is one of unstable equilibrium", letting its owner adapt to "the minutest alterations" in circumstances. Neurons acquired the ability to self-activate early in evolution; filtered by circuits, this becomes exploration, as in the central pattern generators that produce walking or swimming. Dehaene contends that primates (and probably many species) explore cognitively in the same way, generating, testing and discarding plans.
The workspace as a generator of diversity. Spontaneous activity is a "Generator of Diversity" (jokingly, GOD) whose patterns are selected by anticipated reward. Dehaene and Changeux showed that networks built this way solve puzzles such as the Tower of London and, combined with ordinary synaptic learning, learn from mistakes and extract abstract rules. Nothing is magical: the noise comes from random vesicle release, ultimately thermal noise. Digital chips suppress noise with well-separated voltages for 0 and 1; neurons amplify it, probably because randomness helps search, as in Markov chain Monte Carlo and simulated annealing. The stream of consciousness originates in random spikes sculpted by trillions of synapses.
A Catalogue of the Unconscious
Like Linnaeus classifying species, workspace theory sorts the unconscious into mechanistically distinct bins. Dehaene illustrates the first with the grandfather clock in his wife's seventeenth-century family farmhouse: its ticking is loud and clear, yet vanishes whenever he concentrates on writing.
| Category | Mechanism | Can attention make it conscious? | Example |
|---|---|---|---|
| Preconscious | Strong active assembly, but denied entry because the workspace is occupied; fringe inhibition blocks it | Yes, while it lasts in a decaying buffer (then experienced in retrospect) | The ticking clock; inattentional blindness; the blinked second target |
| Subliminal | Activation too brief or weak: top-down amplification returns after the mask has replaced it | No, however hard one tries | Masked words between shapes |
| Disconnected | Strong, lasting firing with no connection to workspace hubs | No; only by recoding through another sense | Brainstem breathing control (CO2 levels); callosal disconnection, where a patient disowns his left hand |
| Diluted | Information spread through a tangled spatiotemporal pattern, never made explicit by a compact assembly | No, until training creates explicit neurons | Gratings too fine or flickering above 50 Hz encoded in V1 but seen as grey |
| Latent | Stored in synaptic weights, not firing | Only when reactivated as firing that reaches prefrontal and cingulate cortex | Memories; contour statistics; grammar rules |
Preconscious is a word Dehaene borrows from Freud's Outline of Psychoanalysis: processes "capable of entering consciousness". The central inhibition that defines each conscious content is what creates the bottleneck: it prevents seeing two things or doing two effortful tasks at once, while early sensory areas still light up almost normally. Subliminal stimuli are different in kind: the sensory wave "dies out before creating a tsunami on the shores of the global neuronal workspace".
Two consequences of the diluted category stand out. Sensory training makes faint sights or sounds conscious because neurons at all levels retune to amplify them. And the brain holds signals its owner ignores: a US military programme flashed satellite photos at ten per second to trained observers and monitored brain potentials for unconscious hunches that an enemy plane was present. Dehaene foresees computer-assisted decoding as "a rigorous form of extrasensory perception".
The latent category is the largest. Synapses (some hundred thousand billion) adapt even before birth; millions form or vanish daily, especially in early childhood. Each stores a bit of statistics: how likely is my presynaptic neuron to fire just before my postsynaptic one? Early visual connections compile contour statistics; years of piano practice change grey-matter density; the hippocampus stores where, when and with whom. A conscious memory is just an old conscious moment reconstructed: retrieval must turn synaptic traces back into firing that invades prefrontal and cingulate cortex.
Grammar: unconscious. "John believes that he is clever" (he can be John); "He believes that John is clever" (he cannot); "The speed with which he solved the problem pleased John" (he can). We know the answers, not the rules. Syntax runs in connections linking left superior temporal cortex and inferior frontal gyrus, sparing dorsolateral prefrontal cortex; under anaesthesia temporal language cortex keeps processing speech.
Arithmetic: conscious. In 24 × 31 every step, its order and even errors are open to introspection. Multi-step calculation runs under prefrontal, cingulate and parietal workspace nodes, where single cells code intentions, plans, steps, their number, errors and corrections. Same difficulty, different format: Dehaene predicts the neural code for grammar will prove radically different from that of mental arithmetic.
Subjective States of Matter
The theory explains why we access "a scrawny portion" of stored knowledge: to be conscious, information must be an organised activity pattern in higher cortex that ignites the inner circle of workspace areas, and the features of that ignition are the signatures. The simulations are not conscious, but Dehaene does not doubt that a program could capture a conscious state. A better one would have billions of differentiated states and do real statistical inference (the probability that a face is present, or that a reach will succeed). Elementary decisions arise from accumulating noisy evidence; during ignition a subset collapses into one interpretation. Picture Selfridge's daemons in a large arena, accumulating evidence at a global level until a coherent answer fits the organism's goals. The machine is driven only partly by input: "Autonomy is its motto." It sets its own goals through spontaneous activity, which then directs memory retrieval, imagery and rule-based transformation, each coherent result moving one step through a mental algorithm that never stops. He looks to the European Human Brain Project and neuromorphic chips to simulate networks of millions of neurons and billions of synapses.
Smaller slips and a debated foundation. The pandemonium model's author was Oliver Selfridge, not "John" Selfridge. The attentional blink and refractory period were introduced in Chapter 1, not Chapter 2. Broadbent's filter theory drew on wartime listening problems but was developed in the 1950s (Perception and Communication, 1958). More substantively, the theory's identification of consciousness with access is exactly what critics dispute: Ned Block's distinction between access consciousness and phenomenal consciousness (1995) holds that experience may overflow what the workspace can report, and Lamme and the IIT camp place experience in posterior cortex rather than in frontoparietal broadcasting. Dehaene's account is a theory of conscious access that denies the gap exists; readers should see that as a position, not a settled result.
| Theory | Core claim | Where consciousness lives | Status since 2014 |
|---|---|---|---|
| Global neuronal workspace (Dehaene, Changeux, Naccache) | Conscious = ignited, globally broadcast information | Long-range frontoparietal network with sensory areas | Restated by Mashour, Roelfsema, Changeux and Dehaene (2020); mixed results in Cogitate (no offset ignition; limited prefrontal content) |
| Integrated information theory (Tononi, Koch) | Consciousness is integrated information (Φ) | Posterior "hot zone" | Predicted sustained posterior synchrony not found in Cogitate; 2023 "pseudoscience" letter and backlash |
| Recurrent processing (Lamme) | Local recurrence suffices for phenomenal experience | Sensory cortex loops | Supported by some no-report data; hard to test separately from report |
| Higher-order theories (Lau, Rosenthal, Brown) | A state is conscious when represented by a higher-order (often prefrontal) state | Prefrontal metarepresentation | Lau's perceptual reality monitoring (2019, 2022) developed; overlaps GNW on prefrontal role |
| Predictive processing (Clark, Seth, Friston) | Experience is the brain's best prediction of causes, including the body | Hierarchy-wide, with interoception | Seth's Being You (2021) popularised the "controlled hallucination" view |
The workspace since the book. George Mashour, Pieter Roelfsema, Jean-Pierre Changeux and Dehaene (2020, Neuron) restated the theory, emphasising ignition as a nonlinear, all-or-none event and reviewing supporting evidence from monkey recordings and anaesthesia. Large-scale, connectome-based models of macaque cortex from Xiao-Jing Wang's group (Madhura Joglekar and colleagues, 2018) reproduced threshold-like, all-or-none propagation of activity into frontoparietal areas, giving the simulation of this chapter a far more realistic successor. The Human Brain Project ran from 2013 to 2023; it did not deliver the promised human-scale cortical simulation but left the EBRAINS research infrastructure and neuromorphic platforms (SpiNNaker in Manchester, BrainScaleS in Heidelberg). In machine learning, workspace ideas have been implemented directly: Rufin VanRullen and Ryota Kanai (2021) proposed a deep-learning global workspace, and Anirudh Goyal, Yoshua Bengio and colleagues (2022) showed that modules competing to write into a small shared workspace and reading its broadcast could coordinate better than all-to-all communication. The adversarial collaboration results are summarised in the Chapter 0 box.
The clipboard with no user. Imagine an office software suite in which every program can write to a single, tiny clipboard, but only the program whose content is most relevant to the current goal wins the slot, after which every other program automatically reads it and the winning content stays put until displaced. No human chooses. Is this suite conscious? Dehaene's theory says it implements access and that, scaled up with rich representations, self-monitoring and spontaneous activity, nothing more is needed. Block would say you have built access without phenomenality, a perfect reporter that may feel nothing. The thought experiment isolates the theory's wager: that once the functional story is complete, no residual question remains.
Broadcast hubs in enterprise agent architectures. The workspace is a design pattern before it is a theory: many specialised processors, one small shared state, a competition for write access, and broadcast to all readers. Multi-agent systems that let every agent message every other agent suffer the cost the brain avoids; a single, capacity-limited shared state with explicit selection keeps coordination tractable and auditable. Dehaene's taxonomy is also a checklist for why information fails to reach the decision layer: it was preconscious (present but never selected), subliminal (too weak to cross the threshold), disconnected (no route to the hub), diluted (never made explicit), or latent (in model weights, not in the context). Each failure needs a different fix.
