Insight Editions · No. 286
Being You
A New Science of Consciousness
Anil Seth
Anil Seth argues that consciousness should be approached as a real problem: explain, predict and control the properties of experience through mechanisms in brains and bodies, rather than first asking why experience exists at all.
A study guide to Being You by Anil Seth. 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
Being You is neuroscientist Anil Seth's account of consciousness as a real problem: not "why is there experience at all?" but how to explain, predict and control the properties of experience through mechanisms in brains and bodies. Its thesis is that everything we experience, the world and the self alike, is the brain's controlled hallucination, a best guess about the hidden causes of its signals, and that this predictive machinery exists to keep a living body alive. We are conscious "beast machines": our experiences of world and self arise with, through and because of our living bodies, so consciousness has more to do with being alive than with being intelligent.
Core claim. Perception is inside-out prediction reined in by sensory error; the self is another perception; and both are rooted in interoceptive, control-oriented prediction of the body's physiological state. Consciousness is tied to life, and separable from intelligence.
Scope. Consciousness science from anaesthesia and coma to perception, selfhood, free will, animals and AI, built on predictive processing, Bayesian inference, cybernetics and Friston's free energy principle, with experiments from Seth's Sussex lab and the wider field (Libet, rubber hands, Owen, Massimini).
Key figures. PCI "zap and zip" separating vegetative from minimally conscious patients; Φ and why it can't be measured; The Dress and Adelson's checkerboard; Good Regulator Theorem (1970); Schurger's 2012 reanalysis of Libet; 500 million octopus neurons; 302 in a nematode; 175 billion GPT-3 parameters.
Best for. Readers who want a mechanistic, testable alternative to both the hard-problem impasse and AI-consciousness hype, and anyone building or governing AI who needs to keep intelligence, seeming-consciousness and consciousness apart.
- The method: the real problem. Seth sets aside Chalmers' hard problem and the "isms", dismisses zombie arguments as failures of imagination, and borrows the strategy that dissolved vitalism: split the mystery into properties (level, content, self) and build explanatory bridges from mechanism to phenomenology.
- Level: measuring consciousness. Conscious level is not wakefulness. Brain-based measures of complexity (Massimini's perturbational complexity index, Sussex's spontaneous Lempel-Ziv measures) track sleep, anaesthesia and coma, find covert consciousness, and rise above waking under psychedelics, because every experience is informative and integrated. IIT turns that insight into an identity (consciousness is Φ) that Seth judges untestable; consciousness is more like life than temperature.
- Content: controlled hallucination. The skull-bound brain receives unlabelled signals and infers their causes with Bayesian best guesses, implemented as prediction error minimisation with generative models, hierarchies and precision weighting (attention). Experience is the content of the predictions; hallucination is uncontrolled perception. Action is active inference, self-fulfilling prediction. Objecthood, change, time, reality and causality are all perceptual inferences, which is why perceptions seem to be "out there" and why the hard problem seems harder than it is.
- Self: another perception. The self splits into bodily, perspectival, volitional, narrative and social levels, each manipulable (rubber hands, full-body illusions, body swaps) or breakable (amnesia, delirium, depersonalisation). At the bottom lie emotions as interoceptive inference and a formless feeling of being alive. By the Good Regulator Theorem, perception of the body is for control, not discovery: we perceive ourselves in order to control ourselves, which also explains the self's felt stability.
- Life and physics. The free energy principle retells the story from the bottom: to exist is to stay in expected states, which means minimising free energy, which equals prediction error: "I predict myself therefore I am". It is a theory for, not of, consciousness.
- Free will. Volition is perception of an action's causes as coming from within; Libet's readiness potential is likely threshold-crossing noise (Schurger); real freedom is the competence to control many degrees of freedom, and the feeling "I could have done otherwise" exists so that next time, you might.
- Other minds. All mammals are likely conscious; birds probably; the octopus is a distributed, alien mind; further out, injury responses are the best clue. Intelligence is neither necessary nor sufficient for consciousness, so smarter AI will not thereby be sentient; the real near-term danger is machines that seem conscious (the Garland test), plus the ethics of organoids and of any artificial suffering.
| Ch | Title | One-line payload |
|---|---|---|
| P | Prologue | Anaesthesia turns people into objects and back; the thesis: experiences of world and self are controlled hallucinations rooted in a living body. |
| 1 | The Real Problem | Consciousness is phenomenology; replace the hard problem with the real problem of explaining, predicting and controlling its properties, as biology did with life. |
| 2 | Measuring Consciousness | Level is not wakefulness; complexity measures (PCI) track it, detect covert consciousness, and rise under psychedelics; experience is informative and integrated. |
| 3 | Phi | IIT identifies consciousness with Φ, but Φ needs a system's full repertoire of possible states and is untestable; keep the insight, drop the identity. |
| 4 | Perceiving from the Inside Out | Perception is the brain's top-down best guess, checked by prediction errors: controlled hallucination; The Dress, Adelson, Mooney images. |
| 5 | The Wizard of Odds | The brain approximates Bayes by prediction error minimisation; attention is precision; action is active inference. |
| 6 | The Beholder's Share | Expectations shape seeing; the hallucination machine; objecthood, change, time, reality are inferences; "seeming real" fuels the hard problem. |
| 7 | Delirium | His mother's hospital-induced delirium shows the self can fragment and return, opening Part III. |
| 8 | Expect Yourself | The self has five levels, all perceptions; rubber hand, full-body and body-swap illusions; Clive Wearing; the social self; self-change blindness. |
| 9 | Being a Beast Machine | Emotions are interoceptive inference; by the Good Regulator Theorem perception of the body is for control; the self begins in the feeling of being alive. |
| 10 | A Fish in Water | The free energy principle: to exist is to stay in expected states by minimising free energy = prediction error; a theory for, not of, consciousness. |
| 11 | Degrees of Freedom | Free will is a self-related perception; Libet reinterpreted by Schurger; freedom as control of degrees of freedom; real, not illusory. |
| 12 | Beyond Human | All mammals conscious; birds likely; the octopus as an alien, distributed mind; consciousness tracks life, not intelligence. |
| 13 | Machine Minds | Functionalism and the intelligence assumption are unjustified; a silicon beast machine is probably not conscious; the danger is AI that seems conscious. |
| E | Epilogue | A hemispherotomy, the three post-Copernican challenges, and consciousness science as the third displacement of human exceptionalism. |
Prologue
Seth opens with general anaesthesia, the routine medical act of "turning people into objects" and back again, as the sharpest proof that consciousness depends on the stuff we are made of. He sets out what a science of consciousness should aim for (explaining its properties through mechanisms in brains and bodies, not first explaining why it exists at all) and previews the book's destination: experiences of world and self are controlled hallucinations rooted in a living body.
Ceasing to Exist
Five years before writing, during a small operation, Seth ceased to exist for the third time in his life. General anaesthesia is not sleep: a surgeon's knife would wake a sleeper. Deep anaesthesia resembles coma and the vegetative state, in which consciousness is wholly absent, and under it the brain's electrical activity is almost entirely quietened, something that never happens in ordinary waking or sleeping life.
The return was the striking part. Waking from deep sleep always carries some sense that time has passed, a continuity between then and now. After anaesthesia there was none: he could have been under five minutes or fifty years. He was simply not there, a premonition of the total oblivion of death that he found strangely comforting. Anaesthesia, he notes, acts on consciousness itself, not merely on brain or mind: by shifting the electrochemical balance of neural circuits, the basic ground state of "being" is temporarily abolished.
Anaesthesia is the art of turning people into objects, and then back into people. That a drug can switch subjective existence off and on is the plainest evidence that experience depends on the material brain, and it frames the central mystery: how billions of neurons, each a tiny biological machine, produce your experience, here and now.
How Seth Got Here
A childhood memory anchors the question. At eight or nine, looking in a bathroom mirror, he realised his experience would one day end, and therefore that consciousness must depend on his body and brain. The career that followed is a compressed history of the field's own rehabilitation:
- Cambridge, early 1990s. Undergraduate physics and philosophy broadened into psychology and neuroscience, fields that then avoided, even outlawed, talk of consciousness.
- PhD detour. Research in artificial intelligence and robotics proved unexpectedly valuable later.
- The Neurosciences Institute, San Diego. Six years with Nobel laureate Gerald Edelman, one of the figures most responsible for making consciousness a legitimate scientific target.
- Sussex. For over a decade, Co-Director of the Sackler Centre for Consciousness Science at the University of Sussex, near Brighton, a mix of neuroscientists, psychiatrists, imagers, VR specialists, mathematicians and philosophers.
Why Consciousness Matters
For each of us, conscious experience is all there is: without it there is no world, no self, no inside or outside. Seth makes the stakes concrete with an offer.
The brain-replacement deal. A future Seth offers to swap your brain for a machine that is its functional equal in every way, immune to decay, perhaps immortal. Nobody could tell the difference from outside, and new-you would say "yes" when asked whether it is conscious. The catch: since nobody yet knows how brains produce experience, he cannot guarantee you will experience anything. If consciousness depends only on functional organisation you are fine; if it depends on specific biological material such as neurons, there may be nothing it is like to be you for the next five hundred years. Most people refuse. The refusal shows that we already treat behavioural equivalence and experience as separable, which is exactly the question the book's final part returns to.
The practical case is just as strong. General anaesthesia ranks among the greatest inventions in history. Disturbances of consciousness accompany brain injury and mental illness, which Seth notes he has encountered himself. Across a life, experience changes from infancy's "blooming, buzzing confusion" through adult clarity (apparent, probably illusory, and not universal) to the dissolution of self in neurodegeneration. The assumption of a single, unique conscious self persisting through time (a soul) may be grossly mistaken.
- Development. When does consciousness begin: at birth, or already in the womb?
- Other animals. Primates and mammals, but also the octopus, nematode worms, bacteria: is there anything it is like to be E. coli or a sea bass?
- Machines. Beyond AI's power over us, when do we owe future machines an ethical stance? Seth recalls his uncanny sympathy for HAL in 2001: A Space Odyssey as its memory banks are pulled, and the deeper empathy for the replicants of Blade Runner, a hint that being a living machine matters for selfhood.
- Self. Is consciousness possible without self-consciousness, and if so would it matter as much?
What a Science of Consciousness Should Do
Seth expects no single eureka moment and no "decoding" like the human genome or a consensus like climate science. His goal is narrower and more tractable: to explain how the properties of consciousness depend on the operations of the neuronal "wetware". The aim is not primarily to explain why consciousness is part of the universe, nor to map the brain in full while sweeping the mystery under the carpet. Accounting for properties through mechanisms should make the deep metaphysical "whys" progressively less mysterious.
The word "wetware" is deliberate: brains are not computers made of meat. They are chemical machines as much as electrical networks, and every brain that has ever existed has been part of a living body embedded in an environment, often one containing other embodied brains.
The Cartesian tradition. From Descartes in the seventeenth century, animals lacked conscious selves because they lacked rational minds; they were "beast machines", flesh automatons unable to reflect on their existence.
Seth's reversal. Consciousness has more to do with being alive than with being intelligent. We are conscious selves precisely because we are beast machines: the experience of being you emerges from how the brain predicts and controls the internal state of the body.
Seth's thesis is that conscious experiences of the world and the self are forms of brain-based prediction, "controlled hallucinations", that arise with, through, and because of our living bodies.
The Plan of the Book
- Part I: Level. Seth's approach to the science, how conscious someone or something can be, and attempts to measure consciousness.
- Part II: Content. What you are conscious of: perception as brain-based best guessing.
- Part III: Self. The experience of being a self, in all its varieties, grounded in the body.
- Part IV: Other. What this view implies for non-human animals and the possibility of sentient machines.
The Fourth Dethronement
Freud, whatever his standing among neuroscientists, identified three scientific "strikes" against human self-importance, each fiercely resisted. Seth accepts the frame and extends it.
- Copernicus. The heliocentric model removed us from the centre of the universe: we became a "pale blue dot".
- Darwin. Common ancestry with all living things, a fact still resisted in places today.
- Freud. The unconscious mind undermined the idea that mental life is under conscious, rational control; wrong in detail, right that a naturalistic account of mind would be a further dethronement.
Seth welcomes each such shift: every advance brings new wonder and lets us see ourselves as less apart from, and more a part of, nature. Conscious experience is part of nature as our bodies are, so when life ends, consciousness ends too. He closes back at the anaesthetic oblivion and with Julian Barnes' meditation on mortality: when the end comes, there is nothing, really nothing, to be frightened of.
Framing to keep in view. The Freud "three strikes" story (from his 1917 essay on a difficulty of psychoanalysis) is a rhetorical device; historians note the Copernican shift was not experienced at the time as a demotion (the centre was then regarded as the lowest, basest place in the cosmos). More substantively, Seth's claim that consciousness tracks being alive rather than intelligence is a minority position among theorists, many of whom are functionalists; the book argues for it from Part III onward rather than establishing it here.
The Real Problem
Opening Part I (Level), Seth defines consciousness as phenomenology, any subjective experience at all, and clears the ground of philosophical "isms" and zombies. His alternative to Chalmers' hard problem is the real problem: to explain, predict and control the phenomenological properties of experience in terms of brain and body mechanisms. The model is the history of life science, where dividing "life" into tractable properties made the mystery of vitalism evaporate; consciousness is divided likewise into level, content and self.
Part I: Level
The book's first part concerns conscious level: how conscious a person (or animal, or machine) is, from coma to vivid waking life, and whether that level can be measured. This chapter lays down the method used throughout; chapters 2 and 3 apply it to measurement and to integrated information theory.
What Consciousness Is
Seth adopts Thomas Nagel's formulation from the legendary 1974 paper "What is it like to be a bat?": for a conscious creature there is something it is like to be that creature. There is something it is like to be you, probably to be a sheep or a dolphin, and almost certainly nothing it is like to be a bacterium, a blade of grass or a toy robot. Nagel's point was that although we could never have a bat's experiences, there would still be something it is like for the bat.
The approach puts phenomenology first: the subjective properties of experience, why a visual experience has the form, structure and qualities it does compared with an emotion or a smell. Philosophers call these properties qualia: the redness of red, the pang of jealousy, the throb of toothache. Wherever there is experience there is phenomenology, and wherever there is phenomenology there is consciousness, even for a creature that exists for one moment of pleasure or pain.
Consciousness is first and foremost about phenomenology, not about function or behaviour. Past confusions equated it with language, intelligence or particular behaviour, but dreaming and total paralysis show it does not need outward behaviour, and a language requirement would deny experience to babies, aphasic adults and most animals.
Seth distinguishes phenomenology from the functional and behavioural properties of consciousness (its roles in cognition and the behaviour it enables). He sets aside two leading theories precisely because they start from function:
Global workspace theory (Baars, Dehaene). Content becomes conscious when it gains access to a "workspace" distributed across frontal and parietal cortex and is broadcast widely, enabling flexible behaviour. Consciously seeing a glass of water, you can drink it, throw it at the computer, write a poem about it or return it to the kitchen; unconscious perception allows no such flexibility. In Ned Block's terms this is a theory of access consciousness rather than phenomenal consciousness.
Higher-order thought theory. Content becomes conscious when a higher-level cognitive process is directed at it, tying consciousness to metacognition ("cognition about cognition"). It also emphasises frontal regions, though it leans on function less than the workspace theory. Seth says little more about either because he starts from experience itself and only then reaches function.
The definition, "any kind of subjective experience whatsoever", is deliberately simple. Premature precision constrains and misleads when a phenomenon is poorly understood; good definitions co-evolve with understanding, as the definition of the gene changed with molecular biology.
The Hard Problem and the Easy Problems
David Chalmers coined the "hard problem" in the early 1990s, and it has set the field's agenda since. The question is why physical processing in organisms that are undeniably subjects of experience should be accompanied by experience at all: the quality of deep blue, the sensation of middle C. Chalmers grants that experience arises from a physical basis but says we have no good explanation of why and how.
The easy problems concern explaining functions and behaviours: sensory processing, action selection, attention, language, anything specifiable as an input transformed into an output. They are not easy in practice (they will occupy neuroscience for decades or centuries) but there is no conceptual obstacle to mechanistic explanation. The hard problem, by contrast, seems to survive even after every function near experience (discrimination, categorisation, internal access, verbal report) has been explained. Seth defines a mechanism as a system of causally interacting parts that produce effects.
- The meta-problem (Chalmers, 2018). Why do people think there is a hard problem? This is actually an easy problem, because it asks about behaviour: the verbal reports of people who profess belief in a hard problem.
- Why Seth likes it. It can be studied whatever metaphysical stance one holds about consciousness itself.
The "Isms"
The roots of the hard problem run back to ancient Greece but are clearest in René Descartes' seventeenth-century split between res cogitans (mind stuff) and res extensa (matter stuff), which inaugurated dualism and has confused discussion since.
| Position | Core claim | Seth's verdict |
|---|---|---|
| Physicalism (materialism) | The universe is physical; conscious states are identical to or emerge from arrangements of physical stuff | His preferred stance and the default of most neuroscientists |
| Idealism (Berkeley) | Mind is the ultimate reality; the puzzle is how matter emerges from mind | The opposite extreme; not pursued |
| Dualism (Descartes) | Mind and matter are separate substances; problem of how they interact | Few now endorse it, but "naive dualism" persists because experience seems non-physical |
| Functionalism | Consciousness depends on what a system does (input-output mappings), not what it is made of | Agnostic and slightly suspicious |
| Panpsychism | Consciousness is fundamental and ubiquitous, like mass or charge | Explains nothing, yields no tests: an empirical dead end |
| Mysterianism (McGinn) | A physical explanation exists but is cognitively closed to humans | Unjustifiably pessimistic |
| Illusionism (Frankish) | Phenomenal properties are an introspective misrepresentation | Rejects the "experience doesn't exist" reading; sympathetic to "experience exists but is not what it seems" |
Why Seth Distrusts Functionalism
Functionalism smuggles in the phrase "information processing", which hides strong assumptions: that the brain is a computer with mind as software ("mindware"), and that we know what information is, a question nearly as vexing as consciousness. Brains are very different from computers, at least the kinds we know. Taken at face value, functionalism implies that simulation is instantiation: get the input-output mappings right and consciousness comes into being.
Go or weather? DeepMind's AlphaGo Zero is really playing Go: for a game, simulation is instantiation. A computer simulation of a storm, however detailed, never gets wet or windy. The question is which kind of thing consciousness is. Seth offers no answer, only the point that the question is real, which is why he stays agnostic. (He notes Searle's Chinese room makes a related argument but targets understanding rather than consciousness.)
Panpsychism and Mysterianism
Mockery of panpsychism (conscious spoons) usually misconstrues its sophisticated versions; the real problems are that it explains nothing and generates no testable hypotheses, offering an easy exit from the hard problem. Seth quotes John Perry: think about consciousness long enough and you "either become a panpsychist or you go into administration", perhaps because people only think rather than experiment.
Mysterianism, associated with Colin McGinn, says the solution is as far beyond us as cryptocurrency is beyond frogs. Seth grants that finite brains must leave some truths beyond any human; nobody fully understands an Airbus A380, yet he happily flew one home from Dubai. But pre-emptively putting consciousness in that zone is unwarranted, because science is cumulative: mystery after mystery has yielded to reason and experiment. Taking mysterianism seriously means giving up and going home.
The test for an "ism" is usefulness, not provable truth. Seth settles on a functionally agnostic physicalism, which he calls the most pragmatic and the most intellectually honest stance.
Getting Rid of Zombies
A philosophical zombie is indistinguishable from a conscious creature but lacks experience. Zombie Anil would look, walk and talk like Seth, would answer "Yes, I'm conscious" and would even have written essays doubting zombies' relevance, all without any inner life.
- Premise 1. We can conceive of a world physically identical to ours with no consciousness in it.
- Premise 2. Whatever is conceivable is possible.
- Conclusion. So consciousness is not fixed by the physical facts, and physicalism is false.
- Reply. Conceivability arguments are intrinsically weak; their plausibility is inversely related to how much the conceiver knows.
The backwards A380. Picture a giant airliner flying backwards: easy. But the more aerodynamics you know, the less conceivable it becomes; planes simply cannot do it (helicopters can, but a helicopter, from "helico" (spiral) + "pter" (wing), is not a plane). Imagining a zombie is likewise easy only superficially. Really conceiving one means grasping 86 billion neurons with roughly a thousandfold more connections (counting one per second would take nearly 3 million years), plus glia, neurotransmitter gradients and a body in a world of other brains. Nobody can. Conceivability is usually a fact about the conceiver, not about reality.
Zombie defenders reply that only logical possibility matters, not conceivability under our physics. Seth rejects this: a backwards A380 may be logically possible under alternative aerodynamics, but that sheds no light on how a real A380 flies, and he wants to know how a real brain in this universe produces experience. He also notes two zombie varieties: behavioural zombies (identical from outside) and neurological zombies (identical inside too, same wetware), both entirely lacking consciousness.
The Real Problem
According to the real problem, the primary goals of consciousness science are to explain, predict, and control the phenomenological properties of conscious experience in terms of physical mechanisms in the brain and body.
The task is to show why a particular pattern of brain activity maps to a particular experience, not merely that it does. It differs from the hard problem because it does not first ask why consciousness exists at all or hunt for a "special sauce". It differs from the easy problems because it targets phenomenology, not function. And its focus on mechanisms aligns it naturally with physicalism.
| Approach | What "explaining redness" means | What it leaves out |
|---|---|---|
| Easy problem | How wavelengths activate the visual system, when we say "that is red", behaviour at traffic lights, emotional responses to red | Why any of this is accompanied by experience |
| Hard problem | Why there is experience of redness rather than none | Any mechanistic answer always leaves a residue: "fine, but why is this mechanism conscious?" (the explanatory gap) |
| Real problem | Why redness has its phenomenological properties: visual, usually attached to objects and surfaces, varies in saturation, forms a category among colours with smooth variation inside it | Postpones (does not abandon) the existence question |
The goal is to learn what about specific brain activity, such as the looping activity in visual cortex (occipital lobe), explains why redness is like this and not like blueness, toothache or jealousy. Explanation, prediction and control are the criteria applied to every other science. Physics has explained, predicted and controlled a great deal while remaining flummoxed about why the universe exists; quantum mechanics is accepted despite being counterintuitive. A mature consciousness science may likewise work without ever giving the intuitive feeling that "of course it has to be this way".
The real problem is not a surrender to the hard problem; it goes after it indirectly. Hurley and Noë's terms make this precise: the real problem works through "comparative" explanatory gaps (why experiences differ as they do) to resolve or dissolve the "absolute" gap (why there is phenomenology at all). Chalmers himself floated similar ideas as the "mapping problem" and the "principle of structural coherence", and Francisco Varela's neurophenomenology matched experience to brain dynamics.
The Neural Correlates of Consciousness
In 1989, psychologist Stuart Sutherland wrote in the International Dictionary of Psychology that consciousness was impossible to specify and that "nothing worth reading has been written on it". Meanwhile in San Diego, Francis Crick (co-discoverer of DNA's structure with Rosalind Franklin and James Watson) and Christof Koch were launching the field's dominant method. (Daniel Dennett's Consciousness Explained, 1991, was a turning point for the young Seth.)
The gold-standard definition: an NCC is "the minimal neuronal mechanisms jointly sufficient for any one specific conscious percept". The recipe is practical: create conditions where people sometimes have an experience and sometimes not, with everything else matched, and compare brain activity with fMRI (blood oxygenation; good spatial detail, indirect) or EEG (electrical signals of large populations; direct and fast, poor spatial precision).
Binocular rivalry. Show a face to one eye and a house to the other. Perception does not settle on a chimera; it flips between face and house every few seconds while the input stays constant. Brain activity that tracks the flips rather than the stimulus marks the NCC for that percept. Subtlety: rivalry studies often highlight regions linked to transitions, and perception of change is not the same as change of perception (a theme of chapter 6).
The method has been productive, but two limits are now clear:
- Confounds. It is hard, perhaps impossible, to separate a true NCC from its prerequisites (such as paying attention) and consequences (such as verbal report). Papers in 2012 (Aru et al.; de Graaf et al.) finally crystallised the issue. "No-report" paradigms, in which volunteers make no behavioural report, often find NCCs that exclude the frontal regions central to global workspace and higher-order theories.
- Correlation is not explanation. Correlation falls short not only of causation but of explanation. The NCC strategy and the hard problem are natural bedfellows: collect only correlations and an explanatory gap will always seem to remain. Build bridges from properties of mechanisms to properties of experience and the gap should narrow and perhaps disappear, the hard-problem intuition vanishing "in a puff of metaphysical smoke".
The correlates hunt has turned into theory-testing. In June 2023, results from the first large adversarial collaboration (the Cogitate Consortium, funded by the Templeton World Charity Foundation) pitted global neuronal workspace theory against integrated information theory with preregistered predictions, using fMRI, MEG and intracranial recordings; the full paper appeared in Nature in 2025. Neither theory came through cleanly: key IIT predictions about sustained posterior synchronisation failed, and GNW's predicted prefrontal "ignition" at stimulus offset was not found. The same week, Christof Koch conceded his 1998 bet with Chalmers that the NCC would be clearly identified within 25 years. No-report designs are now standard, and the field's own reviewers increasingly echo Seth's point that correlations alone will not settle theory choice.
Lessons from Life
Not long ago life seemed as mysterious as consciousness does now. Vitalism peaked in the nineteenth century, backed by Johannes Müller and Louis Pasteur, and lingered into the twentieth: being alive supposedly required a spark, an élan vital. No special sauce was needed. The fatal flaw of vitalism was to interpret a failure of imagination as an insight into necessity, the same flaw Seth finds in the zombie argument. (Even today, most pre-school children prefer vitalist explanations; Patricia Churchland has developed the parallel.)
Biologists made progress by tackling the "real problems" of life: describing properties and explaining, predicting and controlling each mechanistically.
- Properties explained. Reproduction, metabolism, growth, self-repair, development and homeostatic self-regulation all yielded to physics and chemistry, individually and together.
- The concept ramified. "Alive" stopped being all-or-nothing: viruses, synthetic organisms and even collections of oil droplets occupy grey zones with some but not all properties of life.
- Optimism. Today's consciousness researchers may stand where biologists stood a few generations ago.
- Disanalogy. Life's properties are objectively describable; consciousness's are first-person. Seth thinks this mainly makes the data harder to collect, not the project impossible.
The yearning for a single "humdinger eureka" may explain the appeal of quantum theories of consciousness, mostly descended from Roger Penrose's The Emperor's New Mind (1989). Seth sees in them a false syllogism: quantum mechanics is mysterious, consciousness is mysterious, therefore they are related.