The Ultimate Test
The clinic is where a theory of consciousness must prove itself. Dehaene sorts the disorders of consciousness, recounts how Adrian Owen's tennis-imagery test exposed conscious minds inside "vegetative" patients, and then builds theory-driven detectors: the local-global auditory test that isolates a conscious P3, Marcello Massimini's TMS-EEG "ping" of the cortex, and EEG measures of long-distance information sharing that can spot consciousness before the clinician does. He ends with ways to switch the cortex back on: central thalamic stimulation, dopaminergic drugs and the paradox of the sleeping pill zolpidem.
Buried Alive
Dante's line "I was not dead nor living" frames the chapter. In The Count of Monte Cristo (1844), Alexandre Dumas gave probably the first literary description of locked-in syndrome: Monsieur Noirtier, "almost as immovable as a corpse", communicating only through his eyes. The condition is real.
Jean-Dominique Bauby. The editor of French Elle, aged forty-three, suffered a brainstem stroke on December 8, 1995, followed by a twenty-day coma. He woke fully paralysed except for one eye and part of his head. An assistant recited letters in frequency order (E, S, A, R, I, N, T, U, L, O, M...) and he blinked his left eyelid at the right one: about two hundred thousand blinks produced The Diving Bell and the Butterfly (1997), a best seller. The body was the diving bell; the mind, the butterfly. He died of pneumonia days after publication. For Dehaene, there is no better proof of the autonomy of consciousness.
Many such minds go undetected. A survey by the French Association of Locked-In Syndrome (founded by Bauby and run by patients) found that more than half the time a family member, not a physician, first detects consciousness, that diagnosis takes on average 2.5 months, and that some patients wait four years. Involuntary twitches make doctors dismiss genuine blinks as reflexes. Even in the best hospitals, about 40 percent of patients classed as vegetative turn out, on closer examination, to show signs of minimal consciousness. And since half of deaths in intensive care follow a decision to withdraw life support, detection is literally a matter of life or death.
How to Lose Your Mind
| State | Wakefulness | Awareness | Key clinical features |
|---|---|---|---|
| Brain death | None | None | No brainstem reflexes, flat EEG, no spontaneous breathing; cortical metabolism and perfusion abolished; irreversible, equated with death in most countries and by the Vatican |
| Coma | None (eyes closed, cannot be roused) | None | Lasts an hour or more; gag and pupillary reflexes preserved; slow EEG like sleep or anaesthesia, sometimes "alpha coma"; usually resolves in days or weeks |
| Vegetative state / unresponsive wakefulness | Sleep-wake cycle returns | No signs | Autonomic functions intact; random smiles, cries, slow movements; never follows commands; "persistent" after one month, "permanent" after 3 to 12 months (anoxic vs traumatic) |
| Minimally conscious state | Present | Fluctuating, inconsistent | Follows some commands, tracks a mirror, yes/no by nod or word, contextually appropriate emotion |
| Locked-in syndrome | Present | Intact | Lesion in the pons disconnects cortex from spinal outputs; vertical eye movements and blinks spared; not a disorder of consciousness |
- Coma's causes. Head trauma (often car accidents), stroke, anoxia (cardiac arrest, carbon monoxide, drowning) and poisoning (including binge drinking). The word comes from Greek for "deep sleep".
- The alpha wiper. Neurologist Andreas Kleinschmidt calls the alpha rhythm "the brain's windshield wiper": in normal brains it shuts down regions not in use (visual cortex while listening). In some comas, and under propofol (the sedative that killed Michael Jackson), a giant alpha rhythm invades cortex and wipes out conscious states. Because the cells are alive, normal rhythms can return.
- Terri Schiavo. Spent fifteen years in a vegetative state while her family, the state of Florida and President George W. Bush fought legal battles; her feeding tube was removed in March 2005. Her random facial expressions persuaded her parents she could be helped.
- The word "vegetative". Coined by Bryan Jennett and Fred Plum from the Oxford English Dictionary sense of "vegetate": to live "a merely physical life devoid of intellectual activity or social intercourse". Many clinicians prefer the neutral "unresponsive wakefulness", since "permanent" invites premature withdrawal of care, and the category, as the chapter shows, includes conscious but noncommunicating patients.
- Measuring minimal consciousness. Because responses fluctuate within hours and observers' judgements are subjective, Joseph Giacino created the Coma Recovery Scale: standardised, repeated probes of object recognition and use, gaze orienting to command, and reactions to unexpected noise. Minimally conscious patients have a better statistical prognosis than long-term vegetative ones.
What is it like to be minimally conscious? Nobody knows, but the huge fluctuations suggest a haze, like the confused state after being knocked out, anaesthetised or very drunk. Locked-in patients are the opposite: a well-delimited brainstem lesion spares cortex and thalamus, leaving a clear mind. Émile Zola's Madame Raquin (Thérèse Raquin, 1867) speaks only with "two living eyes", and savours revenge as her son's murderers commit suicide before her. Real patients answer e-mail through eye trackers, run nonprofits, or, like executive Philippe Vigand, write two books and father a child. In surveys, most locked-in patients past the first months rate their happiness at the population average. An intact cortex and thalamus suffice to generate a full mental life, even inside a body that cannot move.
Dates and figures to correct. The minimally conscious state criteria were published by Joseph Giacino and the Aspen Workgroup in 2002 (Neurology), not 2005. Bauby died on March 9, 1997, two days after his book appeared, not three. Bryan Jennett and Fred Plum coined "persistent vegetative state" in 1972.
Cortico Ergo Sum
In 2006 Adrian Owen and colleagues reported in Science a twenty-three-year-old woman, injured in a traffic accident with bilateral frontal damage, who five months later met every criterion of the vegetative state. In the scanner her language network (superior and middle temporal gyri) responded strongly to sentences, and Broca's area worked harder for ambiguous ones ("the creak came from a beam in the ceiling"). Language activation alone proves little, since it persists in sleep and anaesthesia, so Owen gave instructions: imagine playing tennis, imagine walking through your home, or relax, in thirty-second bouts each cued by a single spoken word (one for tennis, one for the house tour, and "relax").
- Tennis switched the supplementary motor area on and off every thirty seconds, as instructed.
- The house tour activated a spatial network: parahippocampal gyrus, posterior parietal lobe, premotor cortex.
- Both matched healthy controls performing the same tasks.
Sceptics argued that the words alone might activate motor or spatial meaning unconsciously. Neurologist Allan Ropper quipped that society is not ready for "I have brain activation, therefore I am", which "would seriously put Descartes before the horse". Dehaene calls this wrong. Owen scanned healthy people merely hearing the two cue words without instructions: the two words evoked indistinguishable activity, unlike the imagery networks. And a single word could not drive thirty seconds of activity; on workspace theory, an unconscious activation would dissipate within seconds, whereas sustained, task-specific prefrontal and parietal activity signals conscious content held in working memory.
Freeing the inner butterfly
Monti, Owen and colleagues (New England Journal of Medicine, 2010). Of 54 patients with disorders of consciousness, five modulated their brain activity on the imagery tasks, four of them vegetative. One was asked personal questions ("Do you have any brothers?") and told: imagine tennis for yes, your home for no, starting at "answer" and stopping at "relax". For five of six questions one network lit up clearly (the sixth gave no answer), and all five answers, unknown to the experimenters, were correct according to his family. Asked whether his father's name was Thomas, he activated the spatial (house-tour) network: no, which was right.
The chain of intact functions is long: language comprehension, long-term memory to find the answer, working memory to hold it for minutes, flexible routing of yes and no onto arbitrary imagery tasks (itself evidence of an intact global workspace), and executive task switching across five scans. A "demanding statistician" would want twenty questions, not five, but some patients are vegetative only in appearance.
- What it does not show. Journalists wrote that "coma patients are conscious"; the study included no comatose patients, and only 10 to 20 percent of vegetative and minimally conscious patients responded.
- Asymmetry. A positive result means consciousness is almost certain; a negative one means little, since deafness, aphasia, low vigilance or poor sustained attention can cause failure.
- Cause matters. All responders had traumatic injuries; none of those with massive stroke or anoxia responded, perhaps because of diffuse, irreversible cortical damage like Schiavo's. Using such cases as a blanket argument for unlimited life support would be "utterly irrational".
- The minimally conscious paradox. Thirty of thirty-one minimally conscious patients failed the test, though all had shown bedside signs of awareness. So clinical minimal consciousness is not a normal mind, and the imagery test underestimates consciousness by a wide margin.
Toward bedside interfaces
fMRI is costly and rarely repeated; Owen's own responder was tested once. Owen: "It's hard to open up a channel of communication with a patient and then not be able to follow up." Teams therefore turned to EEG.
- Imagined fist and toes (Damian Cruse, Owen and colleagues). On each beep, patients imagined squeezing their right hand or wiggling their toes; machine learning sorted the motor-cortex oscillations. It worked in three of sixteen vegetative patients, but even healthy volunteers succeeded in only nine of twelve sessions, and chance findings could not be excluded.
- Swimming versus home (Nicholas Schiff's team, New York). Five healthy volunteers and three patients; results seemed reliable but numbers were too small.
- Implants. A quadriplegic patient controlled a robotic arm with electrodes implanted in cortex, and Dehaene anticipates speech synthesis from language areas.
- Clinical routine. At Steven Laureys's Coma Science Group in Liège, brain-computer interfaces were already part of the standard battery for every vegetative admission. Dehaene predicts that within twenty years locked-in patients will drive wheelchairs by thought.
Conscious Novelty Detection: The Local-Global Test
Owen's test needs language, memory and imagination, so a conscious patient can fail it many ways, and it is not derived from any theory. Over lunch in Orsay in 2008, Dehaene, Tristan Bekinschtein, Lionel Naccache and Mariano Sigman asked what the simplest possible consciousness detector would be. Their choices: EEG (cheap, available), sound (hearing is usually preserved, vision often not), and the clearest signature, the P3. An unexpected phone ringing during a quiet concert produces a massive P3.
The obstacle is the mismatch negativity (MMN): a negative voltage from auditory cortex at about 100 ms after a deviant sound. It reflects an automatic, unconscious prediction-and-compare circuit (probably present in every cortical patch) and survives inattention, sleep and even coma, so a simple oddball test cannot separate conscious from unconscious brains.
- Local deviant. Sequences of five tones, beep beep beep beep boop. The final boop is locally novel and always triggers an MMN.
- Habituation. Repeat this sequence; consciously, the surprise disappears, and so does the P3, but the "rather stupid" auditory cortex still fires an MMN at every final boop.
- Global deviant. Rarely, play beep beep beep beep beep. Nothing is locally new, yet the sequence violates the pattern held in working memory, and it evokes a pure P3 with no MMN.
- Controls. Swapping beep and boop rules out acoustic features; asking subjects to count global deviants amplifies the P3.
- Why it works. Detecting a global deviant needs a five-tone pattern held over more than a second and compared across sequences (Chapter 3's link between consciousness and maintenance), and a second-order prediction that expects the local novelty: "the novelty is that there is no final novelty".
- In healthy people. The global P3 appeared in every participant after a short recording, only when attentive and aware of the rule: a demanding visual task abolished it, and mind-wanderers showed it only if they could report the regularity afterwards. EEG, fMRI and intracranial recordings showed global deviants igniting bilateral prefrontal, anterior cingulate, parietal and some occipital areas.
- First patients. Of eight, all four vegetative patients lacked the global response; three of four minimally conscious showed it, and those three later recovered.
- Routine use (Naccache, Salpêtrière, Paris). Of twenty-two vegetative patients, only two showed a global P3, and both reached minimal consciousness within days, suggesting they were already conscious.
- A life-changing case. A young man three weeks in coma after a car crash, with many complications, was under discussion for withdrawal of treatment; his strong global response led Naccache to argue for patience, and he recovered fully to a near-normal life.
The weakness is sensitivity. Even with a machine-learning search for patient-specific responses, the test misses about half of minimally conscious or recovering patients: high specificity, poor sensitivity. Hospital EEG is noisy, and many conscious patients cannot understand, count or attend for long enough.
Pinging the Cortex
If a conscious brain broadcasts over long distances, that capacity should be measurable without any task. In the late 2000s Marcello Massimini (University of Milan) proposed sending a pulse into the cortex, like sonar: with TMS to "ping" and EEG (with amplifiers that recover within milliseconds of the magnetic shock) to record the echo. The patient need not attend or understand.
- Healthy subjects. In deep sleep and anaesthesia the pulse produced a short, focal response, confined to the first 200 ms or so. Awake, or dreaming, the same pulse triggered a complex, long-lasting, widely spreading sequence, regardless of stimulation site, consistent with Dehaene's beyond-300-ms broadcast.
- Patients. In five vegetative, five minimally conscious and two locked-in patients, the test was 100 percent correct: all conscious patients showed complex responses. Of five more vegetative patients followed for months, the three who became minimally conscious were exactly those whose responses regained complexity, and spread into prefrontal and parietal regions tracked the level of consciousness.
From "ping" to the Perturbational Complexity Index. Massimini, Adenauer Casali, Giulio Tononi and colleagues (2013) compressed the TMS-evoked response into a single number, the Perturbational Complexity Index (PCI): how algorithmically incompressible the spatiotemporal pattern of the cortical echo is, capturing both integration (spread) and differentiation (diversity). Silvia Casarotto and colleagues (2016) calibrated a cut-off (PCI* = 0.31) on a benchmark of 150 people whose state was known (awake, dreaming, deep sleep, anaesthesia with different drugs, conscious brain-injured patients), where it classified every case correctly; applied to patients, it was high in about 95 percent of minimally conscious patients and, strikingly, high in 9 of 43 vegetative patients, flagging possible covert consciousness. Faster variants (PCI-ST, 2019) work on other perturbations. PCI was developed by IIT researchers and is often cited in IIT's favour, but it is equally a measure of long-range, differentiated broadcasting, so it does not by itself decide between the theories.
Detecting Spontaneous Thought
A simpler idea: a conscious brain should show long-distance exchange even at rest. PET measures glucose use with a radioactively labelled glucose analogue. Anaesthesia and deep sleep cut cortical glucose use by about 50 percent, as do coma and the vegetative state, and since the 1990s Steven Laureys's team has shown that the drop concentrates in bilateral prefrontal and parietal cortex, cingulate and precuneus: the workspace network.
Isolated regions can stay active without consciousness: vegetative patients with occasional facial movements show active focal motor areas, and one patient who had blurted random words for twenty years had metabolism confined to islands of left-hemisphere language cortex. But metabolism cannot diagnose individuals: some vegetative patients have near-normal cortical metabolism (lesions in ascending diencephalic structures), and patients who recover to minimal consciousness rarely regain normal metabolism. Structural MRI, too, is only indicative.
Information sharing in the resting EEG (Jacobo Sitt, Jean-Rémi King, Lionel Naccache, Dehaene). About 200 high-density recordings (256 electrodes) from vegetative, minimally conscious and conscious patients. Sitt, physicist, computer scientist and psychiatrist, devised a fast measure, weighted symbolic mutual information (wSMI), of how much information two sites share. It sharply separated vegetative patients from everyone else, especially for electrode pairs at least 7 or 8 cm apart; a directional measure showed bidirectional exchange between posterior specialised areas and parietal-prefrontal hubs. Other features agreed: unconsciousness lost high frequencies in favour of sleep-like low frequencies, and conscious cortex showed more synchrony. King trained a classifier to combine all measures; twenty minutes of EEG gave an excellent diagnosis, almost never calling a vegetative patient conscious.
The most important result concerns disagreements with the clinic. For two-thirds of clinically vegetative patients the algorithm agreed, and only 20 percent of them later reached minimal consciousness. For the other third, where it detected consciousness the clinician missed, 50 percent recovered a clinically obvious conscious state within months. Theory-driven brain measures can now detect traces of consciousness before they appear in behaviour, more sensitively than an experienced clinician.
| Test | What the patient must do | Signature used | Strength | Weakness |
|---|---|---|---|---|
| Owen's imagery fMRI | Understand, imagine tennis or home for 30 s | Sustained SMA or spatial network | Positive result nearly proves awareness; allows yes/no answers | Expensive, rarely repeated; misses most conscious patients (30 of 31 MCS failed) |
| EEG motor imagery | Imagine fist or toes on cue | Motor-cortex oscillations, classifier | Cheap, bedside | Unreliable even in healthy people (9 of 12) |
| Local-global test | Listen (counting helps) | Global-deviant P3 and frontoparietal ignition | Theory-driven; very specific | Misses about half of conscious patients |
| TMS-EEG (Massimini) | Nothing | Complexity and spread of the cortical echo | Task-free; worked in every tested case | Specialised equipment |
| Resting EEG (wSMI, classifier) | Nothing | Long-distance information sharing, spectra, synchrony | Cheap; predicts recovery | Fluctuating patients may be misclassified in one session |
Covert consciousness is common, and now measured at scale. Owen's "vegetative only in appearance" has a name: cognitive motor dissociation (CMD), command-following detected by fMRI or EEG without any behavioural sign. Jan Claassen and colleagues (NEJM, 2019) found EEG evidence of command-following in about 15 percent of clinically unresponsive patients in intensive care after acute brain injury, and those patients recovered better at twelve months. The largest study, Yelena Bodien, Joseph Giacino, Nicholas Schiff and colleagues (NEJM, 2024), tested 353 adults at six international centres; among 241 with no behavioural response to commands, about 25 percent showed CMD on fMRI, EEG or both, more often in younger patients and after trauma. Guidelines followed: the 2018 US practice guideline (Giacino and colleagues) recommends "chronic" instead of "permanent" vegetative state and repeated standardised assessment, and the 2020 European Academy of Neurology guideline recommends adding EEG, fMRI or TMS-EEG to behavioural examination. Independent multicentre work (Denis Engemann, Sitt and colleagues, 2018) showed that the EEG classifier described in this chapter generalises across hospitals and protocols.
Toward Clinical Interventions
Shakespeare's "Canst thou not minister to a mind diseased?" is the families' question. The usual answer after a year in a "permanent" vegetative state is pessimistic. But mental faculties sometimes return after years, and Dehaene sees routes to speed this.
Central thalamic stimulation (Nicholas Schiff, Joseph Giacino and colleagues, Nature, 2007). The targets were the central thalamus and intralaminar nuclei, identified by Giuseppe Moruzzi and Horace Magoun in the 1940s as nodes of the ascending arousal system. Their calcium-binding-protein projection neurons reach the cortex broadly, especially frontal, and target upper-layer pyramidal cells, the long-distance workspace neurons; prefrontal and cingulate cortex in turn regulate them, a loop that adjusts arousal to task demands. Most vegetative patients gain little (Terri Schiavo had a thalamic stimulator without lasting benefit). So the team chose a man minimally conscious for six years with a well-preserved cortex (both hemispheres responded to speech) but low global metabolism, monitored him for months with the Coma Recovery Scale, and implanted electrodes in the central lateral nucleus. Forty-eight hours later, stimulation made him open his eyes, raise his heart rate and turn toward voices, though speech stayed unintelligible. After a two-month off baseline with no change, a double-blind on/off alternation month by month raised scores for arousal, communication, motor control and naming whenever it was on, and they dropped only slightly when it was off. Six months later he could drink from a cup unaided and discuss his treatment.