Level, Content, Self
- Conscious level. How conscious a being is, from total absence (coma, brain death) to vivid waking awareness. Part I.
- Conscious content. What we are conscious of: sights, sounds, smells, emotions, moods, thoughts, beliefs. All contents are varieties of perception, brain-based interpretations of sensory signals, and perception can be conscious or unconscious. Part II.
- Conscious self. The specific experience of being you: a body, a first-person perspective, unique memories, moods, emotions, "free will". It is a subset of contents, and must not be confused with consciousness itself: self-consciousness is not consciousness. Part III.
These are not independent, and how they relate is a challenge in its own right. Dividing the problem gives distinct targets for mechanistic explanation and pushes back against treating consciousness as one intimidating thing. There are as many ways of being conscious as there are conscious organisms. Eventually the hard problem may itself succumb, letting us see consciousness as continuous with the rest of nature without any "ism" declaring by fiat how physics and phenomenology relate.
Where the real problem is open to challenge. Chalmers would reply that explaining every structural property of experience still leaves the existence question untouched; the life analogy works because "life" was always a functional concept, whereas the hard problem is defined as what remains after function. Seth's answer is a wager, not a proof: that the absolute gap will dissolve as comparative gaps close. His dismissal of panpsychism as untestable is also contested; defenders such as Philip Goff (Galileo's Error, 2019) argue the same charge applies to physicalism at the level of intrinsic nature. And the historical claim about Pasteur needs care: his vitalism concerned fermentation requiring living cells, a view later overturned by Eduard Buchner's cell-free fermentation (1897).
For AI architects. The functionalism question (Go or weather?) is now practical. Behavioural equivalence in a language model or copilot is the easy-problem layer; claims of machine experience need either a theory that says function suffices or evidence about the substrate. Seth's agnosticism is a sensible governance default: do not treat fluent self-reports ("I am conscious") as evidence, since zombie Anil would say the same.
Measuring Consciousness
The history of the thermometer shows how measurement can turn a mystery into a science; Seth asks whether a "consciousness meter" could do the same. He separates conscious level from wakefulness, then follows the new brain-based measures: Massimini's "zap and zip" perturbational complexity index (PCI), Sussex's measures of spontaneous signal complexity, Owen's "tennis and house" imagery test. The psychedelic results show that level and content are not independent, and Tononi and Edelman's insight that every experience is both informative and integrated points to complexity as the middle ground between order and disorder.
The Lesson of Temperature
In seventeenth-century Paris, a deep cool cellar beneath the Observatoire was proposed as a fixed point for temperature. Natural philosophers faced a circle: you cannot validate a thermometer without a scale, nor build a scale without a reliable thermometer. Candidates such as the boiling point of water varied with air pressure and even the roughness of the glass; Joachim Dalencé proposed the melting point of butter.
Reliable mercury thermometers eventually displaced "calorific" theories (heat as a fluid substance) and enabled thermodynamics, associated with Ludwig Boltzmann and Lord Kelvin. Temperature became mean molecular kinetic energy, and heat became energy transferred between systems at different temperatures. The theory did more than establish a correlation: it said what temperature is. That allowed talk of the sun's surface temperature and an absolute zero, and substance-based scales (Celsius, Fahrenheit) gave way to a physically grounded one (Kelvin). Seth learned the story from Hasok Chang's Inventing Temperature.
Measurement turns the qualitative into the quantitative and the vague into the precise. Even if consciousness proves less like temperature (one underlying quantity) and more like life (many properties), precise measurement is still how explanatory bridges get built. As early as 1870, Thomas Henry Huxley predicted "a mechanical equivalent of consciousness, just as we have arrived at a mechanical equivalent of heat".
At a 1990s conference, David Chalmers pointed an old hairdryer at his head to show how useful a consciousness meter would be: point it at a person, animal or machine and read off the answer. But a meter's real value is quantitative experiment, not yes/no verdicts. If consciousness is like temperature, the payoff would be spectacular: meaningful levels and degrees, even for exotic non-human forms.
Practical Stakes and the Bispectral Index
- Anaesthesia. Applied to millions of people daily (Seth says more than four million a day), anaesthesia must keep patients in oblivion without overdose, often alongside neuromuscular blocking agents that paralyse the body, so outward signs of waking vanish.
- Brain injury. Patients receive grave diagnoses such as "vegetative state" or "minimally conscious state", and deciding whether consciousness remains is urgent.
- The bispectral index (BIS). The most common theatre monitor combines several EEG measures into one continuously updated number; details are hidden in patents. It has been controversial because readings sometimes conflict with behavioural signs (patients opening eyes, recalling surgeons' talk), and, more deeply, because it rests on no principled theory.
A number to check. The "four million a day" figure is attributed to Weiser et al. (2008), but that study estimated about 234 million major operations a year worldwide, roughly 640,000 a day, and not all of those use general anaesthesia. The general point (anaesthesia is routine at enormous scale) stands; the specific figure looks inflated by several times.
Level Is Not Wakefulness
Conscious level (awareness) and wakefulness (arousal) usually move together but can come apart, which shows they rest on different biology. In dreaming you are asleep yet richly conscious. In the vegetative state, now also called unresponsive wakefulness syndrome, patients cycle through sleep and waking but show no behavioural sign of awareness: "the lights are occasionally on but there's nobody home".
What in the brain tracks level rather than arousal?
- Not neuron count. The cerebellum has about four times as many neurons as the rest of the brain combined, yet seems barely involved; people with cerebellar agenesis (no normal cerebellum) live largely normal, presumably conscious, lives.
- Not overall activity. Energy use differs only slightly across levels; the brain never "shuts down" when consciousness fades.
- How regions talk. What matters is communication within the thalamocortical system: the cortex plus the thalamus, oval nuclei below the cortex densely wired to it. The newest measures quantify exactly these interactions, ideally as a single number, like a thermometer.
Zap and Zip: The Perturbational Complexity Index
Marcello Massimini, first with Giulio Tononi at the University of Wisconsin-Madison and later with his own group at the University of Milan, combined transcranial magnetic stimulation (TMS) with EEG: a precisely controlled electromagnet injects a sharp pulse through the skull and EEG records the response, "banging on the brain with an electrical hammer and listening to the echo". People rarely notice the zap unless it triggers a movement (motor cortex), a phosphene (visual cortex) or painful scalp-muscle spasm; a huge disturbance of neural activity usually changes nothing in experience, a reminder that we are not aware of what our neurons do.
- Unconscious states. In dreamless sleep and general anaesthesia the echo is simple: a strong local response that dies quickly, like ripples from a stone in still water.
- Conscious states. The echo ranges widely over the cortex, disappearing and reappearing in complex patterns, implying richer thalamocortical communication.
The complexity is quantified by compression. Any pattern can be written as 1s and 0s; any non-random sequence has a shorter description that regenerates it. The length of the shortest possible description is the algorithmic complexity (formally Kolmogorov-Chaitin-Solomonoff complexity): lowest for a fully predictable sequence (all 1s: "a 1, repeated n times"), highest for a random one (which cannot be compressed at all), intermediate for structured ones. Lempel-Ziv-Welch (LZW) complexity, the algorithm used to zip photo files, approximates and upper-bounds it. Applied to the TMS echo, this gives the perturbational complexity index (PCI).
Massimini's landmark 2013 study (Casali et al.). After validating PCI as reliably lower in dreamless sleep and anaesthesia than in resting wakefulness, the team measured many brain-injured patients. PCI tracked neurologists' independent diagnoses: vegetative-state patients scored lower than minimally conscious patients, whose behavioural signs come and go. They could draw a dividing line between values indicating presence and absence of consciousness. Locked-in patients had PCI indistinguishable from healthy age-matched controls.
Zipping Without Zapping
At Sussex, Adam Barrett and PhD student Michael Schartner measured the algorithmic complexity of spontaneous EEG, with no TMS. Complexity dropped reliably in early sleep and anaesthesia, and during REM sleep (when vivid dreaming is likeliest) it was much the same as in waking. They checked that the drop was not a trivial reflection of the familiar rise in low-frequency "delta" power during sleep. Massimini's group found the same pattern with PCI, supporting the claim that these measures track level, not wakefulness.
Finding Hidden Consciousness
Wakefulness is defined behaviourally (response to a loud noise or a pinch); consciousness is defined by inner experience, so it can only be inferred indirectly from outside. Clinical scales (such as the Glasgow coma scale) still rely on behaviour: obeying a two-part command and stating name and date indicate full consciousness. Behaviour-based diagnosis misses people with inner lives who cannot express them.
- Locked-in syndrome. Damage to the brainstem (usually the pons) leaves consciousness intact but the body paralysed; often only limited eye movements survive. Jean-Dominique Bauby, editor of Elle, locked-in after a 1995 haemorrhage, wrote The Diving Bell and the Butterfly this way. "Complete" locked-in patients lack even that channel. A surprising proportion of locked-in patients report reasonable quality of life, another warning against inferring the inside from the outside.
- The grey zone. Vegetative and minimally conscious states are harder: behavioural signs are absent or inconsistent and scans may be inconclusive. A high PCI in an apparently unresponsive patient flags someone worth a second look.
Conscious, but not in Italian. A young man with severe head injuries in a Milan hospital was unresponsive to simple questions and commands, suggesting a vegetative state, yet his PCI matched a healthy conscious person's although he was not locked-in. Staff traced an uncle from North Africa; when he spoke to his nephew in Arabic, the patient smiled at jokes and gave a thumbs-up at a film. He had been conscious all along. Massimini suspects a kind of "cultural neglect", as if the Italian world had stopped mattering to him.
Tennis and house (Owen et al., 2006). Adrian Owen's team scanned a 23-year-old woman, behaviourally unresponsive after a traffic accident, with fMRI while asking her to imagine playing tennis or imagine walking through the rooms of her house. In healthy people these tasks activate distinct regions (supplementary motor area for fluent movement; parahippocampal gyrus for spatial layout). Her brain showed the same patterns, so she was following instructions and must have been conscious despite the vegetative diagnosis. In 2010, Martin Monti and colleagues used the method for communication: a patient answered yes/no questions with tennis for "yes" and house for "no". The ethics are fraught (asking about pain is reasonable; asking whether someone wants to live is another matter); Seth co-created the 2016 audio drama "The Sky Is Wider" on these dilemmas.
How many conscious people lie unrecognised? A recent analysis suggests 10 to 20 per cent of vegetative-state patients may retain covert consciousness, which would mean many thousands worldwide, and probably an underestimate, because Owen's test requires understanding language and sustaining imagery. PCI's advantage is that it asks the patient to do nothing at all, as a true consciousness meter should.
| Measure | What it does | Strength | Limitation |
|---|---|---|---|
| Bispectral index | Combines EEG features into one proprietary number | Used routinely in theatres | No theory; conflicts with behaviour |
| PCI (zap and zip) | LZ complexity of the TMS-evoked EEG echo | Separates VS from MCS; no task needed; principled | Needs TMS-EEG kit; captures integration only vaguely |
| Spontaneous LZ complexity | Compresses ongoing EEG/MEG | Easy; tracks sleep, anaesthesia, REM, psychedelics | Measures information (randomness) only |
| Owen imagery fMRI | Detects wilful mental imagery to command | Shows covert awareness; allows yes/no answers | Requires language, sustained effort; false negatives |
| Neural complexity, causal density | Quantify integration and information together | Closest to theory | Have not yet worked well on real brain data |
Clinical consciousness detection has advanced. Casarotto and colleagues (2016) calibrated PCI on a benchmark population and set an empirical cut-off, PCI* = 0.31, that separated conscious from unconscious conditions with near-perfect accuracy; applied to patients, it found high-complexity responses in a minority of vegetative/UWS patients, consistent with covert consciousness. Faster variants (such as PCIst, 2019) have followed. In 2024, Bodien and colleagues (New England Journal of Medicine) tested several hundred patients with task-based fMRI and EEG and found cognitive motor dissociation in about a quarter of those who could not follow commands behaviourally, larger than the older 10 to 20 per cent estimate. European clinical guidance (2020) now recommends combining such brain-based tests with bedside examination when assessing disorders of consciousness.
Is Level One-Dimensional?
Asking whether a baby is less conscious than an adult, or a tortoise less than either, risks human exceptionalism: treating the healthy adult human set of properties as the essence of consciousness and putting it at the top of a single scale. This is one place the temperature analogy may fail. (Attempts to rank intelligence on one scale, IQ, foundered on the same variety.)
- All-or-none versus graded. Seth calls the dichotomy misguided. He prefers sharpish transitions from total absence to some experience, after which experience varies in degree and perhaps along several dimensions, whether in evolution, development, waking from anaesthesia or recovery in the clinic.
- Dreaming versus the post-lunch stupor. Dreaming may be more conscious in perceptual vividness and less conscious in reflective insight. Insight survives in lucid dreaming; a recent study communicated with lucid dreamers through eye movements, and they answered simple maths problems and yes/no questions.
- Consequence. If level is multidimensional, the sharp line between level and content disappears, and a one-size-fits-all meter may never suffice.
Psychedelics: Content Changes Level
On 19 April 1943, now "Bicycle Day", Albert Hofmann, the Sandoz chemist who had synthesised LSD, swallowed a small dose and cycled home in Basel, believing he was going insane, before kaleidoscopic images surged behind his closed eyes. Psychedelic states bring vivid hallucinations and "ego dissolution", the shifting of boundaries between self, world and other people, so pervasive that they might alter level as well as content.
- 1943 Hofmann's first deliberate LSD dose ("Bicycle Day").
- 1950s-60s Promising clinical studies, including for addiction and alcoholism.
- Late 1960s Recreational use and counterculture (Timothy Leary) lead to research being shut down.
- 2000s Substantial research restarts after a lost generation.
- April 2016 Seth and Robin Carhart-Harris plan a joint analysis at a conference near Tucson, Arizona.
Classic psychedelics (LSD, psilocybin, mescaline, DMT in ayahuasca) act mainly on the serotonin system, binding strongly to the 5-HT2A receptor. Carhart-Harris's team at Imperial College London had shown that under psychedelics resting-state networks that normally co-vary decouple and normally independent regions link up: a breakdown of usual connectivity that might explain blurred self-world boundaries and mixing of the senses.
Complexity rises above waking (Schartner et al., 2017). With Carhart-Harris and Suresh Muthukumaraswamy (University of Auckland), Schartner and Barrett computed algorithmic complexity of MEG recordings (millisecond resolution, whole-brain coverage) from volunteers on psilocybin, LSD or low-dose ketamine (anaesthetic at high doses, hallucinogenic at low). All three increased complexity relative to placebo. It was the first time any measure of conscious level had risen above normal waking rest; sleep, anaesthesia and brain injury had only ever produced decreases. A later study found the same with DMT (Timmermann et al., 2019), and a follow-up led by Lionel Barnett found reduced information flow between cortical regions, another sign of lost perceptual structure.
The measures index randomness or "signal diversity", so brain activity becomes more random on a trip, matching the freewheeling reorganisation of experience. Since a level measure responds to wholesale changes in content, conscious level and conscious content are not independent aspects of consciousness. But the result raises an absurd extrapolation: would maximally random brain activity give a maximally psychedelic experience? More likely no experience at all, just as free-form jazz at some point stops being music.
Complexity Is Not Randomness
Algorithmic complexity is a poor stand-in for what "complex" ordinarily means. True complexity sits between order and disorder: Nina Simone and Thelonious Monk, not the Bonzo Dog Doo-Dah Band (whose 1967 album Gorilla parodied jazz by playing as badly as possible).
The key paper is Tononi and Edelman (1998), Science, which drew Seth to San Diego. Instead of an NCC for one experience, they asked what is true of all experiences, and answered: every conscious experience is both informative and integrated.
- Informative. Every experience differs from every other you have had or could have; having this one rules out vast numbers of alternatives, and reduction of uncertainty is, mathematically, information. Informativeness is not richness: listening to Nina Simone while eating strawberries on a rollercoaster rules out as many alternatives as sitting in a silent room with eyes closed. An experience is what it is because of all it is not: redness is redness because it is not blue, green, a smell, a thought or regret.
- Integrated. Every experience is a unified scene: colours are not experienced apart from shapes, nor objects apart from backgrounds; computer, coffee cup, a door closing and thoughts about what to write next are bound into one encompassing scene. (Some philosophers, notably Tim Bayne, question whether unity is guaranteed, citing split-brain cases.)
- The key move. If experience has both properties at the level of phenomenology, the mechanisms underlying it should too, and by expressing them they account for, not merely correlate with, experience.
Maximally informative brain. Every neuron fires independently and randomly, as if disconnected. LZW complexity is very high. Lots of information, no integration: no consciousness.
Maximally ordered brain. All neurons fire in lockstep, roughly as in a global epileptic seizure. LZW complexity is very low. Lots of integration, no information: no consciousness either.
A fit-for-purpose measure must therefore track how information and integration are jointly expressed. LZW captures information but says nothing about integration. PCI does somewhat better: the echo must be hard to compress (information) and must travel widely (suggesting integration), but integration is assumed vaguely rather than measured in the same quantitative way. Measures that do quantify both exist: "neural complexity" (Tononi, Olaf Sporns and Edelman, 1990s) and Seth's own "causal density" a decade later, plus newer successors. Yet none has worked well on real brain data. Seth reads this paradox (theory-faithful measures underperform theory-light ones) as calling for better mathematics and better imaging, while conceding the theory itself might be wrong.
Caveats on the complexity programme. Owen-style evidence has had its own controversies: an EEG version of the imagery test (Cruse et al., 2011) was challenged on statistical grounds by Goldfine and colleagues (2013), a reminder that false positives are possible as well as false negatives. Spontaneous LZ complexity is also sensitive to ordinary spectral changes in the signal (power in particular frequency bands), so careful controls (which Seth's team ran) are needed before reading it as a pure measure of level. And the psychedelic "increase in level" depends on interpreting signal diversity as level; the result is equally consistent with altered content read out by a content-sensitive measure, which is Seth's own conclusion.
Temperature or Life?
Progress is real: level is recognised as distinct from wakefulness; several brain-based measures track global states and detect residual awareness; PCI is both clinically useful and grounded in information and integration, a real-problem bridge from mechanism to universal properties of experience. But Seth's intuition is that consciousness will prove more like life than temperature. Integration and information are general properties of most or all experiences; that does not make consciousness integrated information, as temperature is mean kinetic energy. Testing that analogy to breaking point means meeting integrated information theory.
A lesson for model monitoring. The BIS story is the classic trap of a proprietary composite score without theory: it works until it doesn't, and nobody can say why. PCI succeeded because a perturbation plus a principled statistic was validated against known fixed points (sleep, anaesthesia, waking), exactly the thermometry move. The same discipline applies to AI evaluation and risk scoring in banking: anchor metrics to ground-truth reference states, perturb the system rather than only observing it, and be wary when a theory-faithful metric underperforms a crude proxy, since that is a signal about the theory or the data, not a licence to ship the proxy.
Phi
Integrated information theory (IIT), Giulio Tononi's theory, is the purest form of the temperature analogy: consciousness is integrated information, measured by Φ (phi), the information a system generates as a whole, over and above its parts. Seth admires its ambition but argues that its central identity claim is untestable in practice, because Φ requires knowing what a system could do, not just what it does, and that it carries strange consequences: inactive neurons that change experience, information as a fundamental constituent of reality, and restrained panpsychism. His alternative keeps IIT's insight and drops the identity: build practical, observer-relative measures and treat consciousness as more like life.
Gelato in Las Vegas
In July 2006, in the Venetian hotel in Las Vegas, where artificial early evening never ends so that guests keep spending, a jet-lagged Seth spent hours eating gelato with Giulio Tononi and defending his recent paper criticising an early version of IIT. Tononi patiently explained why he was wrong; Seth flew home less sure of himself, resolving to think harder, prepare better and drink less.
IIT claims that subjective experience is a property of patterns of cause and effect, that information is as real as mass or energy, and that even atoms may be a little bit conscious. More than any other neuroscientific theory, it tackles the hard problem head on. It was obscure in 2006 and is now among the highest-profile and most hotly debated theories in the field. Christof Koch, once the champion of the NCC approach, called it "a gigantic step in the final resolution of the ancient mind-body problem".
Two common objections do not move Seth. The theory is mathematically complex, but nobody promised consciousness would be simple. Its claims are counterintuitive, but intuition is a poor guide with a phenomenon this strange. His real objection is different: extraordinary claims require extraordinary evidence, and IIT's ambition makes that evidence unobtainable.
What Φ Measures
Φ measures how much a system is "more than the sum of its parts" in terms of information, loosely like a flock of birds that seems to have a life of its own. A system is conscious to the extent that its whole generates more information than its parts.
IIT asserts an identity, not a correlation. Φ is intrinsic (independent of any observer) and identical to the amount of consciousness: high Φ, lots of consciousness; zero Φ, none. This is why Seth calls IIT the ultimate expression of the temperature view; Tononi and Koch use the temperature analogy themselves.
For a network of simple on/off "neurons" to have high Φ it must satisfy two conditions mirroring two features of experience:
- Information. The network's global state must rule out a large number of alternative global states, as every experience rules out countless others.
- Integration. There must be more information in the whole than in the parts taken separately, however the system is divided, as every experience is unified, "all of a piece".
| System | Information? | Integration? | Φ | Lesson |
|---|---|---|---|---|
| Single photodiode | One state rules out one alternative: at most 1 bit | Trivial | ~0 | Too little information |
| Camera sensor array | Huge: a different global state for every scene | None: photodiodes are causally independent; chop the array up and it works the same | 0 | Information without integration |
| Network split into two halves | Each half may have Φ > 0 | The whole has zero Φ: the cut between halves loses nothing | 0 for whole | Φ depends on the optimal cut |
| Real split brain | Two hemispheres | Severed corpus callosum | Possibly two consciousnesses, none spanning both | Same reason you and I form no collective consciousness |
| Cerebellum | Many neurons (about three quarters of the brain's) | Organised as many semi-independent circuits, like a sensor array | Low | Explains its minor role |
| Cerebral cortex | Rich | Densely interconnected | Likely high | Explains its central role |
The search for the cut that minimises the whole-versus-parts difference (technically the minimum information partition) is what distinguishes Φ from the complexity measures of chapter 2. IIT explains why consciousness fades in dreamless sleep, anaesthesia and coma though wiring is unchanged: the capacity of cortical neurons to interact is compromised so that Φ vanishes.
An Axiomatic Theory
IIT starts from axioms, statements taken as self-evident, like Euclid's "two shapes that fill exactly the same space are the same shape", rather than from data. Its main axioms are that experience is integrated and informative; the full set adds that consciousness exists, that it is composed of many elements, and that it is exclusive to one spatiotemporal scale. From axioms IIT derives the properties any conscious mechanism must have; any system with those properties, biological or not, has non-zero Φ and is conscious. Tim Bayne has questioned whether the axioms, especially exclusion, are really self-evident.
Why Φ Cannot Be Measured
Testing IIT means measuring Φ in real systems, and for most systems this is nearly or actually impossible. The reason is IIT's unusual treatment of information.
Observer-relative (extrinsic) information. Claude Shannon's standard notion: the reduction in an observer's uncertainty on seeing a system's state. Roll a die: each outcome rules out five alternatives. Measured by entropy, S = -Σ pₖ log pₖ, highest when states are equally likely (a fair die: about 2.6 bits; a loaded die less). You can estimate it by watching the system over time and tallying states.
Intrinsic information (IIT). If Φ is consciousness and consciousness is not observer-relative (whether you are conscious does not depend on how anyone measures your brain), then information must be "for" the system itself. Measuring it requires knowing everything the system could do, even what it never does: its maximum entropy distribution, not its observed (empirical) distribution.
The missing twelve. Throw two dice a few times and get a seven, an eight, an eleven, never a twelve. The empirical distribution has no twelve; the maximum entropy distribution does, because a twelve could have happened. So each observed outcome rules out more alternatives relative to the max-entropy distribution. There are two routes to it: perturb a system in every possible way (as a child presses every button on a new toy), or know its complete cause-effect structure (knowing dice have six faces tells you two dice yield 2 to 12 without a single throw). For a brain, neither is available: we can record what it does but cannot know its full physical structure or perturb it in every way. Hence IIT's most distinctive claim is its least testable. (Recent IIT versions define Φ as "irreducible cause-effect power", explicitly a claim about mechanism, not dynamics.)
Further obstacles apply to any flavour of Φ:
- Partitioning. Finding the right cut is easy for split brains but generally hard: the number of possible partitions grows exponentially with system size, and partitions of different sizes are hard to compare fairly.
- Granularity. Should Φ be computed over neurons and milliseconds or atoms and femtoseconds? Could an entire country be conscious, one more than another? Eric Schwitzgebel notes that if Φ peaked at the scale of a nation, its citizens would cease to be individually conscious. Could tectonic plates integrate information over geological time?
- Seth's concession. These are problems for us as observers. On IIT a system simply has its Φ, as a thrown stone traces its arc without calculating gravity. Hard to test is not the same as wrong.
Weird Consequences
The Densuke neurons. Implant a fistful of new neurons wired so they would fire only if you ate a Densuke watermelon, a rare fruit grown only on Hokkaido. You never eat one, so they never fire, and your brain behaves exactly as before. IIT nonetheless predicts that all your experiences change, subtly, because the brain now has more potential states and so a different Φ. The flip side: neurons in visual cortex that are inactive anyway are made inactivated (prevented from firing). Activity is unchanged, but potential states shrink, so IIT predicts a change in experience.
The second scenario may be testable with optogenetics, which genetically modifies chosen neurons to respond to light so lasers or LEDs can switch them on or off (Deisseroth). It would not measure Φ, but it could test an IIT-specific prediction. The "inactive versus inactivated" experiment was proposed by Umberto Olcese and Tononi, and Seth joined discussions within the Templeton World Charity Foundation's structured adversarial collaborations about running it.
- Information is real. By identifying Φ with consciousness, IIT gives information ontological status like mass/energy or charge, echoing physicist John Wheeler's "it from bit".
- Restrained panpsychism. Wherever the right cause-effect structure exists there is non-zero Φ and some consciousness: not spread through the universe "like a thin layer of jam", but present here and there, not everywhere.
Seth's Verdict
IIT's audacity carries the heavy price that its primary claim, the equivalence between Φ and conscious level, may be impossible to test. Weirdness alone is no objection (modern physics is weird and less wrong than what came before), but physics earned its credibility through experimental testability.
Seth's proposal: keep the insight that experiences are informative and integrated; drop the conceit that Φ is to consciousness as mean kinetic energy is to temperature. With Adam Barrett and Pedro Mediano, he has developed versions of Φ based on observer-relative information and the empirical distribution, computable from a system's observed behaviour. Their versions still behave rather differently even on simple model systems, so work remains. Integration and information become properties of experience to be explained, not axioms about what consciousness is: consciousness as life, not temperature. (He notes one intriguing IIT-inspired real-problem proposal: Haun and Tononi (2019) link the spatial feel of vision to the grid-like anatomy of early visual cortex.)
The wider critique of IIT. Seth's testability objection is one of several. Scott Aaronson (2014) showed that simple regular structures such as large grids of logic gates can have higher Φ than a human brain while doing nothing intelligent; Tononi accepted the implication. The "unfolding argument" (Doerig, Schurger, Hess and Herzog, 2019) notes that any recurrent network can be replaced by a feedforward network with identical input-output behaviour but zero Φ, so either IIT's predictions are untestable by behaviour or the theory is wrong. A small correction: Seth dates Shannon's information theory to the 1950s; the founding paper appeared in 1948.
IIT since 2021. - IIT 4.0 (2023). Albantakis, Tononi and colleagues restated the theory, adding intrinsicality as an explicit axiom alongside existence, information, integration, exclusion and composition, and refining how Φ and the "Φ-structure" (the quality of an experience) are computed. It remains intractable for realistic systems. - The adversarial collaboration (2023, published in Nature 2025). The Cogitate study's preregistered tests found evidence against both IIT and global neuronal workspace theory on key predictions; for IIT, the predicted sustained synchronisation within posterior cortex did not appear. - The "pseudoscience" letter (September 2023). Over 100 researchers signed an open letter calling IIT pseudoscience, citing its untestable core and panpsychist implications and objecting to media coverage of the collaboration as a win for IIT. Others, including some researchers sceptical of IIT, objected to the label as damaging; a revised version appeared in Nature Neuroscience in 2025, with replies from IIT's defenders. The dispute sharpened exactly Seth's distinction between IIT's testable predictions and its untestable identity claim.
IIT and AI. IIT has a sharp implication for machine consciousness that functionalists dislike: Φ depends on physical causal architecture, not software. Tononi and Koch argue that a conventional digital computer, even one simulating a human brain perfectly, would have negligible Φ and no experience. On IIT, then, no large language model running on today's hardware is conscious, however human its outputs; on functionalist theories it might be. When colleagues or vendors make claims about "sentient" AI, the first question is which theory of consciousness the claim presupposes.
The chapter closes back in the Venetian, under its fake sky. Linking all the levels surveyed (anaesthesia, coma, the vegetative and minimally conscious states, sleep, wakefulness, psychedelia) is the idea that every experience inhabits the complex middle ground between order and disorder. In Las Vegas it is hard to believe anything is real besides the raw feel of experience, and tempting to think everything is a kind of hallucination, a thought that turns out to contain unexpected truth.
Perceiving from the Inside Out
Opening Part II (Content), Seth inverts the commonsense picture of perception. The brain, sealed in a dark skull, receives only unlabelled, ambiguous electrical signals; what we experience is its best guess about their causes. Perception flows mainly from the top down, or the inside out: predictions are checked against prediction errors, and conscious content is the content of the predictions, not of the sensory signals. Perception is a controlled hallucination, and hallucination is uncontrolled perception. Colour, The Dress, Adelson's checkerboard and two-tone images make the case.
Part II: Content
Part I asked how conscious we are; Part II asks what we are conscious of: the sights, sounds, smells and feelings of our inner universe. Its organising idea is that the brain is a prediction machine, and that everything we perceive is a kind of waking dream reined in by the world, both more than and less than whatever the real world is.
Imagine Being a Brain
On the deck of a ramshackle house in a cypress forest north of Santa Cruz, fog wreathing the trees, Seth tries to persuade himself that this whole world is a construction of his brain. Change perspective and imagine being that brain: no light, no sound, nothing but a barrage of electrical signals only indirectly related to the world. The signals carry no labels ("I'm from a coffee cup"), not even labels of modality (vision, hearing, touch, or less familiar senses such as thermoception for temperature and proprioception for body position). The idea that humans have only five senses, Seth notes, is a completely wrong legacy of Aristotle's De Anima (c. 350 BC).
The "How Things Seem" View
Common sense says: a mind-independent world of objects with real colours, shapes and textures; senses as transparent windows; the brain reading out sensory data; a self "behind the eyes" receiving the results and deciding what to do. Sense, think, act. Many neuroscientists still frame perception as bottom-up feature detection.
Bottom-up feature detection. Signals flow inward through stages extracting ever more complex features: luminance and edges, then object parts (eyes, ears, wheels, wing mirrors), then whole objects and categories (faces, cars, cups), like fishermen catching bigger fish further downriver. Top-down signals only refine or constrain. Evidence: modality-specific cortical regions; hierarchical organisation from primary visual cortex to inferotemporal cortex; neurons in cats and monkeys tuned to edges at early stages, faces at late stages; matching fMRI results in humans; David Marr's 1982 computational theory of vision; and deep learning vision networks built bottom-up and performing near human level.
Top-down inference. Perception comes primarily from the inside out. The brain constantly issues predictions about the causes of its sensory signals; incoming signals serve as prediction errors that keep those predictions tied to the world. What we experience is built from the predictions. The anatomy is equally consistent with this view (perceptual areas are richly endowed with top-down connections, which the bottom-up view struggles to explain), and it changes everything about what perception is.
What would it have looked like if it had looked as if the Earth turned on its axis?
Wittgenstein's point: understanding that the Earth rotates does not change how sunrise looks. Likewise, if perception is a top-down best guess, the coffee cup will look exactly as it always has. The inversion is Copernican: consistent with much existing evidence, leaving appearances unchanged, while changing everything.
A Long Lineage
- c. 375 BC Plato's Allegory of the Cave: prisoners take shadows of hidden objects for reality; our perceptions are indirect reflections of causes we never encounter directly.
- c. 1030 Ibn al-Haytham (Optics): perception depends on "judgement and inference", not direct access.
- 1781 Immanuel Kant: raw sensory chaos is meaningless without prior structuring concepts, including a priori space and time; the noumenon, the thing-in-itself (Ding an sich), lies forever behind a sensory veil.
- Late 1800s Hermann von Helmholtz: perception as "unconscious inference", combining signals with expectations about causes; unconscious because we are aware only of results, not mechanisms. He saw this as a scientific version of Kant.
- 1947 to 1950s The "New Look" in psychology: social and cultural factors shape perception; Bruner and Goodman reported that poorer children overestimated coin sizes. Many such studies were methodologically weak.
- 1970s Richard Gregory: perception as neural hypothesis-testing, like a scientist testing hypotheses against data.
- 1999 onward Predictive coding and predictive processing (Rao and Ballard; later Karl Friston, Andy Clark, Jakob Hohwy): perception as brain-based inference.
Controlled Hallucination
Seth's version takes its name from a phrase he first heard from psychologist Chris Frith (traced back to a 1990s seminar by Ramesh Jain).
- 1. Predictions cascade down. The brain continually predicts the causes of its sensory signals; predictions flow top-down through perceptual hierarchies. Looking at a cup, visual cortex predicts the signals the cup should produce.
- 2. Errors flow up. Sensory signals stream bottom-up as prediction errors, the difference between expected and received at every level. By adjusting predictions to suppress prediction errors, best guesses keep their grip on worldly causes: perception is prediction error minimisation.
- 3. Experience is the prediction. The decisive claim: what we experience is determined by the content of the top-down predictions, not by the bottom-up signals. We never experience sensory signals themselves, only interpretations of them.
What we perceive is a top-down, inside-out neuronal fantasy that is reined in by reality, not a transparent window onto whatever that reality may be.
Hallucination is usually contrasted with "normal" perception of things that are really there. On this view the difference is one of degree: both rest on internally generated predictions and share core mechanisms. In normal perception, predictions are controlled by worldly causes; in hallucination they have partly lost their grip, because predictions are not properly updated by prediction errors. If perception is controlled hallucination, hallucination is uncontrolled perception, and asking where to draw the line is like asking where day ends and night begins.
Colour: Where Brain and Universe Meet
Human vision responds to a thin slice of the electromagnetic spectrum between infra-red and ultra-violet, which alone shows perception cannot be a comprehensive representation of the world: it is both less and more than that. A neurophysiologist might say you see a colour when retinal cone cells are activated in certain proportions. True but incomplete: there is no one-to-one mapping between cone activity and colour experience.
- Discounting the illuminant. White paper looks white indoors (yellowish light) and outdoors (bluish sunlight) although the reflected light differs. The brain infers an invariant property, how the surface reflects light, and whiteness is how that inference appears in experience.
- Colour as a tracking device. Colour is not a property of things-in-themselves but a device evolution hit on to recognise and track objects under changing light. A chair is not red any more than it is ugly or avant-garde; its surface has a way-in-which-it-reflects-light, and redness is the phenomenological aspect of tracking it.
- No red in the head. Nor is redness moved into the brain as some pigment ("figment", Dennett's word) viewed by an inner camera feeding another visual system with its own camera, endlessly. That is Daniel Dennett's "double transduction" fallacy; it explains nothing. The mechanisms of perception are not red. (Likewise, hearing music needs no little band and microphones inside the skull.)
Colour is the place where our brain and the universe meet.
The claim generalises to all perception: the multisensory panorama is "a writing as much as a reading", a reaching out from brain to world. Seth borrows a line from rapper Baba Brinkman: we are all hallucinating all the time; when we agree about our hallucinations, we call that reality.
Three Demonstrations
The Dress (February 2015). A photo of a dress looked blue and black to some and white and gold to others, each side incredulous. Seth, co-author of a recent book on visual illusions (Eye Benders), first suspected a hoax: it looked plainly blue-black to him and to four colleagues, until the fifth said white-gold; as worldwide, his lab split about half and half. Within an hour he was explaining it on the BBC. The consensus: people discount the illuminant differently. The photo is overexposed and context-poor (the dress fills the frame), so the brain must guess the lighting. Assume yellowish indoor light and you infer blue-black; a visual system bathed in bluish daylight might infer white-gold. A cottage industry of tests followed (dim-room-then-daylight, national sunshine rates, age effects). The lesson: confident, opposite experiences of one image show perception is an internal construction shaped by personal biology and history; normally our constructions coincide closely enough to hide this.
Auditory equivalents exist: a clip heard as "Yanny" or "Laurel" (2018), and a 2020 TikTok toy sound heard as "green needle" or "brainstorm" depending on which word you read.
What later work found on The Dress. The physical garment is royal blue and black. Large online studies (Lafer-Sousa, Hermann and Conway, 2015; Pascal Wallisch, 2017, with over 10,000 respondents) confirmed that the split tracks implicit assumptions about the illumination: people who assumed the dress was in shadow (bluish light) tended to see white and gold, and Wallisch found that self-described early risers, with more daylight exposure, were more likely to see white-gold, much as Seth guessed. The case is now a standard demonstration of individual priors in perception.
Adelson's checkerboard shows the same principle with no individual differences: everyone sees A as darker than B, though they are the same grey. The brain's circuitry embodies the knowledge that objects in shadow look darker, so it infers B is lighter. Knowing they match does not help; Seth has stared at it thousands of times. When knowledge fails to alter perception, perception is "cognitively impenetrable". This is not a malfunction: a visual system is not a light meter. Its job is to infer the most likely causes of sensory signals, not to report the signals.
Two-tone (Mooney) images reveal how fast new predictions change experience. A photo is rendered in greyscale and thresholded into pure black and white until it is unrecognisable; after seeing the original, the two-tone image snaps into a scene and is hard to unsee. The sensory signals are identical both times; only the brain's predictions changed. Biyu He's group has tracked how neural dynamics differ when a Mooney image is recognised. Hearing has an equivalent in sine-wave speech (examples by Chris Darwin at Sussex): speech stripped of the frequencies that make it intelligible sounds like whistling until you hear the original, after which the degraded version becomes clear.
Implications and Two Objections
Perception is a generative, creative act, a context-laden engagement with sensory signals, across all senses all the time. So we never experience the world "as it is"; as Kant saw, it is unclear what that would even mean. Judging perception by accuracy alone is a mistake: "veridical" perception is a chimera. Evolution designed the controlled hallucination to enhance survival, not to be a window that makes no conceptual sense anyway. (Peripheral vision is an example: low photoreceptor density, yet it does not look blurry; sharpness and blur are relative to what each part of the system is tuned to, not an "illusion".)
Objection 1: "Then jump in front of a train." Seth denies nothing about trains, cats or cups existing. The "control" matters as much as the "hallucination": not anything goes; how things appear is constructed. John Locke distinguished primary qualities (exist independent of observers: extension, solidity, motion; why the train will hurt you whatever you believe) from secondary qualities (exist only in interaction with an observer: colour). On Seth's view both kinds are perceived through active construction, and in neither case is the experience identical to the quality.
Objection 2: "How can we see anything new?" If perception needs predictions, are we trapped among the expected? Meeting your first gorilla, you still see a gorilla, because the experience is never wholly new: it is assembled from predictions at many levels and timescales (edges, fur, limbs, faces, posture, how objects of that size move and reflect light) learned by you and, over evolution, your ancestors. The long answer requires the machinery of the next chapter.
| Phenomenon | Sensory input | What changes | What it shows |
|---|---|---|---|
| White paper indoors/outdoors | Different spectra | Nothing: stays white | Brain discounts the illuminant |
| The Dress | Same photo for all | Colour seen differs between people | Individual priors about lighting |
| Adelson's checkerboard | A and B identical | Look different for everyone | Built-in shadow prior; cognitively impenetrable |
| Mooney image | Identical before and after | Blobs become a scene | New predictions reshape experience instantly |
| Sine-wave speech | Identical before and after | Whistles become words | Same principle in hearing |
Points to weigh. Seth calls Wittgenstein "the legendary German thinker"; Wittgenstein was Austrian (born in Vienna, later a British citizen). More substantively, the third ingredient (experience is the prediction, not the error) is the most contested part of the view. Some predictive-processing theorists hold that surprising (error-laden) signals are what reach awareness, and laboratory evidence is mixed: expected stimuli often produce weaker but sharper neural responses, which different models interpret differently. Philosophers also dispute Seth's anti-realism about colour (the "chairs are not red" claim): colour realists (Byrne and Hilbert, whom he cites) hold that redness is a real reflectance property of surfaces. Finally, the "New Look" studies he cites were weak, and a 2016 critique (Firestone and Scholl) argued that most claimed top-down effects of cognition on perception reflect methodological pitfalls; Seth's illusions concern perceptual priors, which that critique does not deny.
Bottom-up machines, top-down brains. In 2021 Seth contrasted brains with bottom-up deep-learning vision. Since then mainstream AI has moved toward generative, predictive models: self-supervised vision systems learn by predicting masked parts of images, and large language models are trained purely by predicting the next token. These are predictive in a statistical sense but lack Seth's other ingredient: control by a world the system acts in and depends on. For enterprise AI, the analogy is useful: an LLM's fluent confabulation is prediction without adequate error correction, uncontrolled perception in Seth's terms, and grounding (retrieval, tool results, validation checks) supplies the "control" that keeps the hallucination tied to its causes.
The Wizard of Odds
Bayesian inference is the mathematics of best guessing under uncertainty: combine priors with likelihoods to get posteriors, weighting each source by its precision (reliability). Seth argues that the brain approximates Bayes by prediction error minimisation, built from three components: generative models, perceptual hierarchies and precision weighting, with attention as the turning up of precision. Action is the other way to cancel prediction error: in active inference, movements are self-fulfilling proprioceptive predictions. The conclusion: top-down predictions do not merely bias perception; they are what we perceive.
Reverend Bayes and Three Kinds of Reasoning
The Reverend Thomas Bayes (1702 to 1761), a Presbyterian minister, philosopher and statistician of Tunbridge Wells, never published his theorem. Richard Price presented his "Essay towards Solving a Problem in the Doctrine of Chances" to the Royal Society two years after his death, and Pierre-Simon Laplace did much of the later mathematical heavy lifting. Bayes' name is now tied to "inference to the best explanation". (The chapter title comes from rapper Baba Brinkman's Rap Guide to Consciousness, which Seth advised.)
| Reasoning | Direction | Example | Can it be wrong? |
|---|---|---|---|
| Deduction | Logic alone | Jim older than Jane, Jane older than Joe, so Jim older than Joe | No, if premises true and rules followed |
| Induction | Extrapolate from observations | The sun has always risen in the east, so it always will; three green balls, so all are green | Yes |
| Abduction (Bayesian) | Backward from effects to most likely causes | The lawn is wet: rain or sprinkler? | Yes; it seeks the best explanation of incomplete, ambiguous data |
The Wet Lawn
Looking out one morning, you see your lawn is wet. Did it rain, or did you leave the sprinkler on? Bayes' rule goes from what you already know (the prior) to what you should now believe (the posterior) using what you are learning (the likelihood). These are Bayesian "beliefs": states of knowledge, not of the world, and not necessarily personal beliefs; your visual cortex "believes" there is a coffee cup in the same sense.
- Prior, p(H). Probability of a hypothesis before new data. Rain overnight is very unlikely (you live in Las Vegas); a forgotten sprinkler is unlikely but less so, depending on your habits.
- Likelihood, p(D|H). Reasoning forward from cause to data: if it rained (or the sprinkler ran), how probable is a wet lawn? Assume both make it equally likely.
- Posterior, p(H|D). p(H|D) = p(D|H) × p(H) / p(D). The "prior on the data", p(D), is the same for each hypothesis and cancels when comparing them.
- Decision. A Good Bayesian picks the hypothesis with the highest posterior: here, the sprinkler.
Then new data: your neighbour's lawn is wet too. Under the sprinkler hypothesis only your lawn should be wet; under rain, both. The likelihoods shift, rain now wins, and you change your mind. Two further features matter:
- Reliability. If your window is dirty and you have lost your glasses, the glimpse of the neighbour's lawn is unreliable and you know it, so rain gains only a little and the sprinkler may stay ahead. Reliable information should move beliefs more than unreliable information.
- Iteration. Each posterior becomes the next prior. If the lawn is wet two mornings running, day one's best guess informs day two's.
When intuition fails: base rates. Even a 99 per cent accurate test may only slightly raise the probability of a rare disease, because people overestimate the prior. Seth's example: in 2009 the US government advised against routine mammography for women in their forties. Mammograms then had about 80 per cent sensitivity and flagged about 10 per cent of cancer-free women. Using the low base rate as the prior, the probability of cancer given a positive result is only about 3 per cent: of every 100 women testing positive, 97 are cancer-free, facing anxiety and invasive follow-up.
The base rate in that example is off by a factor of ten. Seth gives the base rate as 0.04 per cent. With 80 per cent sensitivity and a 10 per cent false-positive rate, that prior yields a posterior of about 0.3 per cent, not 3 per cent (0.8 × 0.0004 ÷ (0.8 × 0.0004 + 0.1 × 0.9996) ≈ 0.003). The quoted 3 per cent follows from a base rate of about 0.4 per cent (0.8 × 0.004 ÷ (0.0032 + 0.0996) ≈ 0.031). The lesson is unaffected, and indeed stronger at the lower rate. Policy has since moved too: better mammography led the US Preventive Services Task Force in 2024 to recommend screening every two years from age 40.
Bayesian inference has served everything from medical diagnosis to locating missing nuclear submarines (the history is in Sharon McGrayne's The Theory That Would Not Die). Even science can be seen as Bayesian: hypotheses updated by evidence.
Kuhn and Popper. Thomas Kuhn: whole edifices overturned in paradigm shifts as anomalies pile up. Karl Popper: hypotheses raised and tested one by one, "like balloons released into the sky and then shot down" (falsificationism).
Lakatos and Bayes. The Bayesian view is closest to Imre Lakatos, who analysed how research programmes actually work. It also implies scientists' priors shape how they read data and which experiments they run. Seth admits a strong prior that the brain is Bayesian and wonders how much evidence it would take to overturn it (a point Paul Fletcher and Chris Frith made about their own theory of psychosis).
The Bayesian Brain
The skull-bound brain inferring hidden causes from noisy signals is an ideal Bayesian problem.
- Priors at every scale. From general, fixed priors such as "light comes from above" to situation-specific ones such as "the approaching furry object is a gorilla".
- Likelihoods encode mappings from potential causes to sensory signals: the forward-reasoning component.
- A rolling process. Every fraction of a second a new posterior (a perceptual best guess) forms and becomes the prior for the next moment: perception is a rolling process, not a static snapshot.
- Reliability again. Glimpsing something dark and furry far away, outside a zoo, the prior for gorilla is low and the data unreliable, so the guess moves through large black dog, man in a gorilla suit, actual gorilla as it approaches, hopefully leaving time to run.
Distributions, Means and Precision
Beliefs are better represented as probability distributions than single numbers. The probability that a variable X falls in a range is the area under the curve there, and the total area is exactly 1 (something must happen). A Gaussian (normal, bell-curve) distribution is fully specified by two parameters: its mean (where it peaks) and its precision (how tightly it is concentrated; precision is the inverse of variance). Interpreted as a belief, the mean is its content and the precision its confidence.
For Gaussians, applying Bayes means multiplying prior and likelihood curves and renormalising so the area is 1. The posterior's peak sits nearer the prior because the likelihood has low precision. On the next look the gorilla is closer and clearer, the previous posterior is the new prior, and the best guess shifts towards gorilla, until it is time to run.
The figure misdraws the posterior's width. When two Gaussians are combined by Bayes' rule, precisions add: the posterior is always more precise (narrower and taller) than either the prior or the likelihood. In the book's figure the posterior is drawn wider and lower than the prior, which cannot happen for Gaussian updating. Its position (between the two, nearer the more precise prior) is correct. Formula: posterior precision π_post = π_prior + π_lik, and posterior mean μ_post = (π_prior·μ_prior + π_lik·μ_lik) / π_post, a precision-weighted average.
Bayes supplies a standard of optimality, what a brain should do. It does not say how neurons do it.
Prediction Error Minimisation
The brain generates predictions, compares them with incoming signals, and passes the differences (prediction errors) upward; predictions flow down. Perception is the content of all the top-down predictions once errors have been "explained away" as far as possible. By minimising prediction error everywhere, all the time, the brain approximates Bayes' rule, which is what licenses calling perception top-down controlled hallucination rather than bottom-up read-out. (In Bayesian terms a generative model is the joint probability of hypothesis and data, p(H, D), prior and likelihood together.)
Predictive processing. A theory of mechanism: how brains accomplish perception, cognition and action through prediction error minimisation. Seth uses the term to cover a family that includes predictive coding.
Controlled hallucination. A theory of phenomenology: it takes that mechanism and uses it to explain why conscious perceptions have the properties they do. Both rest on prediction error minimisation.
A terminological twist: people often call bottom-up pathways "feedforward" and top-down ones "feedback". In predictive processing this is backwards, since in engineering the feedback signal is the error signal; here bottom-up connections carry the errors, so they are the feedback. Some stable, global predictions may even be embedded in bottom-up signals (Teufel and Fletcher).
- Generative models. Determine the repertoire of perceivable things. To perceive a gorilla, the brain needs a model able to generate the sensory signals a gorilla would cause; its predictions are compared against data and revised.
- Perceptual hierarchies. Predictions span many scales of space and time: "gorilla" cascades into limbs, eyes, fur, then colours, textures, edges, then brightness variations. Hierarchies are cross-modal and contextual: hearing your mother's voice primes visual cortex to see her; knowing you are at the zoo primes gorillas. (Perceptual cortex is indeed rich in top-down connections.)
- Precision weighting. The brain must infer not only causes but how reliable its signals are, and continually adjusts their estimated precision: down-weighted signals barely update guesses; up-weighted signals dominate. This is done by altering priors on precision (hyperpriors).
"Prediction" need not mean the future. In statistics it means going beyond the data using a model, catering for insufficient data, whether that is because the data concern the future or an unseen present.
Attention and the Invisible Gorilla
Increasing the estimated precision of sensory signals is nothing other than paying attention: turning up their gain. It follows that attending to some signals reduces the influence of others, which explains why we can look straight at something and not see it.
Inattentional blindness (Daniel Simons, 1999). Viewers watch two teams of three, one in white and one in black, each passing a basketball while weaving about, and must count passes between the white players. Many completely miss a person in a black gorilla suit who walks in, makes gorilla moves and walks out. Shown the clip again and told to look for a gorilla, they see it at once and often refuse to believe it is the same video. Attending to white lowers the precision on black-clad signals, gorilla included, so they barely update the best guess.
"Most people" is an overstatement. In Simons and Chabris's original 1999 study about half (46 per cent overall) of observers missed the gorilla, with rates varying by task difficulty and condition. The effect is robust and widely replicated, but it is a coin flip rather than a majority.
Seth once used the argument in court. Turning left onto a short road to the ocean near Del Mar, San Diego, on the way to surf, he was ticketed for ignoring a newly installed "no left turn" sign. He had made the turn hundreds of times, cars ahead had just done the same, and he argued in a written deposition that thanks to precision weighting the sign was not visible to him, though visible "in principle". He took it to California traffic court, landing on the day's "criminal calendar", with a PowerPoint for the judge, which did not help at all.
- Magicians exploit the same mechanism through misdirection, so the queen of spades slipping behind an ear goes unseen.
- Pickpockets too: Seth watched Apollo Robbins relieve perception experts of watches, wallets and purses while they knew exactly what he was attempting.
Action: The Other Half of Inference
"Sense, think, act" is again how things seem, not how they are. Perception and action define each other: every action changes incoming data, and every perception exists to guide action. We perceive the world not as it is but as it is useful for us to perceive it. Action may even come first: brains as fundamentally action-generating systems, calibrated by sensory signals (György Buzsáki's The Brain from Inside Out). The juvenile sea squirt uses a rudimentary brain to find a rock, attaches for life and digests its own brain, an analogy some apply to academics after tenure.
Active inference, Karl Friston's term, minimises prediction error by acting so that sensory data come to match predictions: a self-fulfilling perceptual prediction. Hunting for car keys on a cluttered desk, visual predictions update (empty mug, empty mug, paperclips) while the eyes interrogate the scene until the prediction "keys" is fulfilled. Any bodily action, even high-level ones such as applying for a job or getting married, cascades into movements that change sensory input, so every action participates in perception.