Spontaneous recoveries also happen. Terry Wallis, minimally conscious for nineteen years, suddenly recovered language and memory; diffusion tensor imaging suggested that long-distance connections had slowly regrown. In another patient, depressed frontal-thalamic communication returned to normal on recovery. Dehaene distinguishes three situations: prefrontal neurons dead (no stimulation will help), neurons intact but disconnected, and circuits intact but switched off because circulating activity cannot sustain itself. The third can recover fast.
The mesocircuit and the sleeping pill
- Dopamine. Dopaminergic neurons project massively to prefrontal cortex and the deep grey nuclei. Three persistent vegetative patients suddenly regained consciousness on levodopa; amantadine slightly sped recovery in controlled trials.
- Ambien (zolpidem). A patient mute and immobile for months with akinetic mutism was given the sleeping pill and woke, moved and spoke. A woman made severely aphasic by a left-hemisphere stroke spoke, answered, counted and named objects for a few hours after each evening dose, then fell asleep and woke aphasic again.
The explanation is a loop: frontal cortex excites the striatum, the striatum inhibits the pallidum, the pallidum inhibits the thalamus, and the thalamus excites cortex. After oxygen loss, the striatum's inhibitory cells are among the first to die; the disinhibited pallidum then shuts down thalamus and cortex. The pathways remain, only suppressed, so a circuit breaker can restart them: a thalamic electrode overcomes the inhibition, dopamine or amantadine excites cortex or surviving striatal neurons, and zolpidem, binding inhibitory (GABA-A) receptors abundant in the pallidum, inhibits the inhibitor. All bring cortex back toward normal activity. A good sign is anatomically intact prefrontal cortex with very low metabolism; once switched on, plasticity can strengthen workspace assemblies until they sustain themselves.
For damaged circuits, Dehaene imagines external loops: implanted interfaces that record prefrontal or premotor discharges and replay them to distant regions, electrically or recoded as sounds or images, as sensory substitution lets the blind "see" through sound. Denser loops might supply the self-excitation a brain needs to stay conscious.
Interventions and communication since 2014. Amantadine has the strongest evidence: Giacino and colleagues (NEJM, 2012) showed in 184 traumatic patients that it sped functional recovery during treatment, and the 2018 US guideline recommends it. Zolpidem helps only a small minority (about 5 percent of patients in the largest series). Schiff and colleagues (Nature Medicine, 2023) implanted central lateral thalamus stimulators in five people with chronic moderate-to-severe traumatic brain injury (conscious, but with impaired attention and executive function) and reported gains in processing speed, extending the 2007 logic beyond disorders of consciousness. Communication devices advanced faster than Dehaene's twenty-year forecast: intracortical speech neuroprostheses in 2023 decoded attempted speech at about 62 words per minute in a man with ALS (Francis Willett and colleagues) and drove a talking avatar for a woman paralysed by a brainstem stroke (Sean Metzger, Edward Chang and colleagues), and in 2024 a system for a man with ALS reached over 97 percent word accuracy (Nicholas Card, David Brandman and colleagues). These serve locked-in patients with clear minds; restoring consciousness itself remains far harder.
The two-door test. Imagine two patients, both clinically vegetative. Patient A passes Owen's tennis test but shows a simple, local TMS echo. Patient B fails every task but shows a high-complexity TMS echo and rich long-distance EEG sharing. Which one is conscious? On workspace theory, A is almost certainly conscious (task-following needs broadcast), and B may be: the capacity for broadcast is present, but no content is shown to be accessed. The case shows why Dehaene wants a battery rather than one test, and why tests that need a task and tests that need nothing answer different questions: one detects conscious access in action, the other the machinery that makes it possible.
Asymmetric tests and decisions with irreversible outcomes. Every consciousness test here is highly specific but insensitive: a positive result is strong evidence, a negative result is weak. Where the downstream decision is irreversible (withdrawing life support), the rational policy is to require multiple independent negatives, repeated over time, and to treat any positive as decisive. The same discipline applies to model-based screening in a bank: know your detector's sensitivity and specificity separately, never let a single negative from a low-sensitivity screen close a case, and combine heterogeneous signals with a learned classifier, as King did, rather than trusting the single most familiar one.
The Future of Consciousness
Dehaene turns the signatures outward. Two-month-old infants already show a conscious workspace, but one running three to four times slower than an adult's; monkeys share illusions, blindsight, the local-global response and metacognition, and so do dolphins and rats in rudimentary form. What may set humans apart is a recursive "language of thought" grafted onto the workspace. Schizophrenia and anti-NMDA-receptor encephalitis look like diseases of long-distance top-down connections, and nothing in principle, he argues, stops a machine with global broadcasting, learning and autonomy from being conscious and exercising free will: Chalmers's hard problem will dissolve like vitalism.
From Lab to Open Questions
David Lodge's novelist resents "science poking its nose into this business, my business"; Nabokov replies that "the greater one's science, the deeper the sense of mystery". The black box is open: images can be made visible or invisible at will, conscious ignition has robust signatures, and those signatures now probe residual consciousness in the clinic. Some remaining questions are empirical (when consciousness emerges in development and evolution), others border on philosophy (self-consciousness, human uniqueness, psychiatric vulnerability, artificial consciousness, free will), but Dehaene believes all will yield to an experimental line of attack.
Conscious Babies?
Dehaene opens with a polemic. Michael Tooley (University of Colorado) wrote that newborns "are neither persons nor quasi-persons" and that infanticide up to three months is not intrinsically wrong, because a newborn lacks "the concept of a continuing self". Peter Singer (Princeton) argues that what matters morally is "rationality, autonomy, and self-consciousness", which infants lack, and likewise judges vegetative patients' lives to have no intrinsic value. Dehaene calls these claims "preposterous": a priori "ukases" issued without evidence, and in the case of vegetative patients demonstrably wrong (Chapter 6). "The brain deserves a better philosophy." His alternative is to search infants for the signatures validated in adults, while admitting the method rests on analogy: markers might change with age, disagree with each other, or mature piecemeal.
The argument with Singer and Tooley misses its target. Both philosophers rest their case on personhood (a concept of oneself as a continuing subject over time), not on whether infants have experiences. Singer in particular grants that infants are sentient and can suffer, which is why his views on infants are framed in terms of the wrongness of killing, not of indifference to pain. Evidence that two-month-olds have conscious access therefore does not refute them; it bears on sentience, which they already accept. Dehaene's evidence matters most against the older medical assumption, which he describes next, that infants feel nothing.
The myth of the mindless infant
Last century's paediatricians saw an immature cortex of "scrawny neurons, puny dendrites" and unmyelinated axons, and concluded that only islands of sensory and motor cortex worked: William James's "blooming, buzzing confusion", with prefrontal cortex silent until about one year, which seemed to explain failure on Piaget's A-not-B task. So newborns supposedly felt no pain, and injections and even surgery were done without anaesthesia. Dehaene's diagnosis: the error was to confuse immaturity with dysfunction. From about six and a half months of gestation the cortex forms and folds; at birth, long-distance fibres already connect distant regions, working slowly without myelin, and they already organise spontaneous activity into functional networks.
Language and working memory at two months
Newborns already distinguish their mother tongue from a foreign language, suggesting learning in the womb. Ghislaine Dehaene-Lambertz (a neuropaediatrician and Dehaene's wife) and Dehaene scanned two-month-olds listening to infant-directed speech.
- An adult-like network. Activity was not limited to primary auditory cortex: the classical left-hemisphere language areas lit up, while equally complex Mozart went to the right hemisphere. Broca's area was already engaged; it later proved one of the earliest-maturing and best-connected prefrontal regions.
- Slow. Activation timing showed the network working much more slowly than in adults, especially in prefrontal cortex.
- Working memory. A sentence repeated after fourteen seconds evoked a stronger Broca's response the second time: infants held it in mind, a hallmark of consciousness.
- Awake versus asleep. Auditory cortex always responded, but activity cascaded into dorsolateral prefrontal cortex only in awake babies; asleep, that region was flat.
Signatures in infants
- Local-global test, adapted. Babies get bored by beeps, so Dehaene-Lambertz turned the test into a multimedia show of faces articulating vowels (aa aa aa ee). At two months, global deviants evoked a global novelty response, a signature of consciousness. But every step was slow: about a third of a second for the unconscious mismatch response, and a full second for the prefrontal global response, three to four times longer than in adults.
- Masking (Sid Kouider). Infants orient to faces from birth. Kouider flashed a face for variable durations between scrambled masks and watched gaze. Five-month-olds looked toward the face only if it lasted more than 100 ms, against 40 to 50 ms in adults (two to three times longer); the threshold reaches adult values at ten to twelve months, when prefrontal behaviours emerge.
- Ignition (Kouider, Dehaene-Lambertz, Dehaene). EEG showed the adult two-stage pattern: linear accumulation in posterior cortex, growing with face duration whether or not it was seen, then a slow, all-or-none frontal negativity only for faces above threshold, starting about one second after onset, the infant analogue of the P3. A similar slow frontal negativity appears whenever infants orient to novelty, and to deviant sounds only when awake.