- Fulfils predictions. Actions seek out the data that make predictions come true, guided by conditional (counterfactual) predictions: "if I look over there, what will I sense?" Without them the brain could not choose, among countless actions, the one likely to reduce error (scan the desk, not stare out of the window).
- Improves predictions. Actions harvest data to decide between hypotheses: peeking at the neighbour's lawn, tidying away the mugs to find the keys.
- Drives learning. Over the long run, actions reveal the causal structure of the world, improving generative models: a virtuous circle of better models, better inference and more informative actions. (Seth's group, with Alexander Tschantz, Christopher Buckley and Beren Millidge, builds machine learning on this principle, learning models from little data; fast bottom-up "amortised" inference is a related trick.)
- Is perception. Most counterintuitively, actions are self-fulfilling proprioceptive predictions.
Proprioception tracks the body's position and movement from receptors throughout skeleton and muscles; you can touch your nose with your eyes shut. In active inference, touching your nose means letting predictions of hand position override the evidence that the finger is not there yet. For that, proprioceptive prediction errors must be down-weighted: the opposite of attention, a "disattention" to the body that lets it move. Experiments confirm reduced proprioceptive sensitivity during action (Palmer et al., 2016), and the same sensory attenuation explains why you cannot tickle yourself.
Our perceptual world, alive with colours, shapes and sounds, is nothing more and nothing less than the brain's best guess of the hidden causes of its colourless, shapeless and soundless sensory inputs.
The brain is not isolated: it swims in sensory signals from world and body, issues actions that sculpt that flow, meets it with cascading predictions and sends errors upward, settling and resettling in a "Good-Enough Bayesianism". Action and perception differ only in who leads and who follows in the dance of predictions and errors.
Active inference has grown into a field. Parr, Pezzulo and Friston's textbook Active Inference (MIT Press, 2022) formalised the framework, which is now applied in robotics and reinforcement-learning research. Critics continue to argue that the underlying free energy principle is so general that it is closer to a mathematical framework than a testable theory, and that the evidence for specific neural implementations of prediction error units remains contested. The core behavioural findings Seth relies on, sensory attenuation of self-generated input and attention as precision, have held up well.
Base rates and precision in financial-crime screening. The mammogram arithmetic is the daily reality of sanctions, PEP and fraud screening. If 1 in 10,000 screened customers is a true match and a screen is 99 per cent sensitive and 99 per cent specific, a hit is a real match only about 1 per cent of the time: roughly 99 of every 100 alerts are false. Bayes says the fix is better priors (risk segmentation), better likelihoods (richer matching features), and explicit precision weighting of sources (a verified date of birth should move the score far more than a fuzzy name match), with each posterior becoming the next prior as evidence accumulates. Active inference has an agentic-AI analogue too: an investigator agent should choose the next data request expected to discriminate best between "match" and "not a match", like peeking at the neighbour's lawn.
The Beholder's Share
The art historians' "beholder's share", the part of a perception the viewer contributes, is Seth's bridge from predictive theory to the phenomenology of perception in the wild. Lab experiments show expectations speed and sharpen conscious seeing; a VR "hallucination machine" recreates hallucinations by overweighting perceptual predictions. He then extends controlled hallucination to the deep structure of experience: objecthood, change, time, the felt reality of the world and even causality are all perceptual inferences. Because perceptions present themselves as really out there, the hard problem looks harder than it is; seeing them as constructions starts to dissolve it.
Vienna and the Innocent Eye
Around 1900, Vienna mixed art and science as equals. The cast Seth assembles:
- Gustav Klimt, Oskar Kokoschka, Egon Schiele: pioneers of modernist painting.
- Alois Riegl: art historian who introduced the beholder's share.
- Sigmund Freud: theorist of the unconscious.
- The Vienna Circle: philosophers including Kurt Gödel, Rudolf Carnap and occasionally Ludwig Wittgenstein.
- Ernst Gombrich (born Vienna, 1909): twentieth-century art historian who popularised the beholder's share in Art and Illusion.
- Eric Kandel: neuroscientist who called the period "the age of insight" in his 2012 book of that name.
The beholder's share is the part of perceptual experience contributed by the perceiver and not found in the artwork, or the world, itself. Kandel connects it directly to predictive perception: Gombrich concluded there is no "innocent eye"; all seeing classifies and interprets, and "one cannot perceive that which one cannot classify".
Pissarro's Hoarfrost at Ennery (1873). In the Musée d'Orsay, this painting of a frosted field is, close up, what critic Louis Leroy mocked as "palette-scrapings on a dirty canvas" (in the 25 April 1874 review that coined "Impressionism"). Yet it powerfully evokes a sharply frosted field. Impressionists tried to recover the innocent eye by putting on canvas the raw variations in brightness that feed perceptual inference, not its finished output. Each work is reverse engineering of the visual system, an experiment in predictive perception. Gombrich's phrasing: when brushstrokes "suddenly come to life", we have been led to project a scene into the dabs of pigment.
What Gombrich adds to priors and prediction errors is an appreciation of how perception feels: experiencing the world as really "out there" is a projection, a reaching out from brain to world, not passive revealing. (A cautionary footnote: El Greco's elongated figures were once blamed on astigmatism, but he would have seen his canvas elongated too, cancelling the effect; this "El Greco fallacy" still trips perception researchers. John Berger's Ways of Seeing (1972) later stressed the political and cultural shaping of what we see.)
Expectations in the Lab
Expecting faces and houses (Pinto et al., 2015). Yair Pinto, then Seth's postdoc (now at the University of Amsterdam), used continuous flash suppression: one eye sees a rapidly changing pattern of oblongs that dominates awareness while the other eye's picture is suppressed. Here the mask's contrast ramped down and the picture's (a face or a house) ramped up, so within a few seconds the picture broke into consciousness. A cue word ("face" or "house") preceded each trial and was valid on 70 per cent of trials. People saw expected images faster and more accurately, by about one tenth of a second: small but reliable. Many controls ruled out response and attention biases.
Sharper letters inside words (Heilbron, de Lange et al., Donders Institute, Nijmegen). The word superiority effect: a letter like U is easier to identify in "HOUSE" than in a nonword like "AEUVR". Using noisy displays and fMRI "brain reading" (machine-learning classifiers trained on activity patterns), the team found the neural signature of a letter in early visual cortex was sharper (more distinguishable from other letters) in words than in nonwords. Context-based expectations alter early visual processing to enhance perception, as controlled hallucination predicts.
Hallucination, Pareidolia and the Hallucination Machine
On one summer day, for the first time, Seth took a tiny amount of LSD and lay in the grass. After half an hour the hills, sky and sea began to pulse and come alive; his notes tailed off quickly. Clouds took on definite but shifting forms that he could partly steer: a cloud starting to look like a horse could be pushed further, until a procession of Cilla Blacks (the 1960s Liverpool pop star and TV host) marched across the horizon. For days afterwards he felt he could see "through" his perceptions as constructions, aware of the medium as well as the message.
Pareidolia (Greek "alongside" + "image") is seeing patterns in things, especially faces in clouds, toast or sinks, because faces matter so much that brains come preloaded with strong face priors. Because everyone does it, we don't call it hallucination; a voice that only one person hears commanding self-harm, or an LSD parade of Cilla Blacks, we do.
What we call hallucination is what happens when perceptual priors are unusually strong, overwhelming the sensory data so that the brain's grip on the causes of its sensations starts to slide.
The hallucination machine (Suzuki et al., 2017). In the VR/AR lab (a blu-tacked sign in the old chemistry building), Keisuke Suzuki led an attempt to simulate overactive priors controllably. The team filmed 360-degree video of the Sussex campus square at Tuesday lunchtime, during the pop-up food market, and processed it with a version of Google's "Deep Dream". The network was a deep convolutional neural network, structured roughly like the bottom-up visual pathway and trained (by backpropagation) to recognise over a thousand object categories, many of them dog breeds (it can tell huskies apart). Normally an image goes in and the network says what it contains. Deep Dream runs it backwards: fix the output ("dog") and adjust the image until the network settles, projecting a prediction onto the image, an excess of the beholder's share. Applied frame by frame and replayed in a head-mounted display, it was more compelling than expected: dogs and dog parts emerged organically from every part of the scene, unlike pasted-on pictures.
| Layer clamped in the network | Resulting simulated hallucination | Real-world parallel |
|---|---|---|
| Output layer (whole categories) | Whole objects (dogs) emerge everywhere | Complex hallucinations of people, animals |
| Middle layers (object parts) | A morass of eyes, ears, legs suffusing the scene | Fragmentary, part-based imagery |
| Low layers (edges, textures) | "Geometric" hallucinations: lines, patterns, textures grow vivid | Classic psychedelic geometry |
This is "computational phenomenology": using computational models to bridge mechanisms and properties of experience, here explaining why particular hallucinations have the form they do. The deeper payoff is for normal perception: the machine makes vivid that hallucination is uncontrolled perception, and perception is controlled hallucination. Newer versions by Suzuki, David Schwartzman and Seth build in explicit generative models and capture a wider variety of hallucinations.
The Deep Structure of Perception: Objecthood
Controlled hallucination explains more than "I see a table because that is my best guess". It also explains the deep structure of perception: how contents appear in time, space and across senses. Take objecthood. You perceive the back of a coffee cup you cannot see; a real cup seems to occupy volume in a way a photographed cup does not. Objecthood is pervasive but not universal:
- A clear blue sky does not appear as an object out there.
- A retinal afterimage after glancing at the sun feels like a glitch, not a thing.
- Tinnitus is not experienced as a real sound source, hence "ringing in the ears".
Artists have probed objecthood: René Magritte's The Treachery of Images (1929), the pipe captioned "this is not a pipe", plays on object versus image of an object; Picasso's Cubism shows objects from several perspectives at once, drawing the viewer into constructing objects from a jumble of possibilities. Maurice Merleau-Ponty: the painter investigates how an object makes itself visible.
- Sensorimotor contingency theory (O'Regan and Noë, 2001). Perceptual content comes from "practical mastery" of how actions change sensory input; vision is something an organism does, not a feed to a central mind. It builds on James Gibson's "affordances" (we perceive possibilities for action) and Edmund Husserl's "horizons" of possible perceptions.
- Seth's predictive version (2014). Generative models make conditional or counterfactual predictions: what would happen to the signals if I rotated the tomato. The richer these predictions, the stronger the phenomenology of objecthood. A tomato (red skin all round) supports many; a blue sky or afterimage supports few or none.
- Grapheme-colour synaesthesia. The letter A may evoke a luminous red, consistently and automatically, yet synaesthetes do not confuse it with real colour. Seth's explanation: synaesthetic colour does not change as you move or as lighting changes, so it supports no rich sensorimotor predictions and has no objecthood.
Objects that misbehave (Suzuki et al., 2019). Participants in VR saw deliberately unfamiliar virtual objects (blobs with protrusions), initially hidden by flash suppression. Using a joystick to rotate them, they found either normal object behaviour or random, unpredictable rotation. Normally behaving objects broke into consciousness sooner. Seth concedes the proxy is imperfect (speed of access stands in for objecthood), but valid sensorimotor predictions measurably affect conscious perception.
Perceiving Change
Physical change is neither necessary nor sufficient for perceived change.
- Not necessary. In the "rotating snakes" illusion (Kitaoka), static patterns seen in the periphery persuade visual cortex to infer motion. The waterfall illusion is another: after staring at a waterfall, a rock face seems to move upward while staying in place.
- Not sufficient: change blindness. In one video, the lower half of an image changes from red to purple over about forty seconds, and most viewers do not notice, even looking straight at it, unless primed to look for it. Unlike the invisible gorilla, what goes unseen is change itself.
The supposed puzzle (after the change, are you experiencing red or purple, and if purple, what were you experiencing before?) dissolves once you deny its premise: perception of change is not the same as change of perception. The experience of change is itself a perceptual inference.
Change blindness in the wild. The classic demonstration beyond the lab is Simons and Levin's 1998 "door study": an experimenter asking a pedestrian for directions was swapped for a different person while workers carried a door between them, and about half the pedestrians failed to notice they were now talking to someone else. Change blindness and inattentional blindness are now standard evidence in eyewitness reliability research and in the design of driver and pilot displays, where slow or masked changes routinely escape notice.
Perceiving Time
We have photoreceptors for light and hair cells for sound, but no sensors for time, and apart from the circadian rhythm no evidence of a neuronal clock metering experience (which would anyway be a double transduction, recreating a worldly property for an inner observer). So what plays the role of prediction error for time? Warrick Roseboom, who joined the Sussex centre in 2015, proposed that we infer duration from the rate of change of perceptual content in other senses.
Duration from change (Roseboom et al., 2019; Sherman et al., 2020). The team filmed videos of varying length in different settings (a crowded city street, an empty office, cows grazing near the university). Human judges showed classic biases: long videos underestimated, short ones overestimated (a regression to the mean, known in timing as Vierordt's law and a signature of Bayesian priors), and busy scenes judged longer than quiet ones of equal length. The same network used in the hallucination machine then estimated durations from accumulated change in its activity, with no clock at all, and its estimates matched the human ones, biases included; the fit improved when the network saw only where people were looking. In an fMRI follow-up led by Maxine Sherman, activity in visual cortex, and not in other regions, predicted how long each video seemed.
Free fall and the flickering watch (Stetson, Fiesta and Eagleman, 2007). David Eagleman tested the idea that time slows in moments of high drama because an internal clock speeds up, which should also sharpen perception of brief events. Volunteers wore a wrist display flashing numbers too fast to read, then dropped backwards in a free fall of about 31 metres into a net. They could not read the numbers in mid-air, so there was no sign of a faster clock. Absence of evidence is not evidence of absence, but Seth's verdict: what an experiment. (The same study found people remembered their own fall as about a third longer than others' falls, pointing to a memory effect.)
Seth flags that most time researchers disagree with him: standard models posit a neural "pacemaker", and others locate timing in bodily rhythms such as heartbeat, which his group's work (Suárez-Pinilla et al., 2019) also questions.
The Feeling That It Is Real
Substitutional reality. Alberto Mariola, a PhD student, leads a project (building on Keisuke Suzuki's earlier version in Japan, 2012) to make people experience and believe an environment is real when it is not. Volunteers sit on a stool in the VR lab wearing a headset with a front-mounted camera and look around via the live feed. At some point the feed is switched to pre-recorded footage of the same room. Most people go on experiencing it as real. The set-up lets researchers ask what makes perception feel real and how that fails in depersonalisation and derealisation, where world and self lose their reality.
Even knowing perception is a construction, the red chair still looks really red and really a chair, just as the sun still rises. David Hume made the parallel point about causality: we do not observe it; we project it after repeatedly seeing events in close succession. The mind spreads itself onto the world, "gilding and staining" natural objects with colours borrowed from internal sentiment. Seth generalises: colours, shapes, smells, chairness, change, duration and causality are all Humean projections, exercises in the beholder's share.
| Feature of experience | Phenomenon | Inference behind it | Evidence |
|---|---|---|---|
| Objecthood | A cup has a back; sky and afterimages don't | Richness of counterfactual sensorimotor predictions | VR objects with lawful rotation reach awareness faster |
| Change | Seen without motion; missed despite motion | Change is inferred, not read off | Rotating snakes; slow colour-change blindness |
| Time | Durations stretch and shrink | Rate of perceptual change plus a prior | Network and visual-cortex models match human biases |
| Reality | The world feels really there | A property of generative models built for action | Substitutional reality fools most people |
| Causality | Events seem to cause each other | Projected from regular succession (Hume) | Philosophical argument |
Why should constructions feel objectively real? Because perception exists to guide action. We respond faster and better to things perceived as really existing, so "out-there-ness" is a necessary feature of a generative model that anticipates its sensory flow. We do not experience our models as models: we perceive with and through them (philosophers call this "transparency"), and so out of mere mechanism a structured world is brought forth.
Seth notes a social corollary: if everyone experiences their perceptions as directly revealing reality, accepting that others perceive differently is hard, as the furious Dress arguments showed. Such "perceptual drift", a generalisation of social-media echo chambers, matters for resolving differences between individuals and groups.
Chipping at the Hard Problem
The real-problem approach is paying off. Starting from the brain's need to infer hidden causes, we can explain why experience is filled with cups, colours and causality-ness that seem to belong to an external world, and that "seeming to be real" is itself a property of perceptual inference. That property fuels dualist intuitions and with them the hard problem: it is extraordinarily difficult to accept that perceptual experiences do not, and never directly, correspond to mind-independent things. In Seth's formula: a chair has a mind-independent existence; chairness does not.
The hard problem seems hardest when we take the contents of perception to exist out there in the world, which is exactly what normal perception encourages. Seth's aim is to dissolve the hard problem rather than solve or rebut it, avoiding both veneration of consciousness as magic and dismissal of it as an illusory non-problem, much as the hunt for life's "special sauce" faded as its properties were explained. The next question is who, or what, is doing the perceiving.
Critical notes. Seth's Vienna is compressed: Klimt, Freud and Riegl belong to around 1900, but the Vienna Circle formed in the mid-1920s (Gödel was born in 1906). The dissolution argument is the chapter's boldest move and the least secure: explaining why experiences seem objectively real explains a belief about experience (Chalmers' meta-problem), and hard-problem defenders reply that the question of why there is any experience at all is untouched. Several key studies are from Seth's own lab with modest samples, and he is candid that the VR objecthood test uses speed of access as a proxy. His clockless account of time is a minority view among timing researchers.
Perceptual diversity at scale. Building on The Dress, Seth and philosopher Fiona Macpherson launched the Perception Census in 2023, a large online study collecting data from tens of thousands of participants on individual differences across many perceptual tasks, to map how much people's controlled hallucinations differ. In AI, Deep Dream has been overtaken by diffusion-based image generators that produce images by iteratively refining noise towards a prompted "prediction", an even closer computational cousin of projecting priors onto data.
Hallucination, priors and AI outputs. Seth's definition is useful operationally: hallucination is prior overwhelming evidence. A generative AI system "hallucinates" in exactly this sense when its learned priors (what text usually looks like) outweigh weak or absent evidence in context. Mitigations map onto his framework: strengthen the evidence (retrieval, tool outputs), raise its precision relative to the model's priors (instructions to rely on supplied sources, citations, constrained decoding), and add "substitutional reality" checks, tests where the ground truth is secretly swapped, to see whether a system notices when its inputs stop matching the world.
Delirium
Part III, Self, opens with a personal case: in 2018 Seth's mother, aged eighty-three, developed hospital-induced delirium and for days was not herself. The episode poses the question the next five chapters answer: what is a self, and is it the kind of thing one can leave and return to? Seth's short answer here: the self is not what it seems.
Two Summers at the Radcliffe
In the summer of 2014, Seth's mother slipped into a vegetative state on the surgical emergency unit of the John Radcliffe Hospital, Oxford, from an undiagnosed encephalopathy (a disease of the brain) while being treated for bowel cancer. Seth flew back from a conference in Brisbane; she recovered slowly and remembers little. Four years later, in the heatwave and World Cup summer of 2018, she was readmitted with a recurrence of bowel pain, and two days after admission developed intense hallucinations and delusions. This time the fragmentation was different: not loss of consciousness but a dissolving of her sense of self and world.
The ward, day by day. Seth finds her hunched, dishevelled, empty-eyed, reporting people crawling up the walls, unsure where she is or why. On the worst day, a Friday, she is convinced she is the subject of a grand and cruel experiment run by everyone around her, with her son as ringmaster inducing hallucinations through her medication. Normally charming and kind, she barks at nurses, tries to escape, and orders the doctors to take her "mad-scientist son" away. Subtler changes linger for days after the acute phase: each word spoken separately rather than as a fluent stream. By day fourteen the acute confusion has passed, and she is amazed to hear what she believed.
Delirium comes from the Latin delirare, to deviate or be deranged, and entered English in the sixteenth century. The dictionary defines it as an acutely disturbed state of mind marked by restlessness, illusions and incoherence. Unlike dementia, which is chronic and degenerative, delirium is usually temporary: it waxes and wanes, though it can last weeks. Seth finds it telling that a word redolent of Victorian asylums is still a working diagnosis, a sign of how far psychiatric medicine still has to travel.
| Feature | Delirium | Dementia |
|---|---|---|
| Onset | Acute, over hours or days | Gradual, over months or years |
| Course | Fluctuates, waxes and wanes | Progressive decline |
| Duration | Usually temporary (days to weeks) | Chronic |
| Typical features | Hallucinations, delusions, disorientation | Memory and cognitive loss |
| Relationship | Raises later dementia risk | Raises delirium risk |
Why Hospitals Induce It
- Epilepsy. Neurological instability lowers the threshold.
- Infection and fever. Systemic illness disturbs brain function.
- Major surgery, or a condition that needs it.
- Dehydration, lack of food and lack of sleep. Basic bodily regulation fails.
- Medication side effects. Many drugs act on the brain.
- Unfamiliarity of place. The factor that makes this delirium "hospital-induced".
All six applied to his mother. A surgical emergency unit is close to a machine for dissociation from the real world: constant beeps and flashing lights, barely a glimpse of outside, a world shrunk to a bed, a chair and a corridor, a changing parade of similar but different nurses and doctors, every day like the last. Patients arrive with a disease of the body, and that is what gets treated; problems of mind and brain that develop on the ward are often not recognised as the medical emergency they are.
- Up to one third of elderly patients entering acute care develop hospital-induced delirium; the proportion is higher after surgery.
- Long-term consequences include reduced cognitive capacity, higher mortality in the following months, and raised risk of further delirium and of dementia.
- Seth's mother: 83 years old, acute phase resolved by day 14.
Seth's practical response follows the logic of the risk list: he drives to the village house he grew up in and brings back familiar objects as "flotsam" for reorientation: a framed photograph, her spectacles, a cardigan, and an old stuffed lion from his own childhood.
Delusions Have a Logic
Delusions are rarely random. His mother's fit her situation with twisted accuracy: Seth does experiment on people, and in the hospital he moved fluidly between son and quasi-doctor, reading notes and murmuring to consultants about paraneoplastic encephalopathies. Her brain, deprived of reliable anchors, built the most coherent story available from the evidence it had. In the language of the book so far, the brain is always trying to find closure, to make its best guess, and when inputs are scrambled the best guess can be a persecutory one.
Seth's own life starts to feel unreal at the edges: mornings and evenings on the ward, afternoons swimming in the Thames and walking Port Meadow, whose heat-baked fields look like savannah, where a horse chases him one morning. He hopes his mother will "fully return to her self", and the phrase is the hook.
If a person can be so changed that her son sees a woman who looks like his mother but is not his mother, then the self is not a fixed essence but something the brain constructs, and can construct differently.
The self is the next target of the real-problem approach. Having treated perception of the world as controlled hallucination, Seth now turns the same lens inward: the experience of being a self is also a perception, built from many parts, and so it can fragment, distort and recover, as his mother's did.
Delirium research since 2021. Delirium has gained recognition as a common and serious acute brain dysfunction, given further prominence by high rates among COVID-19 inpatients in intensive care during 2020-2022. Prevention programmes built on non-drug measures (reorientation, familiar objects and people, sleep protection, hydration, mobility, hearing and vision aids), the approach of the Hospital Elder Life Program, have the best evidence base; Seth's instinct to bring the photograph and the stuffed lion matches standard guidance.
The returning self. Imagine a perfect recording of every belief, memory and habit Seth's mother had before admission, and a second recording made on the worst Friday. Both are her brain's best guesses about who and where she is. Ask which one is "really" her and the answer cannot come from the recordings alone; it depends on which model better predicts her body, her history and the people around her. That is a first hint of Seth's view: a self is a model judged by how well it keeps an organism in touch with its circumstances, not a hidden object that can go missing.
Expect Yourself
The self is not the thing that does the perceiving; it is another perception, a controlled hallucination of a special kind. Seth splits selfhood into five levels (bodily, perspectival, volitional, narrative, social), shows each can be manipulated in the lab or broken by illness, and argues they are normally bound into one unified experience that does not signify an actual, indivisible self. The chapter ends on the book's turn from knowing to regulating: we perceive ourselves in order to control ourselves.
The Teletransportation Paradox
Eva and the faulty vaporiser. A century from now, Star Trek-style machines scan a person down to each molecule and build an exact replica elsewhere, say Mars, while the original is instantly vaporised (to stop an explosion of identical people). Eva steps in at London and feels she simply reappears on Mars. One day the vaporiser fails: Eva1 is still in London when a technician shuffles in with a gun and a consent form, explaining she has "safely" arrived on Mars already. The thought experiment, usually attributed to Derek Parfit and independently to Stanisław Lem, exposes our biases about what a self is.
Seth separates two problems. The consciousness-in-general problem (would the replica be conscious at all?) he finds uninteresting: a molecule-perfect copy would be conscious, and conscious in the same way; anything less returns us to the zombie debates of earlier chapters. The interesting one is personal identity: is Eva2 on Mars the same person as Eva1? What seems to matter is psychological continuity, not physical continuity (our cells turn over anyway, most within about ten years, a biological Ship of Theseus). Seth's "strange" answer: both are the real Eva. And from the moment of replication, small differences in sensory input make them diverge like identical twins, laying down different memories and becoming different people, much as you now relate to you ten years ago, or as his mother after delirium is both changed and recognisably herself.
Doubting the Single Self
We treat self-experience differently from world-experience: it seems to reveal an actual self, and an actual self would have to be one, not two or two-thirds. This indivisible, immutable, transcendental self is baked into the Cartesian immaterial soul, but sceptics have pressed on it for centuries.
- Immanuel Kant: in the Critique of Pure Reason, rejected the self as a "simple substance".
- David Hume: the self as a "bundle" of perceptions (hence bundle theory).
- Thomas Metzinger: German philosopher whose Being No One (2003) is "a very brilliant" deconstruction of the singular self.
- Buddhist traditions: deny a permanent self and seek selfless states through meditation.
- Ayahuasca ceremonies: South American rituals that strip away the sense of self using dimethyltryptamine (DMT).
- Oliver Sacks: neurologist who chronicled how selves fall apart after brain damage.
- Krista and Tatiana Hogan: craniopagus twins from British Columbia, joined at the head and sharing brain structures; one can reportedly feel the other drinking orange juice.
Split-brain patients (Chapter 3) raise the possibility of one self becoming two; craniopagus twins raise the reverse, two selves partly sharing one brain. Being you is not as simple as it sounds.
The Five Levels of Selfhood
- Bodily (embodied) self. Identifying with this object as my body (a sense of ownership no other object gets), plus emotions, moods, arousal and alertness, and beneath them a formless "feeling of being alive", of simply being a body, with no clear spatial extent. This bedrock returns in Chapter 9.