Infants have a conscious workspace that is wired but not insulated. The major long-distance tracts exist at birth, but myelin keeps maturing into adolescence; without it, conduction is slow, so integration takes seconds. Dehaene likens the infant to a patient emerging from coma: "a foggy, dawdling, but definitely conscious mind".
He tested no one younger than two months, but would "not be surprised" if consciousness is present at birth: newborns a few hours old already discriminate sets by approximate number. Hugo Lagercrantz and Jean-Pierre Changeux propose that birth is the first access to consciousness: the fetus is sedated by placental pregnanolone (a neurosteroid anaesthetic) and prostaglandin D2 (sleep-inducing), and birth brings a surge of catecholamines and stress hormones, after which the newborn is typically wide awake.
Infant consciousness since 2014. Tim Bayne, Joel Frohlich, Rhodri Cusack, Julia Moser and Lorina Naci (2023) reviewed the markers (attention and memory behaviours, the infant P3-like negativity, network integration) and concluded that consciousness is probably present at birth and possibly in the late fetal period, while stressing that no-report concerns (Chapter 1) apply doubly to subjects who cannot report. The medical premise Dehaene attacks has been abandoned: after Kanwaljeet Anand's studies in the 1980s showed that unanaesthetised neonatal surgery produced massive stress responses and worse outcomes, neonatal analgesia and anaesthesia became standard.
Conscious Animals?
Darwin (1838): "He who understands baboon would do more towards metaphysics than Locke." From cheetahs and moray eels to planners like elephants and geese, problem solvers like magpies and octopuses, and social experts like bats and wolves, Dehaene would be "very surprised" if none shared our experiences. Because the workspace is useful for sharing information among brain areas, it likely evolved long ago, perhaps more than once.
- Anatomy. The long-range prefrontal-association network is evident in macaques and may exist in all mammals; even mice activate tiny prefrontal and cingulate cortices to hold visual information for a second. Vocal-learning birds may have analogous circuits.
- Rivalry. Monkeys trained to report with keys show binocular rivalry, alternating at human rhythms.
- Masking. Monkeys report not seeing masked images while visual cortex still fires transiently and selectively: subliminal perception and a sharp threshold.
- Blindsight. With V1 damage, monkeys point accurately to lights in the blind field but press "no light" when asked to report, like human blindsight patients.
- Working memory and rest. Monkeys pass delayed-response tasks with sustained prefrontal and parietal firing; they activate prefrontal cortex more than humans while watching films (we decouple and mind-wander), and they have a default-mode network.
- Local-global. Bechir Jarraya and Lynn Uhrig found that monkey prefrontal cortex responds only to global deviants, and that this response disappears under anaesthesia. Pilot work by Karim Benchenane suggests mice pass too. Dehaene expects all mammals, and probably many birds and fish, to show a convergently evolved workspace.
| Marker | Adults | Infants | Monkeys |
|---|---|---|---|
| Masking threshold | 40 to 50 ms | >100 ms at 5 months; adult value by 10 to 12 months | Present, sharp threshold |
| Linear then all-or-none ignition | P3 from ~270 ms | Frontal negativity from ~1 s | Reported masking and rivalry match human |
| Local-global response | P3 and frontoparietal ignition | Present at 2 months, ~1 s latency | Prefrontal response; abolished by anaesthesia |
| Awake versus asleep or anaesthetised | Lost in sleep and anaesthesia | Prefrontal response only awake | Lost under anaesthesia |
| Working memory | Seconds to minutes | Sentence held 14 s | Sustained prefrontal firing |
| Metacognition | Rich, recursive | Mental-state generalisation at 7 months | Opt-out confidence, generalises across tasks |
Animal consciousness declarations and new evidence. A UK government-commissioned review led by Jonathan Birch (LSE, 2021) found strong evidence of sentience in cephalopod molluscs and decapod crustaceans, which were then included in the Animal Welfare (Sentience) Act 2022. The New York Declaration on Animal Consciousness (April 2024) stated that there is strong scientific support for conscious experience in mammals and birds, and at least a realistic possibility in all vertebrates (reptiles, amphibians, fish) and many invertebrates (at minimum cephalopods, decapod crustaceans and insects). On Dehaene's specific prediction about birds, Andreas Nieder and colleagues (2020, Science) recorded neurons in the pallium of carrion crows that tracked the bird's report of seeing a near-threshold stimulus rather than the stimulus itself, with a late, report-related component reminiscent of the mammalian signatures, from a brain without a layered cortex.
Self-Conscious Monkeys?
Conscious access is the basic capacity we share; Dehaene asks whether the human workspace adds something. Nabokov thought "being aware of being aware of being" was what made us human, a gap between ape and man "immeasurably greater" than between amoeba and ape. Dehaene says he was wrong: the Delphic "Know thyself" is not ours alone. The relevant competence is metacognition, knowing the limits of one's knowledge. Donald Rumsfeld's known knowns, known unknowns and unknown unknowns capture it.
Natua the dolphin (J. David Smith, State University of New York). At the Dolphin Research Center in Marathon, Florida, Natua classified underwater tones by pressing a left paddle for low and a right paddle for high pitch, with the boundary at 2,100 Hz. Near the boundary he slowed, shook his head and often erred, which by itself is ordinary difficulty. Smith then added an escape paddle: pressing it replaced the trial with an easy 1,200 Hz tone and a small reward, but overuse delayed rewards. Natua spontaneously used the escape paddle only on trials near 2,100 Hz, exactly where he would likely be wrong, as if commenting on his own uncertainty. "Like Rumsfeld, he knows what he doesn't know."
Behaviourist reading. The escape key is just a third trained response whose payoff is highest for certain stimuli; reinforcement learning explains its use without any self-knowledge.
Metacognitive reading. In newer studies of monkeys, rats and pigeons, animals choosing to opt out after a decision but before feedback perform worse on opted-out trials than on kept trials with the identical stimulus, so they track internal states such as momentary distraction, not the stimulus. Macaques who learn the key in a perception task immediately use it in a memory task ("I didn't perceive well" becomes "I don't remember well"). The judgements take seconds, use an arbitrary signal and recruit slow evidence accumulation in parietal and prefrontal cortex, all unlikely for unconscious processes.
Dehaene cautions that self-monitoring can be unconscious (typing errors or wrong saccades corrected without awareness), but concludes, provisionally, that "sapiens sapiens" should no longer belong to Homo alone: several species reflect upon their state of mind.
Uniquely Human Consciousness?
Humans still introspect far better. Dehaene's tentative answer: we share core systems with animals, but the human brain can combine them in a recursive "language of thought". Descartes was right that only Homo sapiens composes "words or other signs... to declare our thoughts". Following Noam Chomsky, Dehaene treats language as primarily a representational device for thinking new thoughts, and only secondarily for communication.
- Composition across domains. Thoughts like "taller than Tom", "left of the red door", "not given to John" bind size, person, space, colour, object, logic and action, each from a different brain circuit, with a syntax that separates "my wife's brother" from "my brother's wife" and "dog bites man" from "man bites dog". Animals associate; humans compose.
- Theory of mind. All languages have rich mental-state vocabularies: of the ten most frequent English verbs, six concern knowledge, feeling or goals (find, tell, ask, seem, feel, try); "I" is the tenth most frequent English word and "you" the eighteenth. We represent our own beliefs and others' in the same format. Seven-month-olds already generalise from their knowledge to others', and two-and-a-half-year-olds surpass adult chimpanzees in understanding social events.
- Recursion. "He thinks that I do not know that he lies" may be beyond other primates, whose metacognition seems to have two levels (a thought and a confidence in it) rather than recursion without limit.
- Expanded prefrontal cortex. Large in all primates, vastly expanded in humans, with the largest dendritic trees of any primate.
- Frontopolar cortex (Brodmann area 10). Active in social and self-oriented reasoning; larger in Homo sapiens than in any other ape, and debated even to exist in macaques.
- Prefrontal white matter. Disproportionately large even after correcting for brain size, supporting long-distance connections.
- Broca's area. Its layer 3 long-projection neurons are more widely spaced in humans than in other apes, leaving room for more connections.
- Von Economo neurons. Giant neurons reported in Broca's area and the anterior cingulate, thought to be unique to humans and great apes, with giant bodies and long axons suited to broadcasting.
Von Economo neurons are not a human-and-great-ape monopoly. They have been found in elephants, whales and dolphins, and Henry Evrard, Thomas Forro and Nikos Logothetis (2012) described them in the anterior insula of the macaque, contrary to the claim that they are absent in other primates. They are concentrated in the anterior cingulate and fronto-insular cortex (layer V), adjacent to but not the same as Broca's area, and their function is still unknown (see the Chapter 5 box). Separately, David Chalmers, described here as at the University of Arizona, had moved to the Australian National University and New York University well before 2014.
Two nested evolutions. In all primates, consciousness evolved as a communication device: prefrontal long-distance circuits break modularity and broadcast across the brain. In humans alone, a second evolution added a language of thought for formulating sophisticated, nested beliefs and sharing them. Human consciousness is the outcome of two nested evolutions: a primate workspace for global broadcasting, later boosted by a uniquely human capacity for recursive, compositional thought.
Diseases of Consciousness?
If the workspace rests on specific biology, disease could reverse the evolutionary trend. Long-distance connections are fragile: an axon can be more than a thousand times the size of its cell body and span tens of centimetres, so gene products made in the nucleus must be shipped centimetres to distant synapses.