- Perspectival self. A first-person point of view, a subjective origin for perception that usually seems located inside the head, between the eyes and slightly behind the forehead. Its emblem is Ernst Mach's 1886 self-portrait, "View from the Left Eye".
- Volitional self. Experiences of intention (intending to do things) and agency (being the cause of things that happen). This is what people mean by "free will", the aspect of selfhood they are least willing to give up to science (Chapter 11).
- Narrative self. Personal identity: a name, a history and a future, a thread of autobiographical memory. It brings sophisticated emotions such as regret (versus mere disappointment) and even "anticipatory regret", the certainty that what I'm about to do will turn out badly and that I'll do it anyway.
- Social self. How I perceive others perceiving me, arising from being embedded in a social network. It develops through childhood, may develop differently in autism, and brings new ways to feel bad (guilt, shame) and good (pride, love, belonging).
The bodily, perspectival and volitional levels can all exist before any concept of personal identity; the higher levels both shape and are partly defined by the richer emotions they make possible. In normal life the levels are bound together as naturally as colour and shape bind in a red chair, which makes the unity easy to take for granted. Seth's warning: just as experiences of redness do not indicate an external "red", experiences of unified selfhood do not signify an "actual self".
| Level | What is perceived | How it comes apart |
|---|---|---|
| Bodily | This body is mine | Phantom limb (felt, often painful, limb that is gone); somatoparaphrenia (my limb belongs to someone else); xenomelia (intense desire to amputate a healthy limb, rarely acted on) |
| Perspectival | I am here, looking out | Out-of-body experiences and other dissociative disorders |
| Volitional | I intend and cause my actions | Schizophrenia and alien hand syndrome (reduced connection with own actions); akinetic mutism (ceasing to interact at all) |
| Narrative | I am a continuing person | Dementia, severe amnesia; warped in delirium |
| Social | Others see me thus | Atypical development (e.g. autism, in Seth's mention); isolation |
The self is not a pilot behind the eyes steering the body like a plane (Douglas Harding's On Having No Head, 1961, is a classic first-person deconstruction of that intuition). It is a collection of perceptions, a tightly woven bundle of neurally encoded predictions geared towards keeping your body alive, and that, Seth says, is all we need to be.
Owning a Body: The Rubber Hand Illusion
Body ownership is precariously assembled, as clinical conditions show and as simple experiments reveal. The best known is the rubber hand illusion, first described by Botvinick and Cohen in 1998 and now a cornerstone of embodiment research (with follow-ups on three-hand and no-hand illusions, hands of different skin colour and implicit racial bias, even a "rubber tail" illusion in mice).
The rubber hand illusion. The volunteer's real hand rests out of sight behind a cardboard partition; a rubber hand lies in front of her where her hand would normally be. The experimenter strokes both with paintbrushes. With synchronous stroking, many people feel the fake hand is somehow part of their body, while knowing it isn't; with asynchronous stroking the illusion fails. A test: threaten the rubber hand with a hammer or knife and a working illusion produces a strong reaction. Experience varies considerably across people.
In predictive terms, seeing the rubber hand touched while feeling (unseen) touch on the real hand supplies enough evidence for a best guess that the rubber hand is part of the body, because of a prior that simultaneous signals share a common source. Asynchrony removes that evidence.
Where Am I? Full-Body Illusions and Out-of-Body Experiences
Full-body illusions (Science, 2007). Two near-simultaneous papers in Science extended the rubber hand to the whole body using VR. In Olaf Blanke's Lausanne study (Lenggenhager et al., 2007), volunteers in a head-mounted display saw a virtual image of their own back about two metres in front, stroked with a paintbrush in or out of time with strokes on their real back. With synchronous stroking, most felt the virtual body was to some extent their own, and when asked to walk to where they felt they were, they drifted toward the virtual body's location. Henrik Ehrsson's companion study (2007) produced similar effects; both worked with a computer-generated avatar too. The upshot: "what is my body" can be partly dissociated from "where I am".
Out-of-body experiences (OBEs), reported in near-death experiences, operating theatres and around epileptic seizures, have long fuelled belief in an immaterial essence of self: if you can see yourself from outside, surely consciousness is separable from the brain? Seth's reply: no "dualistic skyhooks" are needed if the first-person perspective is itself perceptual inference.
Stimulating the perspective. In the 1940s, Canadian neurologist Wilder Penfield stimulated the right superior temporal gyrus of a patient, G.A., who exclaimed she felt "half here and half not here". Decades later a patient of Blanke's, stimulated near the angular gyrus (the temporo-parietal junction), said: "I see myself lying in bed, from above, but I only see my legs." The common factor is disrupted activity in regions handling vestibular input (balance) and multisensory integration, which yields an unusual best guess about where the perspective is, while other aspects of self stay intact.
Seizure-related OBEs divide into autoscopic hallucinations (seeing one's surroundings from another perspective) and heautoscopic or doppelgänger hallucinations (seeing oneself from another perspective), the latter popularised by Dostoevsky, a severe epileptic, in The Double (1846). Seth's stance: take such reports seriously, since people really have these experiences, but they show only that perspectives are assembled in complex, provisional and precarious ways, not that souls have ever left bodies.
Body Swaps and the Machine to Be Another
Body swap (Petkova and Ehrsson, 2008). Two people wear headsets with cameras whose feeds are swapped, so each sees themselves from the other's viewpoint. The effect kicks in properly when they shake hands: seeing and feeling the handshake, plus top-down expectation, leads each to feel located in the other's body, shaking hands with themselves.
In Ojai, California, in winter 2018, Seth tried The Machine to Be Another, shown by Daanish Masood, a UN peacebroker and VR researcher working with BeAnotherLab in Barcelona, whose aim is "empathy generation". The choreography: both partners look down at their laps and see the other's body; make coordinated movements so the new body seems to obey; see each other in mirrors; then the curtain is removed and each looks at themselves from inside the other's body before they hug. Seth swapped with a well-to-do woman in her seventies, noticed "his" glittering sneakers, and found the mirror and hug unexpectedly powerful. Only at dinner did he wonder what it was like for her to become "a mixed-race English neuroscientist with boring shoes".
How Strong Are Body Illusions? The Suggestion Problem
These illusions have limits. Unlike Adelson's checkerboard, where people are astonished to learn the squares match, body ownership illusions almost never produce that surprise: you feel the fake hand is sort-of yours while knowing it isn't, and many people feel little. Even in Ojai Seth never came close to believing he was someone else.
Suggestibility and the rubber hand (Lush, Dienes, Seth et al., 2020). Psychologists Peter Lush and Zoltán Dienes reasoned that the set-up carries a strong implicit expectation of what should be felt. Illusion strength correlated with hypnotic suggestibility on a standard scale: highly hypnotisable people reported strong ownership under synchronous stroking; low scorers were hardly susceptible. The paper calls the trait "phenomenological control" (avoiding the baggage of "hypnosis"); the worry traces to Orne's 1962 "demand characteristics". The methodological sting: contrasting synchronous and asynchronous stroking does not control for this, because people expect different things in each condition.
The finding cuts both ways. It fits controlled hallucination, since a suggestion is a powerful (possibly unconscious) top-down expectation. But it threatens the field: unless studies measure individual suggestibility, which most have not, they cannot say much about mechanism, and the rubber hand illusion might be largely driven by suggestion. The same applies to full-body and body-swap set-ups, and even to physiological measures, since suggestion can alter physiological responses too (as in hypnotic analgesia). Seth turns it into an opportunity: suggestion is a tool for studying how expectations generate experience. The clinical cases (somatoparaphrenia, xenomelia, phantom limbs, seizure OBEs) carry far more conviction and are the stronger evidence that embodiment and perspective are constructions.
The rubber hand debate since 2020. The Lush et al. paper (Nature Communications, 2020) triggered a sustained exchange. Ehrsson and colleagues replied (2022) that the illusion does not reduce to suggestion, pointing to evidence from proprioceptive drift, skin-conductance responses to threat, brain-imaging signatures, and tight dependence on timing and spatial congruence that expectation alone struggles to explain; Lush, Seth and colleagues maintained that expectancy and demand must be measured and controlled. The practical legacy is methodological: newer embodiment studies increasingly measure participants' expectations and suggestibility alongside the illusion.
The Narrative Self: Clive Wearing
Narrative and social selfhood is where we experience continuity across a lifespan, attach the self to a name, and become truly self-aware (what Endel Tulving called "autonoetic consciousness"). These levels are fully dissociable from the embodied self: infants and many animals may have bodily selfhood without personal identity.
Clive Wearing. A British musicologist known for editing Orlande de Lassus, choirmaster in London, and reshaper of BBC Radio 3's music programming in the early 1980s. In March 1985 a herpes encephalitis destroyed much of his hippocampus in both hemispheres (the seahorse-shaped memory-consolidation structure in the medial temporal lobe), leaving severe retrograde and, above all, anterograde amnesia. He lives in a permanent present of seven to thirty seconds, a continual series of mini-awakenings. His diaries repeat a "first" waking, earlier entries crossed out or angrily obliterated: "8:31am Now I am really, completely, awake", then 9:06, then 9:34. He told his wife Deborah (whose memoir is Forever Today, 2005) that it was "like being dead".
What Clive lost is episodic, autobiographical memory, and with it his narrative self. Yet body ownership, first-person perspective and voluntary action are intact; his love for Deborah (they married a year before the illness) persists even when he cannot remember meeting her; and at the piano, singing or conducting, music flows and he seems whole. Oliver Sacks wrote in The New Yorker (2007) that at such times he is "himself again and wholly alive". Seth's lesson: memory is not the whole of selfhood, but perceiving oneself as continuous over time is very hard to do without, as anyone close to dementia knows.
The Social Self: Perceiving Others Perceiving Me
Theory of mind, perceiving others' mental states, runs constantly in the background. Much of it is direct and automatic: when Seth pours himself more wine and his friend nudges her empty glass forward, he perceives her intention (and that he should have poured hers first) as effortlessly, if not as accurately, as he perceives the glass.
- Same mechanism, deeper causes. Social perception is inference about hidden causes, but those causes (beliefs, intentions) reach us through faces, gestures and speech, each stage a new chance for an off-target guess. That is why we misread minds but never mistake a wine glass for a car.
- Social active inference. We reduce "social prediction errors" by changing the data as well as the predictions: we smile partly to change how a companion feels, and speak to insert thoughts into another's mind.
- Reciprocal generative models. To model your mind, my model must include how you model my mind. We perceive ourselves refracted through the minds of others; that nested prediction is the social self.
- Others' minds as "real". With Colin Palmer and Jakob Hohwy (2015), Seth proposed that others' mental states feel real to the extent that we encode rich conditional predictions about how they would change given our actions (the social counterpart of objecthood), a possible lens on social perception in autism.
Seth adds two cautions. Direct social perception is contested (Gallagher and others hold that mental states are inferred from behaviour in a non-perceptual way). And mirror neurons, found in monkeys by Giacomo Rizzolatti's group (firing both when acting and when watching the same act), carry an unreasonable explanatory burden when invoked to explain all social life, the same error as crediting one fMRI blob with "love".
Self-awareness may require other minds. If no relevant minds existed, a brain would have no need to predict others' mental states, and so no need to infer that its experiences and actions belong to a self at all. John Donne's "no man is an island" could be literally true; Chris Frith goes further, holding that the primary function of conscious experience is social.
The Stable Self and Self-Change Blindness
We experience ourselves as continuous and unified across time: the subjective stability of the self. World-perceptions are in constant flux; self-perceptions seem to change very little, an "enduring centre within a changing world". William James called it "the feeling of the same old body always there" (1890). One might say bodies simply change less than the world, and the body always comes with us. Seth thinks there is more to it.
- Change of perception is not perception of change. As Chapter 6 showed, slow changes in the world can go unseen. The same holds inward: you are a slightly different person now than when you started this chapter, but you do not perceive the change.
- Exaggerated self-change blindness. Our blindness to self-change exceeds what slowly changing bodies and brains would warrant. It also feeds the false intuition that the self is an immutable entity, though that is a side effect, not the evolutionary reason.
We do not perceive ourselves in order to know ourselves; we perceive ourselves in order to control ourselves, and that regulatory purpose is what Seth now pursues down to the body itself.
Critical notes. The five-level taxonomy is a useful scaffold but Seth's own; others carve the self differently (Gallagher's minimal versus narrative self, Damasio's proto-self, core self and autobiographical self), and the levels interact more than a ladder suggests. Seth's answer that both Evas are real is defensible (Parfit concluded that identity is not what matters), but many philosophers reject it on physical-continuity or "closest continuer" grounds. Note the honesty of the chapter: Seth reports his own lab's finding that undercuts much embodiment research, including evidence often cited for his view. One small attribution slip: BeAnotherLab is generally credited to an independent Barcelona artist-researcher collective, while Mel Slater, a computer scientist and leading VR-embodiment researcher in Barcelona, is better known for his own lab's work than as its founder.
Expectation effects in AI evaluation. The rubber hand lesson transfers directly to how enterprises judge AI copilots. If users are told a relationship-manager assistant "saves an hour a day", their reported experience will partly follow the expectation, exactly as suggestible participants "own" the rubber hand. Robust evaluation needs the equivalent of Lush's controls: measure prior expectations, use blinded or placebo arms (e.g. a baseline tool presented identically), and prefer behavioural measures (time to complete, error rates, rework) over self-report alone.
Being a Beast Machine
Self-perception is not about discovering what is there, in the world or in the body; it is about physiological control and staying alive. Seth turns Descartes upside down: where Descartes' bêtes-machines made life irrelevant to mind, Seth's beast machine theory holds that we cannot understand conscious experience except in light of our nature as living creatures. Emotions and moods are interoceptive inference, control-oriented predictions about the body's essential variables; beneath them lies a formless feeling of just being alive, the ground state from which all perception, of self and world, follows.
From the Great Chain to the Beast Machine
The Great Chain of Being (Scala Naturae), developed from Plato, Aristotle and Plotinus into its full medieval Christian form, ranked everything: God, angels, humans (with "socially convenient subdivisions", kings above peasants), animals, plants, minerals. Humans sat awkwardly between the spiritual and the physical: immortal souls capable of reason, love and imagination, yet prey to bodily passions such as pain, hunger and desire.
In the seventeenth century, René Descartes collapsed the gradations into two kinds of existence: res cogitans (mind stuff) and res extensa (matter stuff). This created the interaction problem that has framed consciousness research ever since, "largely for worse", and threatened religious and political order: if animals had minds, what stopped them aspiring to the spiritual? Descartes handled the Church carefully, even attempting proofs of a benevolent God in the third and fifth Meditations.
Descartes: the beast machine. Animals lack souls and with them reason and (probably, though the reading is uncertain) the rational and spiritual attributes souls confer. Bodies of man and beast are machines that breathe, digest, perceive and move by the arrangement of their parts; only humans show reason, through true speech. The flesh-and-blood properties of living things are irrelevant to mind: mind and life are as sharply divided as res cogitans and res extensa. (He kept a devoted pet dog, Monsieur Grat; he also vivisected rabbits.)
La Mettrie: l'homme machine. Julien Offray de La Mettrie pushed the argument through the open door: if animals are machines and humans are animals, then humans are machines too (L'Homme machine, 1748). He denied the soul any special immaterial status and questioned God, was forced to flee the Netherlands in 1748 for Frederick the Great's Berlin, and died three years later after eating too much pâté. For him mind is a property of life: the two are continuous.
Seth's sympathies lie with La Mettrie, but his target is consciousness rather than "mind". The core proposal of the beast machine theory: our experiences of the world and of the self happen with, through and because of our living bodies. Animal constitution is not merely compatible with conscious experience; we cannot understand its nature and origin except in light of our nature as living creatures.
Interoception: Perceiving the Body from Within
Below narrative identity, below the first-person perspective and body ownership, lie deeper layers of selfhood tied to the interior of the body: affective experiences (emotions and moods) and, deepest of all, a formless, ever-present sense of simply being a living organism.
- Exteroception. Perception of the outside world: sight, hearing, taste, touch, smell.
- Proprioception. Perception of body position and movement, sitting between the two.
- Interoception. "The sense of the internal physiological condition of the body" (Bud Craig's 2002 definition): signals from the viscera about heartbeat, blood pressure, blood chemistry, gastric tension, breathing, travelling via the brainstem and thalamus to interoceptive cortex, above all the insular cortex (named as an "island" in the cortical sea). Not to be confused with introspection.
The key property of interoceptive signals: they reflect, one way or another, how well physiological regulation is going, how good a job the brain is doing of keeping its body alive.
Theories of Emotion, from James to Interoceptive Inference
In 1884, William James and, independently, Carl Lange argued that emotions are neither "eternal and sacred psychic entities" nor (as Darwin had just proposed) hard-wired circuits, but perceptions of changes in bodily state. James: "We feel sorry because we cry, angry because we strike, afraid because we tremble", not the reverse. The view was resisted partly because it inverts how things seem, which by now we know is no guide to how things are. A more serious objection was that bodily states may be too few and too similar to support the full range of emotions.
The answer that emerged in the 1960s was appraisal theory (the classic is Schachter and Singer, 1962): emotions depend on cognitive appraisal of the context in which bodily changes occur, so closely related emotions (say listlessness and ennui) can share one bodily state and differ in interpretation. Seth suspects every emotion may in fact have a distinct but hard-to-detect bodily signature.
The Capilano bridges (Dutton and Aron, 1974). A female interviewer approached men crossing one of two bridges over the Capilano River in North Vancouver: a 450-foot rickety suspension bridge with low handrails high above rapids, or a sturdy cedar bridge only ten feet above the water upriver. She asked them to complete a questionnaire and offered her phone number for follow-up questions. More men from the rickety bridge called her, many angling for a date: "misattribution of arousal", fear-driven arousal read as sexual attraction. With a male interviewer, bridge type made no difference (assuming heterosexual participants). Seth notes the methodological weaknesses and dubious ethics by modern standards, and a coda: crossing the Sharp Edge ridge on Blencathra in September 2020, he found "Marry me, Maria?" chalked on a stone at its base.
Appraisal theory's weakness is its sharp line between "non-cognitive" readout and "cognitive" evaluation, which the brain does not respect. Around 2010, as his Sussex group began, Seth, learning about interoception from his colleague Hugo Critchley, proposed treating emotions as controlled hallucinations of the body: interoceptive inference (outlined 2011, standard reference Seth, 2013). The causes of interoceptive signals are as hidden behind the sensory veil as the causes of vision, so they too must be inferred.
Emotions and moods are the subjective aspects of predictions about the causes of interoceptive signals, in the same way that redness is the subjective aspect of predictions about how surfaces reflect light.
| Theory | Era, key names | What an emotion is | Main problem |
|---|---|---|---|
| Basic emotions | Darwin (1872); Jaak Panksepp | Output of innate, evolutionarily old circuits | Rigid; poor fit to variability |
| James-Lange | 1884 | Perception of bodily change | Too few distinct bodily states? |
| Appraisal | 1960s; Schachter and Singer | Bodily arousal plus cognitive evaluation of context | Assumes a cognitive/non-cognitive split |
| Constructionist | Lisa Feldman Barrett, Joseph LeDoux | Built by interpretation and prediction | Close to Seth's view |
| Interoceptive inference | Seth (2011-13), with Damasio's and Barrett's influence | Top-down best guess about the body's state and its causes | Hard to test |
Interoceptive inference is more parsimonious than appraisal: one process (Bayesian best guessing at many scales of time and space) rather than two, so it handles emotional range without a bright cognitive line and maps better onto brain anatomy. Seth acknowledges the deep debts to Antonio Damasio and to Barrett, who reached closely related ideas independently.
Testing interoceptive inference. Interoceptive signals are hard to measure and manipulate, so evidence is still thin. - Heartbeat evoked potentials. Frederike Petzschner (2019) showed brain responses to heartbeats, candidate interoceptive prediction errors, are modulated by attention, as the theory predicts. - Cardio-visual synchrony. Keisuke Suzuki (2013) found stronger ownership of a virtual rubber hand that flashed in time with the heartbeat; Jane Aspell and colleagues found stronger identification with a virtual body silhouette pulsing with the heartbeat. Body ownership seems to integrate exteroceptive and interoceptive signals. - Caveats. These studies ignored suggestibility (Chapter 8) and depend on heartbeat awareness, which is notoriously hard to measure. Animal work hints that mouse insular neurons encode something like interoceptive predictions (2019-2020).
The decisive implication: affective experiences are not merely shaped but constituted by interoceptive predictions. Fear, anxiety, joy and regret all come from the inside out.
Cybernetics and the Good Regulator
Why should the body's interior be built into consciousness from the ground up? Seth's answer comes from cybernetics, from the Greek kybernetes, "steersman", defined by Norbert Wiener (1948) as "the scientific study of control and communication in the animal and the machine". It focused on closed-loop feedback (guided missiles), and so on machines that seem to have purposes, a bridge from non-living to living. It lost the funding and headlines to disembodied, abstract-reasoning AI epitomised by chess programs, but its insights are only now being appreciated.
- The theorem. "Every good regulator of a system must be a model of that system."
- System A. A thermostat holding a house at 19°C (about 66°F) by simple feedback: too cold, switch on; otherwise off. It merely is a model, in a fixed, inflexible way.
- System B. Predicts how the house temperature would respond to heating on or off, using room sizes, radiator positions, wall materials and the weather, and can anticipate a cold day and act in advance. It has a model and regulates better.
- System B with noisy sensors. Temperature cannot be read off; it must be inferred from sensor data plus priors. The system needs a model of (i) how readings relate to their hidden cause and (ii) how that cause responds to actions.
- The lesson. Its goal is not to find out the temperature but to regulate it. Perception here is for control, which is active inference: minimising prediction error by acting.
Active inference showed that predictive perception can aim at finding out things or controlling things (in Seth's papers, epistemic versus instrumental). Cybernetics adds that for some systems control comes first: the whole apparatus of prediction emerges from what it takes to regulate. The last piece is Ashby's essential variables (1952): physiological quantities such as temperature, blood sugar and oxygen that must stay within tight limits (human core temperature between about 32°C and 40°C, or death follows quickly).
Emotions and moods are control-oriented perceptions that regulate the body's essential variables. Fear at an approaching bear is a perception of "my body in the presence of an approaching bear" that triggers the actions best predicted to keep essential variables in range: external actions (running, via skeletal muscle and the somatic nervous system) and internal "intero-actions" (raising heart rate, dilating vessels, via smooth and cardiac muscle and the autonomic system).
Why Emotions Feel the Way They Do
Fear, jealousy, joy and pride differ, yet they resemble one another more than any resembles a visual or auditory experience. Seth's explanation: phenomenology depends on the target of a prediction and on its type. Finding-out predictions (how a cup's image would change if rotated) yield objecthood. Control predictions yield something else.
Catching a cricket ball. The intuitive strategy, computing where the ball will land and running there, is not what expert fielders do. They move so the ball "looks the same" in one respect: the angle of elevation of gaze rises at a steadily decreasing rate. This "optic acceleration cancellation" (McLeod et al., 2003) guarantees interception (taken too literally, the ball hits you between the eyes). The fielder perceives not the ball's position but its "catchability": perception is a controlling hallucination as much as a controlled one.
Precedents: James Gibson's affordances (1970s; we perceive opportunities for action, a door for opening) and William Powers' perceptual control theory (1973), whose slogan is that control systems control what they sense, not what they do: we behave so as to perceive things a certain way. Both lacked concrete predictive mechanisms and looked outward, not into the body.
Anxiety doesn't have a back, sadness doesn't have sides, and happiness is not rectangular. Interoceptive predictions give no experience of the shape or location of spleen and kidneys, no objecthood and no spatial movement (metaphors like a "sharp" pang aside). Because they predict how actions will affect essential variables, we experience how well or badly our overall situation is going and is likely to go: desolate, hopeful, panicky, calm.
| Prediction type | Aim | Example | Resulting phenomenology |
|---|---|---|---|
| Epistemic, exteroceptive | Find out what is there | Rotating a coffee cup | Objecthood, things in space |
| Instrumental, exteroceptive | Control how things appear | Catching a cricket ball | Affordance, "catchability" |
| Instrumental, interoceptive | Keep essential variables in range | Heart racing near a bear | Valence and arousal: how things are going |
| Deepest bodily prior | Stay alive | Continuous regulation | Formless feeling of being alive |
The Ground State: The Feeling of Being Alive
Seth builds the core of the beast machine theory in steps.
- The prime imperative. Every organism's primary goal, almost by definition, is to stay alive, maintaining physiological integrity against danger and opportunity.
- Why brains exist. Not for poetry, crosswords or neuroscience, nor fundamentally for perceiving the world: brains exist to keep essential variables within the narrow ranges that survival allows.
- Hidden causes. Those essential variables are the causes of interoceptive signals, and like all causes they are hidden, so they must be inferred.
- Interoceptive inference. Inference happens by prediction error minimisation: interoception is controlled hallucination.
- Control, not discovery. Interoceptive inference is mainly about regulation, an exemplar of active inference: errors are reduced by acting (reaching for food; gastric reflexes; blood pressure changes), as well as by updating.
- Anticipation. Waiting for blood acidity to go out of bounds would be fatal, so control is predictive: fleeing before the bear bites; a transient blood-pressure rise to stand up from the desk. Physiology calls this allostasis, "stability through change" (Peter Sterling), versus homeostasis, a tendency to equilibrium.
- Emotion. Interoceptive predictions underpin emotions and moods, which feel as they do because the predictions are control-oriented and body-directed.
- The ground state. Emotions mostly point at external things (I fear something), but the deepest selfhood, the inchoate feeling of "just being", has no external referent: a formless, control-oriented prediction about the body's present and future condition. This is where being you begins.
- Everything follows. All perceptions, of self and world, are inside-out controlled and controlling hallucinations rooted in predictive machinery that evolved, develops and operates in the service of staying alive.
The conclusion inverts Descartes. Life is not irrelevant to mind; it is the opposite. "We are conscious beast machines, through and through."
Why the Self Feels Stable and Real
The beast machine theory explains the subjective stability left open in Chapter 8.
- Precise priors become self-fulfilling. Regulating the body needs high-precision interoceptive priors, which active inference makes self-fulfilling: body temperature is predicted to be constant, which is why it stays constant. The felt unchanging bodily self comes straight from these strong priors. The brain will never update its prior belief of expecting to be alive.