Schizophrenia as a workspace disorder
Schizophrenia affects about 0.7 percent of adults, typically starting in adolescence or early adulthood, with positive symptoms (delusions, hallucinations) and negative symptoms (reduced intellectual and emotional capacity, disorganised speech, repetitive behaviour). Dehaene and Changeux propose that what unites them is damage to the conscious workspace: the affected functions are exactly its functions (social beliefs, self-monitoring, metacognition, even elementary perceptual access). Patients show dramatic overconfidence in bizarre beliefs, confuse their own thoughts, actions and memories with others', and have poor conscious and metacognitive memory while implicit memory stays intact.
- Raised threshold, intact subliminal processing. Patients need much longer exposure to see a masked word, yet a digit flashed for 29 ms primes them normally: the feed-forward chain from recognition to meaning works; global integration fails.
- Multiple sclerosis parallel. In early MS, a white-matter disease, patients also fail to see flashed words and digits while processing them unconsciously, and the deficit scales with damage to prefrontal-to-visual fibres; a minority of MS patients develop schizophrenia-like psychoses.
- Missing signatures. Early visual and attentional processes can be intact, but the P3 and the long-distance beta-band (13 to 30 Hz) synchrony are deficient.
- Damaged tracts. Diffusion tensor imaging shows anomalies in the corpus callosum and in prefrontal links to cortex, hippocampus and thalamus; at rest, prefrontal cortex loses its hub status.
- Smaller neurons. Layer 2 and 3 pyramidal cells in dorsolateral prefrontal cortex are smaller, with fewer spines.
- Molecules. Many risk genes affect dopamine D2 and glutamate NMDA receptors, central to prefrontal transmission and plasticity.
- Drug models. PCP (angel dust) and ketamine block NMDA synapses, which carry long-distance top-down messages, and induce transient schizophrenia-like psychosis. In Dehaene's simulations, removing NMDA receptors abolished ignition; other models show NMDA receptors are needed for slow evidence accumulation.
Loss of top-down connectivity would spare feed-forward processing and subliminal priming but impair ignition and broadcasting, hence monitoring, top-down attention, working memory and decision: the negative symptoms. For positive symptoms, Paul Fletcher and Chris Frith propose a prediction-error account. Learning normally stops when top-down predictions match bottom-up input and the error signal vanishes. If top-down messages are weakened, errors never go away: the world always seems to hide something, demanding ever more far-fetched explanations.
Ordinarily a predictive signal cancels the sensory consequences of our own movements: grasping a coffee cup feels unremarkable unless it is unexpectedly hot. If prediction fails, even the cup feels subtly wrong, and above all one's own voice sounds strange; from there it is a short step to hearing voices and believing that neighbours or the CIA control one's body. Dehaene stresses that patients are not zombies: conscious broadcasting is impaired more than automatic processes, because "diseases tend to respect the boundaries of the nervous system".
Anti-NMDA-receptor encephalitis (University of Pennsylvania, 2007). Young patients, many of them women with ovarian tumours, arrived with headache, fever or flu-like symptoms, then developed anxiety, agitation, bizarre behaviour, paranoid delusions and hallucinations: an acute, acquired schizophrenia-like state. Within about three weeks, consciousness declined; the EEG slowed as in sleep or coma; patients became motionless, unresponsive and sometimes unable to breathe unaided. Some died; others recovered fully, with no memory of the episode. The cause was an autoimmune attack on the NMDA receptor: patients' serum made NMDA synapses in cultured neurons vanish within hours, returning once the serum was removed. For Dehaene, a single molecule's loss produces first artificial schizophrenia, then loss of consciousness, a possible model disease for the workspace.
Genetics and the NMDA story since 2014. The Psychiatric Genomics Consortium reported 108 schizophrenia risk loci in 2014 and 287 in 2022, with enrichment in genes expressed in neurons and at synapses. The rare-variant SCHEMA study (2022) identified ten genes with large effects, including GRIN2A, which encodes an NMDA receptor subunit, and GRIA3, a glutamate receptor gene, giving direct genetic support to the glutamate-NMDA hypothesis Dehaene invokes. Aswin Sekar, Steven McCarroll and colleagues (2016) linked the strongest common signal, in the complement C4 gene, to excessive synaptic pruning during adolescence, consistent with the loss of prefrontal spines. Anti-NMDA-receptor encephalitis, described by Josep Dalmau's group, is now recognised worldwide and is treatable with immunotherapy and tumour removal, with good recovery in roughly four out of five patients. Dehaene's group later reviewed schizophrenia as a disorder of conscious access (Lucie Berkovitch, Dehaene and Raphaël Gaillard, 2017).
Conscious Machines?
Dehaene sees no logical obstacle to simulating consciousness, and calls it a grand challenge for computer science "over the next decades". His blueprint extends the Chapter 5 simulation into a software architecture: many specialised programs (face recognition, motion detection, spatial mapping, speech, motor guidance), some fed from inside the system for introspection (an error detector predicting deviation from the current goal, as computers already monitor battery, disk and memory).
- Flexible communication. At any moment one program's output is selected for the whole organism and enters a slow, serial, limited-capacity workspace that broadcasts it to every other program: "a sort of universal and autonomous clipboard". Today each application runs in its own memory space and shares only through a user-controlled clipboard.
- Plasticity. Each program learns, by a brainlike rule capturing predictive relations among its inputs, how best to use broadcast information, adapting to the environment and to its own quirks (for instance, the failure of a subprogram).
- Autonomy. A value system decides what deserves slow conscious examination; spontaneous activity keeps sending random "thoughts" into the workspace, kept or rejected by their fit with basic goals, so a serial stream of internal states flows even without input.
Such an organism would set its own goals, explore, learn about its internal states, and at each moment focus on one internal representation, its conscious content. The ideas are "vague", but Dehaene sees "no reason why they would not lead to an artificial consciousness".
The hard problem, according to Dehaene
- Ned Block (NYU) concedes that the workspace may explain access but not qualia, the raw "what it is like".
- David Chalmers distinguishes the "easy problems" (recognising, extracting, reporting: objective mechanisms) from the "hard problem": "how physical processes in the brain give rise to subjective experience", the vivid blue, the distant oboe, the agony of pain.
- Dehaene's reply. Chalmers "swapped the labels": the easy problem is the hard one, and the hard problem only seems hard because it rests on ill-defined intuitions. Educated by neuroscience and simulation, the intuition will fade, and qualia detached from any information-processing role will look like vitalism, the nineteenth-century belief that chemistry could never explain life, which molecular biology dissolved by revealing a self-reproducing automaton. Models of vision already explain why any rational machine facing the same computational problem would suffer our visual illusions.
Chalmers and Block. Even a complete account of access, broadcasting and report leaves open why any of it is accompanied by experience; a functional duplicate without experience (a zombie) is conceivable, so function does not entail feeling.
Dehaene. Conceivability is an intuition, and intuitions about complex machinery are unreliable; as each "easy" problem is solved, the residue labelled "hard" will shrink to nothing, as the "life force" did. Subjective feelings are stable attractors of a brain whose dynamics embed present input in memories and goals.
Free will in a machine
Lucretius asked where free will comes from if atoms never swerve from the chain of causes. Some scientists sought freedom in new physics: John Eccles (Nobel Prize 1963 for synaptic transmission), in books such as How the Self Controls Its Brain, proposed that the immaterial mind tweaks quantum probabilities at synapses; Roger Penrose and Stuart Hameroff proposed the brain is a quantum computer, explaining how mathematicians "see through" Gödel's theorem. Dehaene rejects both as "baroque": quantum computing needs cold to preserve coherence, whereas the brain is a warm bath, and decoherence happens at femtosecond (10⁻¹⁵ s) scales, far from the timescale of awareness. Even if quantum events mattered, randomness is not freedom; Daniel Dennett points out that pure chance gives no freedom "worth having", like the tics of Tourette syndrome.
Autonomy, not indeterminacy. Free will means guiding decisions by higher-level thoughts, beliefs, values and past experience, overriding lower impulses after reviewing options. The workspace collects sensory and remembered information, synthesises, evaluates and ponders it for as long as needed, then acts: that is a willed decision. Decisions are caused by genes, history and learned values; they may be unpredictable because of spontaneous fluctuations, but unpredictability is not the point. "A machine with free will is therefore not a contradiction in terms, just a shorthand description of what we are."
He closes by stressing how far sixteen billion cortical neurons outrun intuition. Each brain's code, sculpted by genes, experience and chance through synapse selection and elimination, is unique to each person and each moment. Its states form a world of inner representations "linked to the environment but not imposed by it", a "remembered present" (Gerald Edelman) thick with memories and forecasts. As you close the book, "ignited assemblies of neurons literally make up your mind".
Machine consciousness after large language models. Dehaene, Hakwan Lau and Sid Kouider (Science, 2017) turned this chapter's blueprint into a test, distinguishing C0 (unconscious computation, where most deep learning sat), C1 (global availability of selected information) and C2 (self-monitoring and confidence), and argued that machines would need both C1 and C2. The 2023 report by Patrick Butlin, Robert Long and colleagues (including Yoshua Bengio, Jonathan Birch and Stephen Fleming) derived fourteen indicator properties from recurrent processing, global workspace, higher-order, attention-schema and predictive-processing theories plus agency and embodiment; the workspace indicators include specialised modules working in parallel, a limited-capacity workspace with selective attention, global broadcast, and state-dependent attention able to query modules in sequence. They judged that no current system was a strong candidate but saw no obvious technical barrier to building one. David Chalmers (2023) estimated the chance that then-current LLMs were conscious at under 10 percent, while expecting successors to be serious candidates within a decade, and Simon Goldstein and Cameron Kirk-Giannini (2024) argued that LLM-based agents with memory and planning modules may already approximate a global workspace. Opposing this, Anil Seth has argued that consciousness may depend on being a living, self-maintaining system, which would make Dehaene's substrate neutrality the real point of dispute. Public controversy (a Google engineer's 2022 claim that the LaMDA chatbot was sentient) and a 2024 report on "AI welfare" (Long, Sebo, Birch, Chalmers and others) moved the question into policy, and some AI developers now run model-welfare research. On C2, language models turned out to be partly calibrated about their own answers (Saurav Kadavath and colleagues, 2022), and interpretability experiments (Jack Lindsey at Anthropic, 2025) found that models can sometimes, unreliably, detect concepts artificially injected into their own activations: limited functional introspection, which says nothing settled about experience.
The Dehaene machine and the zombie. Build Dehaene's three-part machine exactly: specialised learners, a tiny broadcast workspace, a value system and spontaneous "thoughts", plus a C2 monitor that reports its confidence. It passes a machine version of the local-global test, shows ignition-like all-or-none broadcasts, and says "I am not sure I saw that" when its signal is weak. Dehaene would say it is conscious in exactly the sense we are. Chalmers would say we still cannot tell whether anyone is home. Now replace each module, one at a time, with a human volunteer passing notes (the "China brain" variant): functionally nothing changes. If you judge the original machine conscious but hesitate over the paper-passing crowd, your verdict is tracking substrate or speed, not workspace architecture, which is the very assumption Seth and Dehaene dispute.
What this means for AI governance in a bank. The practical question for enterprise AI is not whether a model is conscious but which workspace-like and C2-like properties it has, because those drive behaviour that needs controls: a shared memory that any module can write to (a prompt-injection and data-leakage surface), autonomous goal generation and spontaneous action (needs explicit bounds and approvals), and self-reported confidence (useful only once calibrated against outcomes). Dehaene's architecture is a good map of agentic systems; his C2 point is the operational one: insist on measured calibration and error monitoring, and treat any model's self-reports as data to validate, not as testimony.
Master Takeaways
Taken whole, the book is the most complete case for a functional theory of consciousness built bottom-up from experiments: define the target as conscious access, find its signatures, explain them with a broadcasting architecture, then prove the theory's worth at the bedside. Its lasting achievement is to make "is this person conscious, and of what?" an empirical question with measurable answers, while its boldest bet, that explaining access will leave no further mystery, remains the point on which the field divides.
The takeaways
- Consciousness is global information sharing. To be aware of something is for that representation to enter a limited-capacity workspace that broadcasts it to every specialised processor, so it can be held, reported, remembered and used to plan.
- Conscious access is a late, all-or-none ignition. Seen and unseen stimuli evoke identical early waves; from about 270 to 300 ms the conscious ones trigger a nonlinear, self-amplifying spread into prefrontal and parietal cortex, a phase transition, not a gradual brightening.
- There are four signatures of conscious access. Ignition of frontoparietal networks, the P3 wave, a late sustained gamma burst, and long-distance synchrony with bidirectional information flow; together they index whether a stimulus reached awareness.
- The unconscious is deep but brief and local. Unseen inputs are recognised, bound, understood, valued, attended to and even prepare motor responses, but their effects are small, decay within about a second, stay in local circuits and cannot be chained across steps.
- Consciousness can now be detected at the bedside. The local-global test, TMS-EEG complexity and EEG information sharing reveal consciousness in some "vegetative" patients and predict recovery, sometimes before the clinician can see it in behaviour.
- A broadcasting architecture is substrate-neutral. If consciousness is what the workspace does, nothing in principle stops a machine with global broadcast, learning and autonomous spontaneous activity from being conscious and exercising a meaningful free will.
- Define the target narrowly. Separating vigilance, attention and conscious access made the problem tractable; attention is often a gateway to awareness but operates unconsciously too.
- Use minimal contrasts. Hold the stimulus fixed and let awareness vary (masking, attentional blink, binocular rivalry, inattentional blindness); what differs in the brain is the candidate signature.
- Reports are data, not method. What people say they saw defines the phenomenon; their theories about why they saw it are often confabulation, as the stockings study shows.
- Consciousness collapses probabilities into one sample. Unconscious circuits compute distributions; consciousness draws one interpretation, alternating between percepts in proportion to their likelihood, so the organism can commit and act.
- Consciousness holds and chains. Only conscious information survives seconds of delay robustly (trace conditioning, working memory) and can be passed from one operation to the next, making the conscious mind a slow, serial "human Turing machine".
- Consciousness is a social sharing device. Language compresses experience to about 40 to 60 bits per second; pairs who share answers plus confidence outperform their best member, which is why every conscious content carries a certainty label.
- Correlates are not signatures. Many brain events correlate with awareness, including pre-stimulus fluctuations; a signature must be present whenever consciousness occurs, absent otherwise, and must encode the reported content.
- Content-specific neurons track experience. Anterior temporal concept cells fire for a specific person or place only when the patient reports seeing it, during imagery and recall too, so ignition is not a generic alarm.
- Late long-distance loops are causal. Stimulation can create percepts and disruption can erase them; local early activity is unconscious until it reaches and returns from parietal and prefrontal cortex.
- The workspace is built from long-axon pyramidal neurons. Thick layers II and III, huge dendrites and dense spines in prefrontal hubs, especially in humans, supply the anatomy for broadcasting.
- The P3 mostly says what the thought is not about. Active neurons coding the content are few; the inhibited majority sum to the large positive P3, while focal negativities carry the content.
- The brain is autonomous. Spontaneous ignitions, born of amplified neural noise, form the stream of consciousness, compete with outside input (inattentional blindness), and act as a Darwinian generator of diversity for new plans.
- There are five ways to stay unconscious. Preconscious (unattended but strong), subliminal (too weak), disconnected (no route to the hubs), diluted (never made explicit) and latent (stored in synapses); each needs a different route to awareness.
- Vigilance gates access. A second, nested threshold set by ascending neuromodulators and the thalamus decides whether ignition is possible at all, which is why sleep, anaesthesia and coma abolish access.
- "Vegetative" can hide a conscious mind. Owen's tennis-imagery fMRI found command-following in some vegetative patients; tests are highly specific but insensitive, so negatives prove little.
- A silenced cortex can be switched back on. Central thalamic stimulation, dopaminergic drugs and paradoxically zolpidem act on a cortex-striatum-pallidum-thalamus loop that anoxia tips into self-inhibition.
- Infants are conscious, slowly. Two-month-olds show working memory, awake-only prefrontal responses and a global novelty response, with ignition three to four times slower than adults' because their tracts lack myelin.
- Animals share the workspace and some metacognition. Monkeys show rivalry, masking thresholds, blindsight and the local-global response; dolphins, monkeys and rats opt out when unsure, a rudimentary knowing that they do not know.
- Human uniqueness may lie in recursion. A compositional, recursive language of thought lets humans nest beliefs ("he thinks that I do not know that he lies"), boosted by an expanded, densely connected prefrontal cortex.
- Schizophrenia may be a disease of long-range top-down links. Patients have intact subliminal priming but a raised threshold for access, weaker P3 and long-range synchrony, smaller prefrontal neurons and NMDA-related vulnerabilities; anti-NMDA-receptor encephalitis moves from psychosis to coma.
- Free will is autonomy, not indeterminacy. Quantum randomness would give no freedom worth having; a will is free when conscious deliberation proceeds without impediment to a decision, which a deterministic machine could also do.