- Change inferences damped. Because perceived change is itself an inference, the brain may attenuate predictions of bodily change to keep regulation tight: self-change blindness at the physiological level (just as proprioceptive signals are attenuated during movement).
- Useful misperception. Effective regulation may depend on systematically misperceiving the body as more stable and less changing than it is. Seth suspects this generalises to every level of self, which is why we perceive ourselves in order to control, not to know, ourselves (the Stoics' self-control over Socrates' "know thyself"). A new field, computational psychosomatics (Petzschner et al., 2017), asks whether illness lifts this blindness.
The self also feels real, as the red chair does: inward-directed prediction makes it seem there is a stable essence of "me" at the centre. That reality can fail. In depersonalisation the self's reality is attenuated or abolished; in the Cotard delusion, first described by French neurologist Jules Cotard in 1880, sufferers believe they do not exist or are already dead. No essence has left; control-oriented predictions at the deepest layers of bodily regulation have gone awry, perhaps because generative models lose their rich counterfactual interoceptive predictions (Seth and Tsakiris, 2018), the bodily counterpart of losing objecthood.
What the Theory Does and Doesn't Claim
Seth is explicit: he is not claiming life is necessary for consciousness, that only flesh, blood, guts or biological neurons can be conscious. That may or may not be true, and nothing here settles it. (He avoids the label "biological naturalism" because it has been used in different ways.) His claim is explanatory: to understand why experiences are as they are, appreciate the roots of all perception in the physiology of the living.
On the brainstem: tying consciousness to regulation raises its importance beyond a "power cable" enabling factor. Some (Mark Solms, 2021; Björn Merker, 2007) argue consciousness arises in the brainstem without cortex; Seth thinks this extremely unlikely given the evidence for cortex and thalamus, though the brainstem may shape conscious states more than the analogy suggests (Parvizi and Damasio, 2001).
For the hard problem, extending controlled hallucination to the deepest self erodes another intuition: that the conscious self is an immaterial inner observer looking out on a material world. That is one more confusion between how things seem and how they are.
We are not cognitive computers, we are feeling machines.
Seth even finds an echo of the soul: not an immaterial quiddity or distilled rationality, but the perceptual expression of a deep continuity between mind and life, the experience of the inchoate feeling of being as really existing. It recalls older conceptions such as the Hindu Ātman, which saw our innermost essence more as breath than as thought.
The perfect thermostat. Build a house regulator with a superb generative model: it predicts weather days ahead, infers temperature from a hundred noisy sensors, and keeps every room at 19°C. It is a System B par excellence. Does it feel cold? Seth's framework says its predictions are control-oriented and concern its essential variable, yet nothing is at stake for it: if the temperature drifts, the thermostat does not die. The thought experiment isolates what the beast machine theory adds to predictive processing: regulation of a system whose continued existence depends on that regulation. Whether that is the missing ingredient for feeling is exactly the question Chapter 13 takes up.
Critical notes. The Dutton and Aron study is a classic but rests on small samples and a single female interviewer, and misattribution-of-arousal effects have a mixed replication record; Seth himself flags its weaknesses. Interoceptive inference is, by his own admission, thinly tested: heartbeat-detection tasks are unreliable measures, and the cardio-visual studies did not control for suggestion. The theory's attractiveness comes partly from its unifying scope, which also makes it hard to falsify. The account of the "feeling of being alive" as a formless control prediction is a philosophical identification rather than an empirical result. Seth's reading of Descartes on animal consciousness (that he probably did not deny it outright) is a defensible but contested scholarly interpretation.
Seth names his position. Since 2021 Seth has developed the life-mind link into an explicit stance against conscious AI, now adopting the term "biological naturalism" he avoided here. In a 2024 preprint published in Behavioral and Brain Sciences (2025), "Conscious artificial intelligence and biological naturalism", he argues that consciousness may depend on properties of living, self-maintaining (autopoietic) systems and that computation alone is unlikely to suffice. Interoception research has meanwhile expanded rapidly, with growing work on heart-brain and gut-brain signalling and on interoceptive dysfunction in anxiety, depression and eating disorders, though reliable measurement of interoceptive accuracy remains a known difficulty.
Regulators need models. The Good Regulator Theorem is a design principle for any control system, including enterprise ones. A fraud or credit-risk control that only reacts to thresholds is System A; one that models how the portfolio would respond to interventions and acts ahead of a downturn is System B. Seth's twist matters too: a regulator's internal state estimates should be judged by how well they support control, not by how faithfully they describe reality, so model monitoring should track decision outcomes, not only predictive accuracy.
A Fish in Water
Karl Friston's free energy principle (FEP) starts from a near-trivial statement, things that exist must keep themselves distinct from their environment, and derives from it the predictive brain. Living systems resist the second law of thermodynamics by staying in the statistically expected states that define them (a fish in water), which they do by minimising free energy, which in practice equals sensory prediction error. For Seth the FEP grounds the beast machine theory in physics, tells the same story in reverse, and adds a mathematical toolbox; but it is a theory for consciousness science, not a theory of consciousness.
A Very Big Idea
In September 2007, Seth travelled from Brighton to a Barcelona summer school on "brain, cognition, and technology", arriving too late for Karl Friston's three-hour masterclass on the FEP. At the rooftop bar that evening he met a sea of bemused faces: Friston had flown straight back to London, and after three hours of mathematics and neuroanatomy most of the audience was more confused than before.
The scale is the problem. The FEP draws on biology, physics, statistics, neuroscience, engineering and machine learning, and for Friston it explains all features of living systems: self-organisation of a single bacterium, the details of nervous systems, animal body plans, even the broad strokes of evolution. It is as close to a "theory of everything" in biology as has been proposed. Ten years later, in 2017, Seth with Chris Buckley, Simon McGregor and Chang-Sub Kim published a review of "the free energy principle in neuroscience" in the Journal of Mathematical Psychology, "about nine years longer than anticipated". The inscrutability is legendary: Scott Alexander's blog post "God help us, let's try to understand Friston on free energy", and the parody account @FarlKriston ("I am, whatever I think I am. If I wasn't, why would I think I am?").
Existence, Boundaries and Entropy
- To exist is to have a boundary. Without a difference between a thing and everything else, there would be no things. The boundary must also persist over time.
- Passive persistence. A drop of ink in water disperses at once; a drop of oil stays distinct for a while; rocks last longer still. None actively maintains its boundary; each simply disperses slowly enough to be noticed. (Tornadoes and whirlpools are intriguing edge cases.)
- Active persistence is life. Living systems actively maintain their boundaries, by moving or growing. The FEP's starting point: by virtue of existing, they must resist the dispersion of their internal states. A puddle of undifferentiated mush is no longer alive.
- Low entropy. Entropy (Chapter 2) measures disorder or uncertainty. Being alive means occupying a low-entropy condition.
- The second law. The entropy of any isolated system increases, so organised matter is intrinsically improbable and unstable; "in the long run we're all doomed". Living systems temporarily fend off the second law, existing out of equilibrium with their environment, which they can do only because they are open systems, harvesting resources, nutrients and information.
Expected States: The Fish in Water
To resist the second law, a living system must occupy states it expects to be in, "expect" in the statistical, not psychological, sense. A fish in water is in a statistically expected state, since most fish are in water most of the time; a fish out of water is unexpected unless it is turning to mush. A body temperature of about 37°C is likewise expected. Being alive means proactively and repeatedly revisiting a particular set of states: temperatures and heart rates (the essential variables of Chapter 9), or protein complexes and energy flows in a bacterium.
How can expected states also be improbable? A living system inhabits a small "attracting set" of states out of a vast number possible. The set is statistically expected because that is where the system is usually found, yet improbable because far more states lie outside it. In Seth's line: there are many more ways of being mush than there are of being alive.
What matters to an organism is the entropy of its sensory states, those that bring it into contact with the environment. A bacterium needing a nutrient expects to sense high concentrations and moves to keep sensing them, so it stays in the states that define it as alive. The FEP generalises: all organisms stay alive by minimising sensory entropy over time.
Enter Free Energy
The catch: to minimise a quantity a system must measure it, and sensory entropy cannot be measured from the sensations alone. (Is the number 6 surprising? It depends on context.) Free energy solves this: it approximates sensory entropy and can be measured by the organism, so it can be minimised. The name comes from nineteenth-century thermodynamics, where free energy is energy available to do work at constant temperature; the FEP's version is variational free energy, from machine learning and information theory.
- Surprisal (self-information). How statistically unexpected a sensory event is. Under nonequilibrium steady-state assumptions, the long-term average of surprisal is entropy: uncertainty is the average surprise you expect.
- Two densities. A recognition density (the current best guess about hidden environmental causes) and a generative density (a probabilistic model of how those causes generate sensory input).
- Decomposition. Free energy = surprisal + KL divergence between the recognition density and the true posterior. Because the Kullback-Leibler divergence ≥ 0, free energy ≥ surprisal: it is an upper bound.
- Two ways down. Lowering free energy means either better inference (shrinking the divergence: perception) or lower surprisal (sampling different sensory input: action).
- Prediction error. With Gaussian densities and further assumptions (e.g. separated timescales), free energy maps onto precision-weighted prediction error, the currency of predictive processing.
- Model evidence. Minimising free energy is equivalent to maximising Bayesian model evidence; free energy is the negative of the machine-learning evidence lower bound (ELBO). Jakob Hohwy calls this "self-evidencing".
- Markov blankets. The FEP formalises a boundary as a Markov blanket: a statistical partition into internal, external and blanket states, with internal and external states conditionally independent given the blanket.
Because predictions require models, the FEP licenses the claim that living systems have, or are, models of the causes of their sensations; only relative to a model can sensations be judged surprising (if you know the 6 came from a die, you know exactly how surprising it is). The resulting picture: a living system models its world and body so that its defining states keep being revisited, from each heartbeat to "commiserating my birthday every year". Friston's phrasing is that organisms maximise the sensory evidence for their own existence; Seth's: "I predict myself therefore I am."
Two Objections
The dark room problem. If organisms minimise prediction error, why not sit in a dark, silent room where inputs are perfectly predictable? Raised early (Friston et al., 2012) and again while Seth was writing (Sun and Firestone, 2020).
Reply. Other signals, such as blood sugar, soon deviate: you get hungry, sensory entropy rises and non-existence looms. Organisms must let some things change so others stay the same: getting up to make breakfast, blood pressure rising so you don't faint. This is allostasis. Long-run minimisation means leaving the dark room, or at least switching on the lights (Seth et al., 2020).
Unfalsifiability. No experiment can prove the FEP wrong.
Reply. True, but neither unique nor fatal. The FEP is mathematical philosophy, a first-principles answer to Hohwy's question "What are the conditions for the possibility of existence?", as Kant asked about the conditions for perception. Like Hamilton's principle of stationary action, which yields testable equations of motion and even general relativity, it motivates specific, falsifiable theories: predictive processing fails if brains don't use prediction errors. It will be judged on usefulness, not truth.
What the FEP Adds to the Beast Machine
The beast machine theory stands on its own without the FEP. (A statistical-mechanics textbook opens by noting that Boltzmann died by his own hand in 1906, Ehrenfest in 1933: "Now it is our turn to study statistical mechanics.") But the FEP adds three things.
| Benefit | What it means |
|---|---|
| 1. Grounding in physics | The beast machine's "drive to stay alive" reappears as the deeper imperative to remain in expected states against the second law; Friston even holds that a Markov blanket's existence implies active inference (contested by Bruineberg et al., 2020). |
| 2. The story in reverse | Seth's chapters went outside in, from inferring the world to the body's core. The FEP goes inside out, from "things exist" to body and world. Two routes to one destination strengthen coherence and expose parallels (free energy and prediction error). |
| 3. A mathematical toolbox | Staying alive means minimising expected future free energy, which requires seeking sensations now that reduce uncertainty about action outcomes: a curious, sensation-seeking agent. The maths quantifies the exploration-exploitation balance, with consequences for what we perceive (e.g. agents that learn adaptively biased models, Tschantz et al., 2020). |
The FEP is not a theory of consciousness. Despite its "theory of everything" billing, it relates to consciousness exactly as predictive, Bayesian brain theories do: a theory for consciousness science (the real problem), not of consciousness (the hard problem). In return, controlled hallucination and the beast machine give the FEP's austere mathematics relevance to consciousness: "what good is a theory of everything unless it has something to say about that?" Two further cautions: the equivalence of free energy and prediction error holds only under specific assumptions (Gaussian densities and more), and critics such as Bruineberg and colleagues argue that FEP papers slide between Markov blankets as modelling tools and as real boundaries of organisms. The book's FEP account is also unusually reliant on analogy; readers wanting rigour must go to the technical literature.
Aegina: Two Grand Theories in a Storm
In September 2018, shortly after his mother's delirium, Seth joined Friston and about twenty neuroscientists, philosophers and physicists on the Greek island of Aegina, an hour by ferry from Athens, to discuss the FEP's relation to integrated information theory (IIT). A rare Mediterranean "medicane" tossed tables and chairs into the sea; inside, ideas were floated "in semi-darkness".
FEP. Starts from "things exist" and derives neuroscience and biology, but not consciousness.
IIT. Starts from "consciousness exists" and launches a direct assault on the hard problem.
Unsurprisingly they talk past each other. By 2020 an adversarial collaboration was being planned in which proponents commit in advance to what outcomes would support or undermine their theory; this one pits IIT against active inference (not the FEP itself). One proposed test (Chapter 3): IIT predicts that inactivating already-inactive neurons changes experience; active inference predicts no change. Seth's forecast: we will learn much, but neither theory will be explicitly ruled out. His own theory charts a middle course: from the FEP, grounding in the nature of the self and predictive machinery; from IIT, a clear focus on phenomenology, though aimed at the real problem. He hopes the beast machine theory can bring them together.
On the calm ferry back, at peace with a hard decision to miss personal events in Brighton, he began wondering how anyone makes any decision, and what it means to feel in control of our choices. Once you start thinking about free will, there's really no stopping.
The FEP since 2021. The Markov-blanket critique appeared as a major target article, "The Emperor's New Markov Blankets" (Bruineberg et al., Behavioral and Brain Sciences, 2022), with extensive commentary sharpening the distinction between blankets as modelling tools and as real boundaries of organisms. Friston and colleagues have since restated the principle in more general mathematical terms (for example a 2023 Physics Reports review), which defenders see as clarification and critics as making it even less constraining. Seth's forecast about adversarial collaborations held for the first to report: the IIT versus global workspace study eliminated neither theory (see Chapter 1).
Free energy, ELBO and the AI connection. The FEP's core quantity is the same evidence lower bound that trains variational autoencoders and underlies diffusion models' training objectives, so the mathematics is familiar to ML engineers. The dark room problem has a direct analogue in agent design: a system rewarded only for predictability will seek boring, controllable inputs, which is why reinforcement-learning agents use curiosity or information-gain bonuses, the "expected free energy" exploration term in another guise. For enterprise agentic platforms, the lesson is that objectives which reward low surprise must be balanced by explicit incentives to reduce uncertainty, or agents will avoid exactly the edge cases that matter.
The obedient fish. Imagine a fish that, by a quirk of wiring, finds dry land statistically expected: its model predicts air on its gills. It would "minimise prediction error" by leaping onto the bank, and die. The thought experiment shows the FEP's hidden premise: the expected states must be the ones compatible with survival, which in real organisms is guaranteed by evolution and development shaping the priors. Free energy minimisation explains how a viable creature stays viable; it does not by itself say which states are worth expecting.
Degrees of Freedom
Experiences of free will do not reveal an immaterial self pulling strings in the physical world; they are self-related perceptions associated with voluntary actions, Bayesian best guesses that the causes of an action come from within. Seth dismisses "spooky" free will, shows why the determinism debate does not matter, reinterprets Libet's readiness potential through Aaron Schurger's reanalysis, and grounds real freedom in the brain's competence to control many degrees of freedom. Free will is not an illusion: the experience is as real as seeing red, and as useful.
Briony's Finger and Buck-Stopping
The epigraph comes from Ian McEwan's Atonement: thirteen-year-old Briony Tallis bends her finger, hunting for "the crest" of the breaking wave where intention takes effect, hoping to find the part of her that was really in charge; when the finger moves, the action seems to start in the finger itself. Few topics are as inflammatory as free will, and there is no clarity even about the experience. One stable intuition remains, in Galen Strawson's words: a feeling of "radical, absolute, buck-stopping up-to-me-ness in choice and action", absent in a reflex such as pulling a hand from a nettle. This experience, more than any other, makes us feel an immaterial self is at work. That is how things seem.
Spooky Free Will and Determinism
Seth first says what free will is not: an intervention in the flow of physical events that makes things happen that otherwise wouldn't. This "spooky" free will invokes Cartesian dualism, demands freedom from cause and effect, and explains nothing.
- Libertarian free will. The philosopher's version of spooky free will (no relation to the political philosophy): genuine choices escape prior causation.
- Hard determinism. Determinism is true, so no reasonable conception of free will survives.
- Compatibilism (Seth's view). Reasonable conceptions of free will are compatible with determinism. Seth is agnostic about whether the universe is deterministic and a compatibilist either way.
With spooky free will gone, the determinism question stops mattering. Determinism says every event is fixed by prior physical causes; its alternative builds chance in from the ground up (say, quantum fluctuations). If free will is a perceptual experience, no disruption of causal flow is needed and a deterministic universe "chugs along just fine". If determinism is false, nothing changes either, because voluntary actions neither feel random nor are random (neural noise may matter to brain function, but that is beside the point). Gerald Edelman's one-liner: "Free will: whatever you think about it, we're determined to have it."
Libet and the Readiness Potential
Libet's volition experiment (UCSF, early 1980s). Benjamin Libet exploited the readiness potential, a slow, ramp-like EEG signal over motor cortex that precedes voluntary movements (first recorded as the Bereitschaftspotential by Hans Kornhuber and Lüder Deecke, 1965). Participants flexed their wrist whenever they chose, while EMG timed the movement and EEG recorded the brain. They also watched a dot rotating on an oscilloscope and reported its position at the moment they felt the urge to move. Averaged over many trials, the readiness potential began hundreds of milliseconds before the reported urge (in Libet's data roughly 550 ms before movement, versus about 200 ms for the urge).
The standard reading is that Libet "disproves free will". It is certainly bad news for spooky free will, since the conscious urge cannot have caused an action already under way. Libet himself, worried, proposed a rescue: enough time remains for "free won't", a conscious veto (Libet et al., 1983). Seth calls it a cute trick that doesn't work: conscious inhibition is no more a miracle than conscious intention. (Philosophers distinguish urges from intentions, which answer to reasons and norms; for a finger flex they amount to much the same thing.)
Schurger's Reanalysis
Half a second is a very long time in brain time. In 2012, Aaron Schurger proposed that readiness potentials may be artefacts of how they are measured (philosopher Al Mele had made a similar conceptual point in 2009). RPs are found by time-locking backwards from the moments actions occurred, which ignores all the times actions didn't happen.
The high striker. At the circus, punters swing a mallet to send a puck up toward a bell. A "circus scientist" who examined puck trajectories only on the swings that rang the bell would conclude that a rising puck always leads to the bell. To understand the machine she must also look at the swings that fell short. Averaging EEG backwards from actions is examining only the bell-ringers.
Schurger, Sitt and Dehaene (2012). Participants made spontaneous Libet-style movements but were occasionally interrupted by a loud beep prompting the same movement immediately. When they responded quickly to the beep, the preceding EEG showed something like a readiness potential, reaching back long before the beep, though they had prepared no voluntary act. Slow responses showed little sign of one.
Schurger's interpretation: the RP is not a signature of the brain initiating an action but fluctuating activity that occasionally crosses a threshold, triggering action when it does. Looking back from actions reveals a rising slope; beep responses are fast when activity happens to be near threshold and slow when far. Seth adds his own gloss: the fluctuations look like the brain accumulating evidence for a Bayesian best guess, the neural fingerprints of a special kind of controlled hallucination.
Three Features of Volition
Seth's test case: "I just made a cup of tea."
| Feature | What it feels like | What is really going on |
|---|---|---|
| 1. Doing what I want | The act aligns with my beliefs, values, desires, physiology and affordances (thirsty, tea available, nobody force-feeding me hot chocolate) | I cannot choose my wants: Schopenhauer (1839), "Man can do what he wills, but he cannot will what he wills" (the root, Seth notes, of all addiction) |
| 2. Could have done otherwise | I made tea and not coffee, though coffee was there | Given the exact state of universe, body and brain, I could not have; replaying the tape gives the same result bar randomness. The feeling is not a window on causality |
| 3. Comes from within | Unlike a reflex (foot jerking from a stubbed toe), a deliberate backswing to kick a ball feels self-originated | An inference about where an action's causes lie |
Putting them together: We perceive an action as freely willed when we infer that its causes come predominantly from within, aligned with our beliefs and goals, detached from alternative causes in body or world, and suggesting we could have done otherwise. That is how volition feels from the inside and how voluntary action looks from outside. (Seth separates this from willpower, the felt effort some actions need; many voluntary acts take none.)
Degrees of Freedom and the Volition Network
In engineering, degrees of freedom are the number of ways a system can respond. A rock has essentially none; a train on one track has one (forward or back); an ant has quite a few; humans have vastly more. Voluntary behaviour is the competence to control these degrees of freedom in line with beliefs, values and goals, adaptively detached from immediate bodily and environmental pressures (a view Seth credits to Daniel Dennett, 1984). It links to Ross Ashby's Law of Requisite Variety (1956): a controller must be able to take at least as many states as the environment perturbing it; "only variety can force down variety".
There is no single "volition" region. Following Patrick Haggard (UCL), the distributed network performs three processes:
- What. Selects one action out of many by integrating hierarchical beliefs, goals and values with perception (thirsty, like tea, right time of day, kettle in reach, no wine available). Concentrated in frontal cortex.
- When. Sets the timing, and is most closely tied to the subjective urge to move that Briony wondered about and Libet measured. It localises to readiness-potential regions: gentle stimulation of the supplementary motor area produces an urge to move without movement (Fried et al., 1991); stronger stimulation produces the movement too.
- Whether. A last-minute check that can cancel the action (out of milk): intentional inhibition, also in frontal regions.
These processes run in a continual loop through brain, body and world, with no beginning or end, funnelling many potential causes into a single flow of voluntary action. The experience of volition is the perception of this network's operation. And since action is self-fulfilling perceptual inference (Chapter 5), the experience of volition and the competence to control degrees of freedom are two sides of one prediction-machine coin: volition is a controlling hallucination.
What Volition Is For: Learning for Next Time
Experiences of volition also serve future behaviour. They flag voluntary acts so we attend to their consequences: try a shortcut on Monday, get lost and arrive late, take the reliable route on Tuesday. The feeling "I could have done otherwise" matters most here. It does not mean I actually could have; but circumstances are never identical twice (Heraclitus: no one steps in the same river twice), and at minimum my brain has changed by having had and attended to Monday's experience. The usefulness of feeling "I could have done otherwise" is that, next time, you might. The "you" is the bundle of self-related priors, values, goals, memories and best guesses, with volition one more strand.
Is Free Will an Illusion?
Daniel Wegner's The Illusion of Conscious Will (2002) set the tone. Seth's answer: it depends.
In what sense it is not real. Spooky free will is not real, and may not even qualify as an illusion: the phenomenology concerns actions that seem to come from within, not uncaused causes, so spooky free will is "an incoherent solution to a problem that doesn't exist" (Sam Harris: it makes no sense subjectively either). Often there is no felt volition at all: playing piano, making tea, being "in the moment" or in flow (Csikszentmihalyi), actions and thoughts "just happen".
In what sense it is real. With a reasonably intact brain and upbringing, we have a genuine capacity to execute and inhibit voluntary action. It is freedom from immediate bodily and worldly causes and from coercion by authorities, hypnotists, mesmerists or social-media pushers, and freedom to act on our beliefs and values. It is not freedom from the laws of nature. And the experience is as real as colour experience: a perceptual best guess, constrained by priors and data, that projects causal power onto the self as we project redness onto surfaces.
The capacity's reality shows in how it can be lost: anarchic hand syndrome (voluntary actions not felt as one's own), akinetic mutism (no voluntary action at all).
Tumours and responsibility. Charles Whitman, an engineering student, became the "Texas Tower Sniper"; a tumour was found pressing on his amygdala. A previously blameless teacher developed paedophilic urges that vanished when a tumour was removed and returned when it regrew (Burns and Swerdlow, 2003). If Whitman did not choose his tumour, is it "brain tumours all the way down" for everyone? Einstein, in a 1929 interview with George Sylvester Viereck (Saturday Evening Post, 26 October), said that since he did not believe in free will he "claimed credit for nothing".
Law requires both a guilty act and a guilty mind (mens rea). Bruce Waller argues moral responsibility is incoherent because we don't choose our brains; Dennett defends it (see the Dennett and Caruso debate, 2021). Seth leans toward a threshold view: past a certain competence to control our degrees of freedom, we can be held responsible.
Experiences of volition are real and indispensable. They are self-fulfilling perceptual inferences that bring about voluntary actions and let us learn from them. Knowing that we project causal power into them, to channel Wittgenstein, changes everything and leaves everything just the same.
The chapter closes by widening the frame: flexible voluntary behaviour seems distinctively human, but free will may come in degrees across species (Björn Brembs finds its traces in the spontaneous, variable escape responses of cockroaches). If freedom extends beyond us, so may consciousness. It's time to look beyond the human.
Critical notes. Seth's compatibilism is mainstream among philosophers but hard determinists (and incompatibilists generally) reply that redefining free will as a perceptual competence changes the subject: the folk notion many people care about includes could-have-done-otherwise in the strong sense Seth discards. Libet's timing method has its own well-known weaknesses (subjective clock reports are unreliable and shift with attention), which Seth does not discuss. The Whitman case is often told with more causal confidence than the evidence allows: the tumour (a glioblastoma) was found at autopsy and its role in the shootings remains uncertain. A small slip in the source's footnote: the legal term is actus reus, not "actus rea".
The readiness potential since 2021. Schurger's accumulator account has become a leading interpretation; Schurger, Hu, Pak and Roskies reviewed the debate in Trends in Cognitive Sciences (2021), concluding the RP is a poor marker of a decision to act. Maoz, Yaffe, Koch and Mudrik (eLife, 2019) showed the RP appears for arbitrary choices but is absent or much reduced for deliberate choices that matter, limiting what Libet-style studies can say about real decisions. On the philosophical side, Robert Sapolsky's Determined (2023) argued at length for hard determinism and against moral responsibility, reviving exactly the debate Seth sidesteps; Daniel Dennett, compatibilism's leading voice, died in 2024.