Frameworks & mental models
| Name | What it says | Use it when |
|---|---|---|
| Three ingredients | Narrow definition (access), experimental manipulation of awareness, subjective reports as data | Turning a vague mental phenomenon into an experiment |
| Vigilance / attention / access | Three distinct processes with distinct mechanisms; only access is consciousness proper | Diagnosing whether a failure is arousal, selection or awareness |
| Minimal contrast | Same stimulus, different awareness; subtract shared processing | Isolating the neural basis of any subjective state |
| Introspection as data vs method | Trust reports of what was experienced, not explanations of why | Evaluating self-reports from people or models |
| Attentional blink / refractory period | One conscious chunk at a time; others queue in a decaying buffer | Explaining missed events in dual-task situations |
| Hadamard's four stages | Initiation, incubation, illumination, verification | Thinking about where unconscious work fits in problem-solving |
| Consciousness as sampler | Unconscious inference yields distributions; consciousness draws one sample | Explaining ambiguity, alternation and the need to commit |
| Human Turing machine / production system | Parallel unconscious competition, then serial conscious ignition that rewrites the shared state | Analysing multi-step reasoning and its failure without awareness |
| Four signatures | Ignition, P3, late gamma, long-distance synchrony | Judging whether a stimulus reached awareness |
| Correlate vs signature | A signature must be necessary, sufficient and content-encoding | Avoiding over-reading early or pre-stimulus effects |
| Ignition as phase transition | Recurrent excitation gives two stable states and a sharp threshold | Explaining all-or-none perception and its variability |
| Global neuronal workspace | Long-range hubs select, hold and broadcast one content | The core theory; designing broadcast-hub architectures |
| Five ingredients of the code | Cell assemblies, pandemonium, attractors, convergence zones, reentry | Understanding how a conscious content is represented |
| Dehaene-Changeux simulation | Thalamocortical columns, four areas, feedback, vigilance current | Linking architecture to observable signatures |
| Nested phase transitions | Vigilance controls whether ignition is possible; input decides each stimulus | Explaining sleep, anaesthesia and coma |
| Generator of diversity | Spontaneous activity proposes; reward-based evaluation selects | Explaining autonomy, exploration and the stream of thought |
| Taxonomy of the unconscious | Preconscious, subliminal, disconnected, diluted, latent | Diagnosing why information never reached awareness or a decision layer |
| Disorders of consciousness | Brain death, coma, vegetative state, minimally conscious state, locked-in | Clinical classification and prognosis |
| Local-global paradigm | Local deviants give unconscious MMN; global deviants give a conscious P3 | Detecting consciousness without language or instruction |
| Specificity vs sensitivity | Consciousness tests rarely give false positives but often false negatives | Interpreting any asymmetric diagnostic before an irreversible decision |
| Mesocircuit loop | Cortex excites striatum, which inhibits pallidum, which inhibits thalamus, which excites cortex | Understanding how thalamic DBS, dopamine and zolpidem restore arousal |
| Metacognitive opt-out | Choosing to escape hard trials reveals knowledge of one's own uncertainty | Testing self-knowledge in animals, infants or machines |
| Two nested evolutions | Primate broadcasting workspace, then a human recursive language of thought | Explaining what humans share with animals and what they add |
| Prediction-error account of psychosis | Weak top-down predictions leave errors unexplained, breeding delusions | Linking connectivity deficits to positive symptoms |
| Three missing functions for machines | Flexible broadcast, plasticity, autonomy | Assessing whether an AI architecture is workspace-like |
| Free will as autonomy | Freedom is unimpeded conscious deliberation, not randomness | Thinking about agency in people and machines |
Strengths. The book is a model of how to make consciousness science operational: a clear definition, an elegant experimental logic, convergent evidence from scalp, intracranial and single-neuron recordings, a mechanistic theory with simulations, and real clinical payoffs. It is written by the person who ran many of the key experiments, and its catalogue of unconscious processing remains a standard reference. Blind spots. By defining consciousness as access, Dehaene sets aside phenomenal consciousness rather than refuting it; his dismissal of the hard problem as a soon-to-vanish intuition, and of qualia as vitalism, is a philosophical bet, not an empirical result. Nearly every experiment contrasts reported seen and unseen trials, so the signatures may partly reflect what reporting requires (task relevance, decision, working memory). Integrated information theory, recurrent processing, higher-order and predictive theories get little hearing. Contested or dated. The no-report paradigms (Pitts, Frässle, Tsuchiya) showed that the P3b and some frontal activity drop out when stimuli are not task-relevant, shifting attention to an earlier posterior negativity; the Cogitate adversarial collaboration (published 2025) found no ignition at stimulus offset and limited prefrontal coding of some conscious features, a setback for GNW (and a worse one for IIT's posterior-synchrony prediction). Bold subliminal-processing claims elsewhere in the literature (scene semantics and arithmetic under continuous flash suppression, unconscious thought) have replicated poorly, and the theta-burst metacognition result failed replication. Several details need correcting: the Sergent study was 2005, the minimally conscious state was defined in 2002, the pandemonium was Oliver Selfridge's, von Economo neurons exist in macaques, and the human FOXP2 changes are shared with Neanderthals. How the field moved. Clinical detection has grown far beyond the book: the Perturbational Complexity Index and large studies of cognitive motor dissociation (about a quarter of behaviourally unresponsive patients in Bodien and colleagues, 2024) now inform guidelines. Animal consciousness has broadened (the 2024 New York Declaration; crow neurons tracking reported perception). And the machine question is no longer hypothetical: the C1/C2 framework (Dehaene, Lau and Kouider, 2017) and the 2023 indicator-properties report apply workspace criteria to large language models, while critics such as Anil Seth argue that consciousness may require life, not just architecture.
Who's who
| Name | Who | Why they matter here |
|---|---|---|
| Stanislas Dehaene | Cognitive neuroscientist, Collège de France and NeuroSpin | Author; co-founder of the global neuronal workspace theory |
| Jean-Pierre Changeux | Neurobiologist, Institut Pasteur | Co-developer of GNW and its simulations; spontaneous activity, schizophrenia hypothesis |
| Lionel Naccache | Neurologist, Salpêtrière | Co-author of the 2001 manifesto; unconscious number priming; local-global test in patients |
| Bernard Baars | Psychologist | Proposed the cognitive global workspace theory (1989) and minimal contrasts |
| René Descartes | Philosopher, 17th century | Mechanist and dualist; set the two challenges (report, flexible reason) |
| William James | Psychologist, Principles of Psychology (1890) | Attention, stream of consciousness, brain as unstable equilibrium |
| Daniel Dennett | Philosopher | "Fame in the brain"; homunculus fallacy; free will worth wanting |
| Francis Crick and Christof Koch | Neuroscientists | Launched the search for neural correlates via visual illusions; 40 Hz and "coalition" ideas |
| Claire Sergent | Cognitive neuroscientist | Attentional-blink EEG timing of conscious access; later report-free signatures |
| Antoine Del Cul | Researcher, Dehaene lab | Masking phase-transition experiment; raised threshold in schizophrenia |
| Raphaël Gaillard | Psychiatrist | Intracranial markers of conscious access (70% vs 25% of sites) |
| Rafi Malach | Neuroscientist, Weizmann Institute | Late gamma in seen pictures; human single-neuron work |
| Rodrigo Quian Quiroga and Itzhak Fried | Neuroscientist; neurosurgeon | Concept cells ("Jennifer Aniston", "Bill Clinton") that track reported perception |
| Hakwan Lau | Neuroscientist | Equalised-performance prefrontal finding; theta-burst metacognition; C1/C2 paper |
| Moti Salti and Dominique Lamy | Psychologists | Same stimulus, same response, different awareness: P3 isolates access |
| Victor Lamme | Neuroscientist | Local recurrent processing theory, Dehaene's main foil on loops |
| Giulio Tononi | Neuroscientist | Differentiation of conscious states; integrated information theory; PCI |
| Donald Hebb | Psychologist | Cell assemblies (1949), the basis of ignition |
| Oliver Selfridge | AI pioneer | Pandemonium model of competing daemons (1959) |
| Gerald Edelman | Neuroscientist | Reentry; "the remembered present" |
| Edward Vogel | Psychologist, University of Oregon | Parietal negativities tracking working-memory contents |
| Adrian Owen and Martin Monti | Neuroscientists | Tennis-imagery test; yes/no communication with a "vegetative" patient |
| Marcello Massimini | Neuroscientist, Milan | TMS-EEG "ping" of the cortex, later PCI |
| Steven Laureys | Neurologist, Liège | Coma Science Group; metabolism of the workspace in disorders of consciousness |
| Tristan Bekinschtein and Mariano Sigman | Neuroscientists | Co-inventors of the local-global paradigm |
| Jacobo Sitt and Jean-Rémi King | Researchers | wSMI information sharing and machine-learning EEG diagnosis |
| Nicholas Schiff and Joseph Giacino | Neurologist; neuropsychologist | Central thalamic DBS; Coma Recovery Scale; minimally conscious state |
| Jean-Dominique Bauby | Editor of Elle | Locked-in author of The Diving Bell and the Butterfly |
| Ghislaine Dehaene-Lambertz | Neuropaediatrician | Infant language fMRI and infant local-global test |
| Sid Kouider | Cognitive scientist | Masking thresholds and ignition in infants; C1/C2 co-author |
| J. David Smith | Comparative psychologist | Natua the dolphin's opt-out metacognition |
| Paul Fletcher and Chris Frith | Psychiatrist; neuropsychologist | Prediction-error account of hallucinations and delusions |
| Ned Block and David Chalmers | Philosophers | Access vs phenomenal consciousness; the hard problem |
| Roger Penrose, Stuart Hameroff, John Eccles | Physicist; anaesthesiologist; neurophysiologist | Quantum accounts of mind and free will that Dehaene rejects |
| Peter Singer and Michael Tooley | Philosophers | Personhood views on infants that Dehaene attacks |
Connect it
- Broadcast hubs for agentic systems. The workspace is a design pattern: many specialists, one small shared state, competition for write access, broadcast to all readers. Multi-agent AI systems that adopt it gain coordination and auditability (what was selected, when, and why), and they inherit its risks: anything that can write to the shared state can steer every agent.
- Score versus decide. Dehaene's split between fast unconscious scoring and slow conscious commitment maps onto models that score and workflow layers that decide, record and chain steps. Multi-step tasks fail when intermediate results are not explicitly committed and passed on, the "one step yes, two steps no" lesson that chain-of-thought rediscovered.
- Calibrated confidence at every handoff. Bahrami's dyads and the C2 argument point the same way: combining human and model judgements works only when each carries a calibrated confidence, which must be measured against outcomes rather than taken from self-report.
- Asymmetric detectors and irreversible decisions. Consciousness tests are specific but insensitive; credit, fraud and financial-crime screens often are too. Know sensitivity and specificity separately, never let one negative from a weak screen close a case, and combine heterogeneous signals with a learned classifier, as King did for coma patients.
- Correlation versus intervention in explainability. A probe that decodes a concept is a correlate; only intervening on it (ablation, activation patching) shows it drives the decision, which is the standard Dehaene applies to brain signatures and regulators increasingly apply to model explanations.