Degrees of freedom in agentic AI. Seth's framing transfers directly to governance of AI agents in banking. An agent's "what, when, whether" should be explicit and separable: what action (policy-constrained selection), when (scheduling, rate limits), and above all a whether stage, a last-minute veto or human-approval gate before irreversible actions such as payments or customer communications. Responsibility frameworks can mirror his threshold view: an agent is trusted to act autonomously only once its competence to control its degrees of freedom has been demonstrated in a given scope, and Ashby's requisite variety warns that a controller simpler than the environment it faces will fail on the cases it cannot represent.
Beyond Human
Part IV, Other, opens by asking how far the circle of consciousness extends. Seth's rules: language and metacognition are not tests, beware anthropomorphism and anthropocentrism, and above all do not confuse consciousness with intelligence. On the strength of shared brain mechanisms, he is confident all mammals are conscious; birds make a strong case; the octopus offers a distributed, alien consciousness without human-style body ownership; further out, responses to injury are the best clue, and somewhere (perhaps by the 302-neuron nematode) the lights go out. Consciousness tracks being alive more than being smart, and intelligence can exist without consciousness at all.
Pigs on Trial
From the early ninth century to the mid-1700s, European ecclesiastical courts prosecuted animals. E. P. Evans' 1906 history documents almost two hundred cases: pigs (the commonest defendants, since they roamed medieval villages) were hanged or burned for crimes from eating children to consuming consecrated biscuits, sometimes charged with abetting by grunting, occasionally acquitted; bulls, horses, eels, dogs and once dolphins also stood trial.
Smaller pests were harder to try. In a celebrated sixteenth-century case the French lawyer Bartholomew Chassenée won acquittal for rats by arguing they could not be expected to attend court given the cats lying in wait on the route; weevils were served written orders to leave a barley crop by a given day and hour. The legal fiction was absurd, but it carried a recognition that animals have inner universes and can in some sense decide, the opposite of Descartes' animal-robots (which became dominant only as the Enlightenment spread). Today it would be strange to claim only humans are conscious; the question is how far the circle extends and how different other inner worlds are.
Ground Rules for Looking
- No report needed. Absence of language is not evidence of absence of consciousness; nor is absence of metacognition (reflecting on one's own thoughts), though higher-order theorists would disagree.
- Anthropomorphism. Attributing humanlike qualities to non-humans: seeing human consciousness where it may not be (the dog that "understands what we're thinking"). Darwin's emotions-across-species stance leaned this way, feeding basic-emotion research (Panksepp, Paul Ekman).
- Anthropocentrism. Interpreting the world in human terms: missing non-humanlike consciousness where it is, as in Descartes' beast machines.
- Consciousness is not intelligence. Using intelligence as a litmus test commits both errors and breeds methodological laziness, since language and metacognition are easier to assess than consciousness itself.
- But intelligence matters. It opens new kinds of experience: sadness needs little cognition, regret needs the ability to consider alternatives. A 2014 study (Steiner and Redish) found rats looked back at the unchosen option after a disappointing choice, a possible rodent regret, though what they felt is unclear.
- Start from humans anyway. We know we are conscious and increasingly know the mechanisms, so humans are the firm shore from which to extrapolate.
The beast machine view implies consciousness is more widespread than an intelligence criterion suggests, though not that wherever there is life there is consciousness (Ernst Haeckel's 1892 "biopsychism", distinct from panpsychism). Seth's image: stepping onto an icy lake from a frozen shore, one careful step at a time.
Mammals: Confident on Shared Mechanisms
Seth believes all mammals are conscious, on grounds of shared mechanisms, not superficial similarity. Set aside raw brain size (mostly a function of body size; even the encephalisation quotient says little about consciousness) and mammalian brains are strikingly alike.
Seventeen testable properties (Seth, Baars and Edelman, 2005). With cognitive scientist Bernard Baars and animal-cognition expert David Edelman (Gerald's son), Seth listed seventeen properties of human consciousness testable in other mammals, an arbitrary number but proof that animal consciousness can be studied experimentally. - Anatomy. All mammals have a six-layered cortex, a thalamus densely interconnected with it, a deep brainstem, and shared neurotransmitter systems. - Dynamics of conscious level. Waking brains show irregular, low-amplitude, fast activity; sleep switches all mammals to regular, large-amplitude dynamics, as in humans. - Anaesthesia. Produces a similar breakdown of communication between brain regions with behavioural unresponsiveness across mammals.
- Seals and dolphins sleep with half the brain at a time.
- Koalas sleep about 22 hours a day; giraffes under 4.
- Newborn killer whales do not sleep at all for their first month.
- Nearly all mammals have REM sleep, though seals only on land and dolphins apparently not at all.
Contents will differ hugely with dominant senses: whiskers for mice, echolocation for bats, smell for naked mole rats. Each lives in its own Umwelt (Jakob von Uexküll's term).
Selfhood and the Mirror Test
Human mirror self-recognition emerges between 18 and 24 months, a marker of high-level self-consciousness, not of consciousness itself. Gordon Gallup Jr's test (1970): anaesthetise an animal, mark it where it cannot normally see, then give it a mirror; if it investigates the mark on its own body rather than the image, it passes.
- Passes among mammals: some great apes (one paper: "Another gorilla (Gorilla gorilla gorilla) recognizes himself in a mirror"), a few dolphins and killer whales, and a single elephant.
- Fails so far: pandas, dogs, various monkeys. (Rhesus monkeys passed only after weeks of training, Chang et al., 2017, which is not spontaneous use.)
- Non-mammals: no convincing pass; manta rays and magpies may come close; the cleaner wrasse is disputed.
Failure can have many causes (disliking mirrors, not understanding them, avoiding eye contact), so researchers adapt tests to each perceptual world, such as olfactory mirrors for dogs (who still do poorly; the field is called "dognition"). Even so, the short pass list suggests dramatic differences in how mammals experience being themselves.
Monkeys: Conscious, Not Furry Little People
Monkeys are standard primate models, especially in vision, and can be trained to "report" by lever whether they saw something, a primate version of a core human method (Cowey and Stoerig, 1995). For Seth there is no doubt they have conscious selfhood of some kind.
Cayo Santiago and the cucumber protest. In July 2017 Seth spent a day with Yale psychologist Laurie Santos (and a film crew) on Cayo Santiago, "monkey island" off eastern Puerto Rico, home to more than a thousand rhesus macaques transplanted from Kolkata in 1938 by zoologist Clarence Ray Carpenter, tired of travelling to India. Two monkeys took turns leaping from a tree into a pond for the sheer fun of it. (Hurricane Maria hit soon after; most monkeys survived, much infrastructure did not; the footage is in the 2018 documentary The Most Unknown.) In Frans de Waal's famous video, a capuchin paid in cucumber for passing a stone watches her neighbour get a grape for the same task, then throws her cucumber back and rattles the cage.
Fun and tantrums are powerful intuition pumps: we infer not just a conscious being but one like us. Yet monkeys consistently fail the mirror test. They are conscious, with some selfhood, but "they are not furry little people".
The Octopus: The Most Other Mind
In summer 2009, Seth and David Edelman spent a week with about a dozen Octopus vulgaris in Graziano Fiorito's lab at the Stazione Zoologica, a dank basement beneath a public aquarium in Naples. Matching one animal's changing appearance against Fiorito's A Catalogue of Body Patterning in Cephalopoda (heavy enough to break his Ryanair baggage allowance), Seth heard a splat and a slither: a lid left ajar, an escape, and his lasting conviction that the octopus had lulled him into false security. Another day, as Edelman lowered shapes (some paired with a crab) into one tank, every octopus across the walkway was pressed to its glass, watching, apparently out of sheer interest, though octopuses are solitary and even cannibalistic.
- Last common ancestor of humans and octopuses: about 600 million years ago, perhaps a flattened worm.
- O. vulgaris: eight arms, three hearts, blue blood, ink, jet propulsion; soft-bodied except for a bony beak.
- About 500 million neurons, roughly six times a mouse; about three fifths in the arms.
- Central brain of about forty distinct lobes, with no myelin.
- Nematode worm: 302 neurons, where Seth finds conscious status hard to ascribe.
Octopus minds are an independent evolutionary experiment, "as close to the mind of an alien as we are likely to encounter" (Seth chose to write about octopuses for a 2015 book chapter on alien consciousness; the film Arrival draws the same parallel). Peter Godfrey-Smith: the minds of cephalopods are "the most other of all". Their nervous system is more distributed and less integrated than a mammal's, so octopus consciousness, if any, may lack a single centre. Even their genetics is odd: in 2017 researchers found octopuses and some other cephalopods edit their RNA extensively before translation, much of it in nervous-system genes, as if rewriting parts of their genome on the fly, possibly underpinning their cognition.
Their abilities: retrieving crabs from nested Plexiglas cubes, solving mazes, trying varied actions on a problem, learning by observing other octopuses (Fiorito and Scotto, 1992), and, in Blue Planet II, an octopus cloaking itself in shells to hide from a shark (see also My Octopus Teacher, 2020).
Camouflage Without Colour Vision
Octopuses match colour, shape and texture so well they are invisible from a metre or two, using chromatophores, elastic skin sacs producing red, yellow or brown, opened by commands from chromatophore lobes. They must hide from predators' visual systems, so camouflage implicitly encodes what predators see. Yet octopuses are colour-blind: their eyes have one photopigment (they sense polarisation), and light-sensitive cells in their skin are colour-blind too. Chromatophore control is thought to be open-loop, with no internal copy of the commands: the central brain may not know what its skin is doing.
They also taste with their suckers, another sign of decentralisation. For body ownership, the challenge is formidable: eight flexible arms furling in all directions, with control devolved; a severed arm can still grasp food for some time. So octopuses may not bother with unified body ownership at all. They still distinguish self from other: suckers grip almost anything except the same octopus's arms and body, thanks to a skin chemical suckers detect. In admittedly macabre experiments (Nesher et al., 2014), detached arms gripped skinned arms but never intact ones (arms regrow, and octopuses seem to notice little). An octopus may know what is itself without knowing where its body is, perceive its body only hazily (without experiencing the haze), and there may even be something it is like to be an octopus arm.
Further Out: Birds, Fish, Insects, Worms
Birds make a strong case. Their brains differ but map closely onto mammalian cortex and thalamus (the pallium), though they lack a corpus callosum, possibly a "natural split brain". Many are clever (parrots count, cockatoos dance, scrub jays cache food for future needs), but non-dancing, non-speaking birds are likely conscious too.
Further out, Seth swaps similarity for his beast machine logic: since conscious perception is rooted in physiological regulation and integrity, look at responses to injury. This is also ethically right: welfare should rest on capacity for pain and suffering, not similarity to us or a cognitive threshold (Jonathan Birch's precautionary "benefit of the doubt"; the EU brought cephalopods under animal-welfare law in 2010, directive 2010/63/EU).
| Group | Key evidence in the chapter | Seth's assessment |
|---|---|---|
| Mammals | Shared cortex-thalamus anatomy, sleep and anaesthesia signatures | All conscious (confident) |
| Birds | Pallium maps onto cortex and thalamus; rich cognition | Strong case |
| Octopus | Huge distributed nervous system, learning, observation | Likely conscious, very differently |
| Fish | Injured zebrafish pay a cost (leave an enriched tank for a barren, brightly lit one laced with analgesic) | Suggestive |
| Insects | Ants don't limp, but insects have opiate-like systems; injured Drosophila show chronic-pain-like hypersensitivity (2019) | Suggestive, unclear |
| All animals | Anaesthetics work from single cells to primates | Not conclusive |
| Nematode (302 neurons), paramecium | Minimal nervous system or none | Probably outside the circle |
"All of this is suggestive, none of it is conclusive." Seth's intuition, "no more than an intuition", is that some animals are not conscious at all, partly because even in mammals unconsciousness is easy to achieve.
- Perceptual richness. How fine-grained experience is in each sense.
- Evaluative richness. How varied the valenced (good/bad) states are.
- Unity. How integrated experience is at a time (an octopus or a bird may score low).
- Temporality. How experience is integrated across time.
- Selfhood. How far the animal experiences itself as a distinct self.
Two Benefits, and a Stronger Claim
Studying animal consciousness gives recognition that ours is not the only way to experience world and self (we occupy a tiny region of a vast space of possible minds, explored with "a few flares cast into the darkness") and humility, with renewed motivation to minimise suffering wherever it appears.
Consciousness can exist without much intelligence (you don't have to be smart to suffer), and intelligence can exist without consciousness. That second possibility leads to the final question: machines.
Animal sentience since 2021. The LSE review led by Jonathan Birch (2021) for the UK government found strong evidence of sentience in cephalopod molluscs and decapod crustaceans, and the Animal Welfare (Sentience) Act 2022 recognised them as sentient. In April 2024 the New York Declaration on Animal Consciousness, launched at NYU and signed by hundreds of scientists and philosophers, stated that there is strong scientific support for conscious experience in mammals and birds and at least a realistic possibility in all vertebrates (including reptiles, amphibians and fish) and many invertebrates (at minimum insects, decapod crustaceans and cephalopods), and that ignoring this possibility in welfare decisions is irresponsible. Birch's The Edge of Sentience (2024) develops the precautionary framework. Behavioural work has also pushed outward, for example studies suggesting bumblebees engage in play-like behaviour (2022), and further cleaner wrasse studies by Kohda and colleagues claiming mirror self-recognition via self-face recognition, still debated.
Critical notes. The "single Eurasian elephant" that passed the mirror test was an Asian elephant, "Happy", at the Bronx Zoo (Plotnik, de Waal and Reiss, 2006). Seth's split between confident (mammals) and intuitive (outer circle) judgments is honest, but his assessment still leans on similarity of mechanism to mammals, which may itself be anthropocentric for animals like insects. The injury-response criterion faces the problem that nociception can be unconscious (even humans withdraw from pain reflexively), so behavioural tests need flexibility and trade-offs, as in the zebrafish design, not mere withdrawal. Numbers for octopus neurons vary by source (often about two thirds in the arms). The beast machine view predicts consciousness is tied to life; it is not yet clear how it would decide hard cases such as insects any better than rival criteria.
The octopus arm's point of view. Suppose a severed arm, still grasping food for minutes, were briefly conscious. It would have touch, taste through its suckers, and a local motor model, but no eyes, no central self-model and no narrative. If there is something it is like to be that arm, then consciousness does not require a unified self at all, and Seth's five levels of selfhood (Chapter 8) come apart into separable modules distributed across one animal. The thought experiment shows why the octopus is such a hard case: it tests whether consciousness and selfhood scale together, or whether a single creature might host several partial perspectives.
Intelligence is not the marker. For AI governance the chapter supplies a clean principle: capability benchmarks do not measure moral status or experience. Enterprise discussions that slide from "the model reasons like an expert" to "the model understands or feels" commit the anthropomorphism Seth warns against; conversely, dismissing all concern because systems are "just statistics" can be anthropocentric if criteria are never stated. The defensible stance is Birch's: state the criteria, assess evidence against them, and apply proportionate precaution.
Machine Minds
The fear of conscious AI is "the myth of the golem made silicon", and it rests on two unjustified assumptions: that functionalism is true, and that consciousness comes along for the ride with intelligence. Seth's view: making computers smarter will not make them sentient; even a "silicon beast machine" built on predictive principles is probably not conscious, because self-maintenance in living systems goes all the way down to cells, and "it is life, rather than information processing, that breathes the fire into the equations". The urgent problems are machines that seem conscious (the Garland test), the ethics of possibly creating artificial suffering, the open question of cerebral organoids, and the techno-rapture that drives the whole fantasy.
The Golem and Its Descendants
In late-sixteenth-century Prague, Rabbi Judah Loew ben Bezalel shaped a golem, Josef ("Yoselle"), from Vltava clay to defend his people from pogroms. Once animated by incantation it could move, was aware, and obeyed, until it turned to violent monstering; the rabbi revoked the spell and it fell to pieces in the synagogue grounds, its remains perhaps still hidden somewhere in Prague. The lesson is hubris, and fiction repeats it: Frankenstein's monster, Čapek's robots, the Terminator, Blade Runner's replicants, Kubrick's HAL, and Ava in Alex Garland's Ex Machina. (Norbert Wiener's God and Golem, Inc., 1964, used golems to think about AI risk.)
AI is now in phones, fridges and cars, often powered by brain-inspired neural networks, and the worries are real: jobs, social fabric, even extinction, whether through AI self-interest or a lack of programming foresight (Bostrom's paperclip maximiser, which turns the Earth into paperclips). Beneath the most apocalyptic of these runs the assumption that AI will at some point become conscious.
Two Unjustified Assumptions
- Necessary condition: functionalism. Consciousness depends on what a system does, not what it is made of (neurons, silicon, Vltava clay). It bundles two claims: substrate independence and sufficiency of input-output relations, which usually but not always go together. It is popular, even a default, but there are no knock-down arguments either way; Seth's stance is "suspicious agnosticism".
- Sufficient condition: intelligence. Even if functionalism is true, information processing alone is not enough; the further assumption is that the processing that yields intelligence is the processing that yields consciousness. This is the anthropocentrism of the previous chapter: we are conscious and intelligent and "species-proud", so we assume the two are linked. The risk is attributing consciousness too eagerly to clever machines and too stingily to animals.
Other forces amplify the myth. Ray Kurzweil's "Singularity", motivated by exponential growth in computing: on an exponential curve, as the COVID-19 pandemic taught, wherever you stand the road ahead looks impossibly steep and the road behind irrelevantly flat, so the local view tells you nothing. Promethean fears packaged and sold back by science fiction. And sloppy use of "consciousness": some AI researchers call anything that responds, learns or maximises a reward conscious, a "nonsensical overextension".
Seth's alternative picture: consciousness is not determined by intelligence, intelligence can exist without consciousness, and each has many dimensions (echoing Birch's consciousness profiles and the multidimensional levels of Bayne, Hohwy and Owen). He placed current AI low even on intelligence: "sophisticated machine-based pattern recognition, perhaps spiced up with a bit of planning". Neither artificial general intelligence nor post-Singularity AI warrants assuming consciousness, and many non-humanlike intelligences may complement ours without it. Some forms of intelligence might need consciousness, and all conscious creatures may be a little intelligent (consciousness and intelligence are not quite doubly dissociable), but intelligence is not the royal road to consciousness.
What Theories Say a Conscious Machine Needs
If not intelligence, what? It depends on which theory of consciousness you hold.
| Theory | What would make a machine conscious | Verdict on machines |
|---|---|---|
| Global workspace plus higher-order (Dehaene, Lau and Kouider, Science, 2017) | "Global availability" of information and "self-monitoring" of performance | Achievable in computers; the authors equivocate on real versus as-if consciousness and admit leaving out the experiential part ("too much", says Seth) |
| Integrated information theory (Tononi, Koch) | Non-zero Φ from the system's cause-effect structure, whatever its material | Any Φ-generating machine is somewhat conscious; but feedforward networks generate no Φ, so a behaviourally perfect one is a "behavioural zombie" (IIT accepts substrate independence but rejects input-output sufficiency) |
| Beast machine (Seth) | Rooted in living self-maintenance, all the way down | Probably not, even for a predictive, "interoceptive" robot |
Both GWT/HOT and IIT avoid equating consciousness with intelligence, but each licenses conscious machines only if you accept the theory.
The Silicon Beast Machine
A near-future robot. Give it a silicon brain and a humanlike body with rich sensors and effectors, controlled by a network built on predictive processing and active inference. It runs a generative model of its world and body, makes Bayesian best guesses, and its controlled and controlling hallucinations keep it in an optimal functional state, "alive" by its own lights. It has artificial interoception (battery levels, actuator and synthetic-muscle integrity) and control-oriented guesses about these produce synthetic emotions that guide behaviour. It acts autonomously, doing the right thing at the right time. Its mechanisms map onto exactly the machinery Seth says underlies human embodiment and selfhood. Is it conscious?
Seth's answer: "I don't know for sure, but probably not." Consciousness, on his theory, arose in evolution, emerges in development and operates moment to moment in ways bound up with being self-sustaining living machines that care about their own persistence. His intuition, "only an intuition", is that the materiality of life will matter for all consciousness. The reason: in living systems regulation and self-maintenance are not confined to one level such as the whole body. Every cell continually regenerates the conditions of its own integrity, which no current or near-future computer does, nor would the silicon beast machine. This links to autopoiesis ("self-creation"), Chilean biologist Humberto Maturana's concept of a system that produces the components of its own continued existence (Seth met him in Santiago in January 2019; Maturana died in May 2021, aged 92). Cells are not thereby conscious; rather, the regulation underpinning consciousness is bootstrapped from life processes that go all the way down.
On this view, it is life, rather than information processing, that breathes the fire into the equations.
The Garland Test: Seeming Conscious
Real machine consciousness may be distant or impossible, but machines that seem conscious are coming soon.
From Turing to Garland. In Ex Machina (2014), billionaire Nathan invites programmer Caleb to judge whether his robot Ava is conscious. In the standard Turing test, a judge exchanges typed messages with a hidden human and a machine; the machine passes if the judge can't reliably tell which is which. Nathan's version: "the challenge is to show you that she's a robot, and see if you still feel she has consciousness." This turns a test of intelligence into a test of consciousness, and reveals it as a test of the human, not the machine: what matters is whether Ava makes a conscious person feel she is conscious. The "Garland test" was coined by Murray Shanahan, whose Embodiment and the Inner Life (2010) inspired the film.
Gullible judges. In 2014 the chatbot "Eugene Goostman", posing as a thirteen-year-old Ukrainian boy with poor English, fooled ten of thirty judges (one third) and was hailed as passing the Turing test. Impersonating a foreign teenager is far easier than impersonating a peer; on winning it said: "I feel about beating the turing test in quite convenient way." It was a test of human gullibility, and the humans failed.
In May 2020 OpenAI released GPT-3 ("Generative Pre-trained Transformer"), trained to predict the next word, with 175 billion parameters; it writes fluent prose in many styles, including a Guardian essay (8 September 2020) on why humans should not fear AI, with the line that AI "should not waste time trying to understand the viewpoints of people who distrust artificial intelligence for a living". Seth, with Gary Marcus, holds that it does not understand what it writes and thought a sophisticated interlocutor could still catch it out, conceding this "may not be true for GPT-4, or GPT-10". Even a GPT that aces the Turing test would show only narrow, disembodied linguistic intelligence, not the embodied "right thing at the right time" intelligence of animals. The Garland test, he wrote, "remains pristine."
Geminoids and the Uncanny Valley
Hiroshi Ishiguro builds Geminoids, robots as humanlike as possible: copies of himself, his six-year-old daughter, and a Japanese-European TV anchorwoman blended from about thirty people, made from 3D body scans with pneumatic actuators for expressions. They contain little AI; they are about mimicry and telepresence (Ishiguro once delivered a 45-minute remote lecture to 150 students through his). Meeting one is the opposite of meeting a cat: a cat or octopus gives an immediate sense of another conscious presence despite its different look; the Geminoid's near-likeness accentuates otherness. In a 2009 study, visitors' commonest feeling was fear.
This is Masahiro Mori's "uncanny valley" (1970): affinity rises as robots grow more humanlike (C-3PO) until, strikingly human in some ways but short in others, they provoke revulsion and fear, recovering only near indistinguishability. Physical robots remain trapped; virtual humans are climbing out. Generative adversarial networks produce photorealistic faces of people who never existed (via thispersondoesnotexist.com, using techniques akin to the hallucination machine), deepfakes animate them (Obama, 2017; Tom Cruise on TikTok, 2021), and language models give them speech. Seth predicts flatly that a video-enhanced Turing test will be passed; doubting it shows human exceptionalism, a failure of imagination, or both. The open questions are whether such agents will cross the uncanny valley into the physical world, and whether the Garland test will fall, and what feeling that machines are conscious, while knowing they are code, would do to us.
Ethics: From Seeming to Being
- Economic and social disruption. Self-driving cars and automated factories (though, as colleague Ryota Kanai put it, "a horse is basically a self-driving horse").
- Opaque, biased delegation. Decisions handed to systems whose workings are opaque even to their designers; at the extreme, AI in charge of nuclear weapons or the internet backbone.
- Psychological and social manipulation. Deepfake privacy invasion, behaviour modification by predictive algorithms, filter bubbles and echo chambers: identity and autonomy ceded to "faceless data-corporations" in a vast uncontrolled global experiment (see Stuart Russell, Human Compatible; Nina Schick on deepfakes).
- Machines that pass the Garland test. We may learn to separate feeling from action (caring for humans but not robots both felt to be conscious), with unknown psychological cost; Westworld-style abuse of robots felt conscious but known not to be risks mind-fracturing sociopathy; or empathy distorted by similarity, caring more for Geminoid twins than for other humans or animals. Dennett: we are building "intelligent tools, not colleagues".
- Real machine consciousness. Conscious status brings moral status. We would have to minimise its suffering (and we do poorly with animals), possibly facing forms of suffering with no human equivalent, or minds where positive and negative valence do not even apply: we would not know what the ethical issues were.
We do not know what it would take to create a conscious machine, but neither do we know what it would not take. In June 2019, at a Leverhulme Centre for the Future of Intelligence meeting in Cambridge where Seth also spoke, Thomas Metzinger called for an immediate thirty-year moratorium on research aiming at "synthetic phenomenology". Hard to follow literally (much computational psychology could fall under it), but the thrust is right: the best ethics is preventative ethics.
A historical parallel: in the heyday of élan vital-style vitalism, the ethics of artificial life would have seemed absurd; a century later we have CRISPR (Nobel Prize in Chemistry 2020 for Emmanuelle Charpentier and Jennifer Doudna) and, in 2019, Jason Chin's Cambridge lab built E. coli with a fully synthetic genome.
Cerebral Organoids
Biotechnology may bring us closer to synthetic consciousness than AI. Cerebral organoids are tiny brain-like structures of real neurons grown from human pluripotent stem cells. They are not "mini brains", but resemble the developing brain enough to model developmental disorders. One study (Trujillo et al., 2019) found coordinated waves of electrical activity resembling those of premature babies.
Reasons to worry. Organoids are made of the same stuff as brains, removing one obstacle, and they can be wired to cameras and robotic arms. The scale is the ethical multiplier: Alysson Muotri, "We want to make farms of these organoids."
Reasons for doubt. They are extremely simple, completely disembodied, and do not interact with the world. Seth (with Tim Bayne and Marcello Massimini, 2020) judges current organoids highly unlikely to be conscious, but the question will stay disconcertingly open as the technology advances. In summer 2020 he spoke to a US National Academies committee on regulating organoid and chimera research.
Techno-Rapture
Why is machine consciousness so alluring? Seth diagnoses techno-rapture: a wish to escape messy, mortal biology by uploading minds into ageless circuitry, the favourite trope of transhumanists "for whom one life is not enough". Nick Bostrom's simulation argument (2003) holds that if far-future civilisations run many ancestor simulations, we are statistically more likely to be simulated than original; Seth objects that it assumes functionalism (that simulation equals instantiation). Add a fast-approaching Singularity, and the technological elite can cast themselves as the decisive actors in history, with immortality the prize: "human exceptionalism gone properly off the rails", a symptom of alienation from our biological nature.
The beast machine view reverses the techno-rapture. Human experience arises because of, not in spite of, our being self-sustaining organisms that care about their own persistence. It does not exclude conscious machines, but it undercuts the narrative of soon-to-be-sentient computers. Understanding consciousness places us increasingly within nature, not further apart from it.
AI consciousness after ChatGPT. The book went to press before large language models went mainstream. ChatGPT (November 2022) and GPT-4 (March 2023) overtook Seth's GPT-3 benchmark; in a 2025 UC San Diego study (Jones and Bergen), GPT-4.5 prompted to adopt a persona was judged to be the human about 73 per cent of the time in a three-party Turing test, so the text Turing test has effectively fallen. The Garland test is under pressure too: in June 2022 Google engineer Blake Lemoine publicly claimed the LaMDA chatbot was sentient and was later dismissed. David Chalmers (2022-23) put the chance that then-current LLMs were conscious at well under 10 per cent, higher for extended successors. In August 2023, Patrick Butlin, Robert Long and 17 co-authors (including Yoshua Bengio, Jonathan Birch, Stephen Fleming and Chris Frith) published Consciousness in Artificial Intelligence, which assumed computational functionalism for the sake of argument and derived 14 "indicator properties" from recurrent processing, global workspace, higher-order, attention schema and predictive processing theories, plus agency and embodiment. They concluded no current AI system is a strong candidate, but that no obvious technical barrier prevents building systems that satisfy the indicators. "Taking AI Welfare Seriously" (Long, Sebo et al., 2024) argued that AI companies should begin preparing for possible AI moral patienthood, and at least one major lab (Anthropic) launched a model welfare research programme in 2025. Seth has pushed the other way: his 2024-2025 work, notably the Behavioral and Brain Sciences target article "Conscious artificial intelligence and biological naturalism", now openly adopts biological naturalism, arguing that computational functionalism is doubtful, that features of living, autopoietic systems (multiscale self-maintenance, no clean separation of "software" from "hardware") may be required, and that the bigger near-term danger is AI that seems conscious and exploits our anthropomorphism. Metzinger's moratorium appeared in print as "Artificial Suffering" (2021), proposing it run to 2050.
Organoid intelligence. In 2022, Cortical Labs' "DishBrain" (Kagan et al., Neuron) showed cultured neurons on a multi-electrode array learning to play a simplified Pong through closed-loop feedback, explicitly framed in free energy principle terms. In February 2023, Smirnova, Hartung and colleagues (Johns Hopkins) launched "organoid intelligence" as a field in Frontiers in Science, proposing biocomputing with brain organoids and calling for embedded ethics from the outset; a 2023 Nature Electronics study ("Brainoware") used an organoid as a reservoir computer for simple speech recognition. These developments bring together exactly the ingredients Seth flagged: living neurons, closed-loop interaction with an environment, and intended scale. Consciousness in such systems remains highly unlikely on current evidence, but the ethical case for precautionary oversight has strengthened.
Critical notes. Seth's core claim is an intuition, as he says, and his opponents' functionalism is also unproven; the debate is a standoff of priors, which is why the Butlin report adopted functionalism as a working assumption while Seth adopts the opposite. Biological naturalism faces the "which property?" challenge: if not computation, what exactly about metabolism or autopoiesis matters, and why could it not in principle be engineered? Some predictions have aged quickly: "current AI" sits low on the intelligence axis of his figure, and the claim that a GPT could be caught out by a sophisticated interlocutor no longer holds for text. Small slips: the "45 terabytes" figure for GPT-3 refers to raw Common Crawl text before filtering (the training set was far smaller after filtering); and the book's footnote garbles Turing's original imitation game, in which a man and a woman are questioned and the machine takes the man's role, trying to pass as the woman. "Chimeras" are organisms containing cells from another organism, not simply animals genetically modified to express human traits.
Design for the Garland test, not against it. For a bank deploying an RM copilot or customer-facing assistant, Seth's analysis gives concrete guardrails. Do not design for perceived sentience: avoid persona features that invite users to attribute feelings (first-person claims of emotion, simulated distress), since perceived consciousness is a lever on trust that regulators will treat as manipulation. Disclose machine status clearly, as AI-transparency rules increasingly require. Treat intelligence and consciousness as separate risk axes: capability drives operational and conduct risk; seeming-consciousness drives customer-vulnerability and dependency risk. And keep humans in decisions that need care about persistence and consequence, which on Seth's view no current system has.
Epilogue
Watching a surgeon disconnect half of a child's brain, Seth feels the hard problem's full force again, then answers it with the real problem: explain, predict and control the properties of experience rather than its bare existence. He recaps the book's three post-Copernican challenges (perception as controlled hallucination, the self as perception, and prediction rooted in staying alive) and argues that the science of consciousness, like Copernicus and Darwin before it, displaces us from the centre but gives back more than it takes.
Looking Into the Mechanics of a Self
In January 2019, some twenty years into his research and three after the anaesthesia that opened the book, Seth saw a living human brain for the first time. His host was Michael Carter, a paediatric neurosurgeon at the Bristol Royal Hospital for Children, operating on a boy of just over six with severe epilepsy since birth, originating in a right hemisphere damaged during premature birth; all medications had failed. (The old joke came to mind: a brain transplant is the only operation where it is better to be the donor.)
The hemispherotomy. The surgeon enters from the right, removes (resects) the temporal lobe, and cuts all white-matter tracts linking the right hemisphere to the rest of brain and body. The hemisphere stays in the skull with its blood supply: a living but isolated island of cortex, an extreme version of split-brain surgery that stops seizures spreading. Done early, the young brain adapts and the remaining hemisphere takes up most or all of the slack; outcomes are generally good. The operation began at about noon and lasted just over eight hours, Carter working without pause (Seth admits he struggles to last five minutes before checking email or the cricket score). Invited to the microscope midway, Seth found the crisp cortical hierarchies and top-down/bottom-up flows of his models nowhere to be seen: the brain was "newly inscrutable", the view almost transgressive, "looking directly into the mechanics of a human self".
Could that disconnected but living hemisphere sustain its own isolated consciousness? Seth, Tim Bayne and Marcello Massimini (2020) analysed such potential "islands of awareness", alongside ex cranio reanimated pig brains and cerebral organoids.
The Real Problem, Restated
Driving home through the winter dark, Seth recalls Chalmers' hard problem: we agree experience arises from a physical basis, but why and how it does "seems objectively unreasonable", and yet it does. Philosophy offers a set menu from panpsychism (consciousness more or less everywhere) to eliminative materialism (no consciousness as we conceive it). Consciousness science is instead "cooking with whatever you can find in the fridge": philosophy, neuroscience, psychology, computer science, psychiatry and machine learning recombined into something new.
- Accept that consciousness exists, then ask how its phenomenological properties relate to properties of brains in bodies in worlds.
- Start from correlations, but don't stop there: build sturdy explanatory bridges so that links between mechanism and phenomenology make sense, meaning they explain, predict and control.
- Follow the precedent of life: science beat vitalism by splitting life into properties and explaining each mechanistically. The aim is to dissolve, not solve, the hard problem.
The Journey in Three Challenges
| Stage | How things seem | How they are, on Seth's account |
|---|---|---|
| Conscious level | Awake versus not is a simple switch | Measurable through causal density and integrated information, which track experience being unified and distinct |
| Perception | A passive window on an objective world | An active, action-oriented construction, less and more than reality, designed for utility, not accuracy |
| Self | The thing that does the perceiving | Another controlled hallucination, from personal identity down to the inchoate feeling of being a living body |
| Function | Perception exists to represent | Prediction exists for control and regulation of the body's physiological condition |
| Self versus world | Utterly different kinds of experience | The same predictive principles; finding-out predictions give objects, control predictions give feelings |
This is the beast machine theory, a twenty-first-century version, or inversion, of La Mettrie's l'homme machine. Tying mind to physiology gives new substance to the continuity of life and mind, buttressed by predictive processing and the free energy principle, and brings us closer to other animals and further from "the fleshless calculus of AI".
As consciousness and life come together, consciousness and intelligence are teased apart.
The Third Displacement
- Copernicus. Removed Earth from the centre and gave us a universe, since expanded beyond imagination.
- Darwin. Removed humans from special creation and gave us a family: kinship with all species and deep time.
- Consciousness science. Breaches "the last remaining bastion of human exceptionalism", the presumed specialness of conscious minds, and shows them inscribed in the wider patterns of nature.
Everything in experience is a perception of sorts, and every perception a controlled or controlling hallucination: free will, the flow of time, perhaps even three-dimensional structure and the felt objective reality of what we perceive. The tools of consciousness science edge closer to Kant's noumenon, the unknowable reality of which we are part. All of this is testable, and every step erodes the intuition that consciousness is "one big scary mystery in search of one big scary solution".
The practical payoffs are already arriving:
- Consciousness meters. Theory-based measures of conscious level detect "covert consciousness" in behaviourally unresponsive patients.
- Psychiatry. Predictive models of hallucinations and delusions point toward treating causes, not symptoms.
- Technology. New directions for AI, brain-machine interfaces and virtual reality.
A Little Mystery
The search remains personal: a science of consciousness must illuminate our own inner lives and those of others. The real problem's promise is to show our inner universe as part of, not apart from, nature, and to offer a new peace with what happens when the controlled hallucination of being you "finally breaks down into nothingness": not an anaesthetic interruption, but a return to the eternity each of us emerged from. Seth's last line: when life in the first person ends, "perhaps it's not so bad if a little mystery remains".
The island hemisphere. After the operation, the boy's right hemisphere lies in his skull, perfused and alive but cut off from every input and output. Suppose it still showed complex, integrated activity. IIT would say it could be conscious; a global workspace theorist would ask whether it broadcasts anything; Seth's beast machine view adds that without any body to regulate or any prediction errors to receive, its "experience", if any, would lack the control-oriented grounding of every ordinary experience. The thought experiment shows how sharply theories diverge on islands of awareness, and why they are a natural target for future tests.
Meters and islands since 2021. The "consciousness meters" of the epilogue have moved toward the clinic: the 2024 finding that about one in four behaviourally unresponsive patients shows cognitive motor dissociation (Chapter 2) has made brain-based testing part of recommended assessment. The ethics of "islands" has sharpened too: Yale's BrainEx (2019) and OrganEx (2022) restored some cellular and molecular function in pig brains and organs hours after death (with no sign of organised, consciousness-like activity), keeping the questions raised by Bayne, Seth and Massimini live, while organoid research (Chapter 13) supplies the other half of the problem.
Critical notes. The epilogue's confidence that the hard problem is being dissolved is a programme rather than a result: the real-problem approach explains structure and content, and critics hold that the existence of experience is untouched. The Copernicus-Darwin framing (borrowed in spirit from Freud's "three blows" to human narcissism) is rhetorically strong but also self-flattering for the field. And the beast machine theory's distinctive empirical claims, especially about interoceptive inference and the life-consciousness link, remain thinly tested compared with its unifying scope.
Master Takeaways
Being You adds up to a single inversion of how things seem: we do not passively receive the world and then act; we predict it, from the inside out, in order to stay alive. Perception, selfhood, emotion and free will all become forms of controlled (and controlling) hallucination, and consciousness lands closer to life and further from intelligence than either common sense or AI culture assumes.
The takeaways
- Swap the hard problem for the real problem. Instead of asking why experience exists at all, explain, predict and control its properties (level, content, self) through brain and body mechanisms; the hope, as with life and vitalism, is that the "why" question dissolves as the "how" questions are answered.
- Perception is controlled hallucination. The brain never meets the world directly; it predicts the causes of its sensory signals and uses prediction errors to keep those guesses honest, and what we experience is the prediction, not the signal. Hallucination is simply perception that has lost its control.
- The self is another perception. There is no inner pilot; the experience of being you is a bundle of predictions across five levels (bodily, perspectival, volitional, narrative, social), any of which can be manipulated in the lab or broken by illness.
- We perceive ourselves to control ourselves. By the Good Regulator Theorem, a system that regulates must model what it regulates; emotions are interoceptive inference, control-oriented predictions about the body's essential variables, and the deepest self is the formless feeling of being alive.
- Consciousness is tied to life, not intelligence. Intelligence is neither necessary nor sufficient: you don't have to be smart to suffer, and making computers smarter will not make them sentient. Seth's intuition is that self-maintenance "all the way down" in living systems matters.
- The near-term risk is AI that seems conscious. The Garland test (does a machine make us feel it is conscious, even knowing it's a machine?) is a test of humans, and it is the one we are about to fail, with consequences for empathy, manipulation and moral concern.
- Conscious level is not wakefulness. Dreaming (conscious, asleep) and the vegetative state (awake, unconscious) prove the two come apart; level tracks thalamocortical communication, not neuron count or overall activity.
- Measure complexity between order and disorder. Experience is always informative and integrated, so its signature is complexity, as in the perturbational complexity index ("zap and zip"), which separates vegetative from minimally conscious patients and finds covert consciousness.
- Behaviour can hide a mind. Owen's "tennis and house" test and PCI show that some apparently vegetative patients are conscious; inferring the inside from the outside is unreliable in clinics, animals and machines alike.
- Level and content are not independent. Psychedelics raise brain-signal diversity above normal waking, the first measure of level ever to do so, because wholesale changes in content register as changes in level.
- IIT's boldest claim is its least testable. Identifying consciousness with Φ requires a system's full repertoire of possible states, unmeasurable for brains; keep "integrated and informative" as properties to explain, not an identity.
- Bayes is the brain's logic. Best guesses combine priors and likelihoods weighted by precision, and each posterior becomes the next prior; base rates matter enormously (positive tests for rare conditions are mostly false alarms).
- Attention is precision. Turning up the estimated reliability of some signals turns down others, which is why we miss gorillas, signs and pickpockets in plain view.
- Action is self-fulfilling prediction. In active inference, moving means letting proprioceptive predictions override current evidence, which requires attenuating bodily signals (why you can't tickle yourself).
- Perceived reality is itself an inference. Objecthood, change, time and causality are constructed, and the felt "out-there-ness" of perception is what makes the hard problem seem harder than it is: a chair exists mind-independently; chairness does not.
- Body ownership is fragile but its illusions are weak. Rubber-hand and full-body illusions show ownership and perspective are inferences, but suggestibility drives much of the effect; the clinical cases (somatoparaphrenia, out-of-body seizures) are stronger evidence.
- Self-change blindness is a feature. We perceive ourselves as more stable than we are because regulation needs high-precision, self-fulfilling priors: the brain never updates its prior of expecting to be alive.
- Emotions don't have backs. Finding-out predictions give objects; control predictions give feelings of how things are going, which is why affect has no shape or location.
- The free energy principle grounds the story in physics. Living things persist by staying in expected states ("a fish in water"), which means minimising free energy, equal to prediction error under standard assumptions; it is mathematical philosophy, useful rather than falsifiable.
- Free will is real, but not spooky. Volition is the perception that an action's causes come from within; determinism doesn't matter; Libet's readiness potential is likely accumulated noise crossing a threshold; and "I could have done otherwise" helps you do otherwise next time.
- All mammals are probably conscious; the octopus is the most alien mind. Shared cortex, thalamus, sleep and anaesthesia signatures support mammals; octopuses run distributed, open-loop bodies that may lack human-style body ownership.
- Welfare should track capacity to suffer, not similarity to us. Responses to injury and analgesia are the best cross-species evidence; precaution should guide policy where evidence is thin.
- Functionalism is an assumption, not a fact. Whether consciousness is like Go (simulation is the real thing) or weather (simulation never gets wet) is open; claims about machine sentience inherit whichever theory they assume.
- Organoids may be closer than computers. Lab-grown human neurons remove the substrate objection, and the intended scale ("farms") makes preventative ethics urgent.
- Consciousness science is the third displacement. After Copernicus and Darwin, it places our minds within nature, and gives back more wonder than it takes.
Frameworks & mental models
| Name | What it says | Use it when |
|---|---|---|
| Real problem vs hard and easy problems | Explain, predict, control properties of experience, not its bare existence or mere functions | Framing any consciousness claim as tractable science |
| Level, content, self | Three targets for mechanistic explanation | Breaking "consciousness" into testable parts |
| Functionally agnostic physicalism | Physicalist, but neutral on whether function suffices | Avoiding premature commitment on AI and substrate |
| Zombie argument and the backwards A380 | Conceivability tracks ignorance, not possibility | Defusing armchair arguments |
| Level vs wakefulness | Awareness and arousal dissociate | Clinical diagnosis; anaesthesia monitoring |
| PCI (zap and zip) | Perturb cortex, compress the echo; complexity indexes level | Detecting covert consciousness |
| Informative and integrated | Every experience rules out alternatives and is unified | Choosing measures that sit between order and disorder |
| IIT and Φ | Consciousness is integrated information, intrinsic and measurable in principle | Assessing identity claims and machine-consciousness verdicts |
| Temperature vs life | One quantity vs many explained properties | Deciding what kind of explanation to expect |
| Controlled hallucination | Experience is the content of top-down predictions controlled by errors | Explaining illusions, hallucination and AI confabulation |
| Bayes' rule (prior, likelihood, posterior) | Update beliefs in proportion to evidence and its reliability | Diagnostics, screening, alert triage |
| Prediction error minimisation (models, hierarchies, precision) | The brain's approximate route to Bayes | Designing predictive systems |
| Attention as precision | Raising the gain of some signals lowers others | Explaining inattentional blindness; alert design |
| Active inference | Act to make predictions come true; actions as proprioceptive predictions | Agent design; exploration strategies |
| Objecthood as counterfactual prediction | Rich "what if I moved?" predictions make things feel like objects | Explaining why some experiences feel real |
| Beholder's share / Humean projection | Perceivers project colour, causality and reality onto the world | Understanding perceptual disagreement |
| Five levels of selfhood | Bodily, perspectival, volitional, narrative, social | Diagnosing which aspect of self is altered |
| Teletransportation paradox | Psychological continuity; both copies "real" | Thinking about identity of copies and clones |
| Interoceptive inference | Emotions are predictions about the body's internal state | Explaining affect, anxiety, depersonalisation |
| Good Regulator Theorem; System A vs B | Every good regulator must be a model of what it regulates | Control-system and risk-model design |
| Essential variables; allostasis | Keep life-critical quantities in bounds by anticipatory change | Predictive control and early-warning systems |
| Epistemic vs instrumental prediction | Finding out vs controlling; objects vs feelings | Explaining why emotions feel unlike vision |
| Beast machine theory | All experience is rooted in prediction for staying alive | Grounding consciousness in biology |
| Self-change blindness | Precise priors make the self feel more stable than it is | Understanding resistance to self-change |
| Free energy principle; Markov blankets | Persist by minimising free energy (prediction error) within a boundary | Unifying life and mind; exploration-exploitation |
| Dark room problem | Minimising surprise long-term requires exploration | Avoiding degenerate objectives in agents |
| Spooky vs compatibilist free will | No uncaused causes; freedom as control of degrees of freedom | Responsibility debates |
| Schurger's accumulator model | Readiness potential is noise crossing a threshold | Interpreting Libet-style data |
| Three features of volition | Doing what I want; could have done otherwise; from within | Analysing experiences of agency |
| Haggard's what, when, whether | Selection, timing, last-minute veto | Structuring agent decision pipelines |
| Anthropomorphism vs anthropocentrism | Over-attributing vs under-recognising non-human minds | Judging animal and AI minds |
| Consciousness profiles (Birch et al.) | Perceptual and evaluative richness, unity, temporality, selfhood | Comparing species without a single scale |
| Consciousness vs intelligence axes | Separate, multidimensional | AI capability vs moral-status claims |
| Turing test vs Garland test | Passing as human vs making us feel it is conscious | Evaluating anthropomorphic AI products |
| Uncanny valley | Near-human likeness provokes revulsion | Avatar and robot design |
| Preventative ethics | Act before synthetic consciousness or suffering arrives | Organoids, AI welfare policy |
Strengths. Seth gives the most coherent single-author synthesis of predictive processing and consciousness for general readers, with a disciplined method (the real problem), honest calibration of confidence (he labels intuitions as intuitions), and real experiments, several from his own lab, including some that cut against his view (the suggestibility results on the rubber hand). The life-not-intelligence thesis has proved prescient as LLMs forced the AI-consciousness question into public view.
Blind spots. The dissolution of the hard problem is a wager: explaining why experiences seem real and structured is arguably the meta-problem, and critics hold the existence of experience is untouched. Interoceptive inference, the hinge of the beast machine theory, is thinly tested, and heartbeat-based measures are unreliable. The framework's unifying scope makes it hard to falsify, and the free energy principle is admitted to be unfalsifiable. Rival theories (global workspace, higher-order, recurrent processing) get little space, and the case for biological naturalism rests on an unspecified "which property of life?".
Contested or dated. Some numbers and attributions slip: anaesthesia volume (four million a day is several times too high), the mammogram base rate off by a factor of ten, a Gaussian posterior drawn wider than its prior, Wittgenstein called German, a "Eurasian" elephant that was Asian, actus rea for actus reus. Claims about current AI (low on intelligence; GPT easily caught out) dated within two years.
How the field has moved. The Cogitate adversarial collaboration (2023; Nature 2025) challenged both IIT and global workspace theory, and a 2023 open letter branding IIT pseudoscience split the field. Clinically, cognitive motor dissociation in about a quarter of unresponsive patients (2024) vindicated the push for brain-based meters. Animal sentience gained legal and scientific standing (UK Sentience Act 2022, New York Declaration 2024). On AI, ChatGPT and successors passed text Turing tests; the 2023 Butlin, Long et al. report derived consciousness indicator properties assuming computational functionalism and found no current system a strong candidate; AI-welfare research began at major labs; and Seth himself now argues explicitly for biological naturalism (Behavioral and Brain Sciences, 2025). Organoid intelligence (2023) and Pong-playing DishBrain (2022) made his organoid worries concrete.
Who's who
| Name | Who | Why they matter here |
|---|---|---|
| Anil Seth | Neuroscientist, University of Sussex | Author; controlled hallucination and beast machine theory |
| David Chalmers | Philosopher | The hard problem, easy problems and meta-problem |
| Thomas Nagel | Philosopher | "What is it like to be a bat?" definition of consciousness |
| René Descartes / Julien Offray de La Mettrie | Philosophers | Beast machines with mind separate from life / L'homme machine, mind as a property of life |
| Gerald Edelman / Giulio Tononi | Neuroscientists | Experience as informative and integrated; Tononi founded IIT |
| Christof Koch / Francis Crick | Neuroscientists | The neural-correlates programme; Koch later an IIT champion |
| Marcello Massimini | Neuroscientist, Milan | Perturbational complexity index ("zap and zip") |
| Adrian Owen / Martin Monti | Neuroscientists | Tennis-and-house fMRI; communication with unresponsive patients |
| Adam Barrett, Michael Schartner, Robin Carhart-Harris | Researchers | Spontaneous complexity; psychedelics raising signal diversity |
| Hermann von Helmholtz / Immanuel Kant | Physicist; philosopher | Unconscious inference; the noumenon behind the sensory veil |
| Thomas Bayes | Minister and statistician | Bayes' rule for inference to the best explanation |
| Karl Friston | Neuroscientist | Active inference and the free energy principle |
| Andy Clark, Jakob Hohwy, Chris Frith | Philosopher; philosopher; psychologist | Predictive processing; "self-evidencing"; "controlled hallucination" phrase |
| Daniel Simons / Edward Adelson | Psychologist; vision scientist | The invisible gorilla; the checkerboard illusion |
| Keisuke Suzuki, Warrick Roseboom | Sussex researchers | Hallucination machine, VR objecthood, time perception |
| Ernst Gombrich | Art historian | The beholder's share; no innocent eye |
| Derek Parfit | Philosopher | Teletransportation paradox and personal identity |
| Thomas Metzinger | Philosopher | Being No One; moratorium on synthetic phenomenology |
| Olaf Blanke / Henrik Ehrsson | Neuroscientists | Full-body and body-swap illusions; out-of-body experiences |
| Peter Lush, Zoltán Dienes | Psychologists | Suggestibility as a driver of the rubber hand illusion |
| Clive and Deborah Wearing | Musicologist; author | Amnesia and the loss of the narrative self |
| William James / Carl Lange | Psychologists | Emotions as perceptions of bodily change |
| Hugo Critchley / Lisa Feldman Barrett / Antonio Damasio | Neuroscientists | Interoception; constructed emotion; body-based self |
| W. Ross Ashby, Roger Conant, Norbert Wiener | Cyberneticists | Good Regulator Theorem, essential variables, cybernetics |
| Benjamin Libet / Aaron Schurger | Neuroscientists | Readiness potential; its threshold-crossing reinterpretation |
| Patrick Haggard | Neuroscientist, UCL | What-when-whether model of volition |
| Daniel Dennett | Philosopher | Double transduction; compatibilism; "intelligent tools, not colleagues" |
| Graziano Fiorito / Peter Godfrey-Smith | Biologist; philosopher | Octopus cognition; cephalopods as "the most other" minds |
| Jonathan Birch | Philosopher | Consciousness profiles; precautionary sentience policy |
| Murray Shanahan / Alex Garland | AI researcher; filmmaker | The Garland test; Ex Machina |
| Hiroshi Ishiguro / Masahiro Mori | Roboticists | Geminoids; the uncanny valley |
| Alysson Muotri | Organoid researcher | Brain organoids and their ethical scale |
| Nick Bostrom / Ray Kurzweil | Philosopher; futurist | Simulation argument and paperclips; the Singularity |
Connect it
- Hallucination as prior over evidence. Seth's definition maps directly onto LLM confabulation: learned priors overwhelming weak context. Grounding through retrieval, tool outputs and validation supplies the "control" in controlled hallucination; substitutional-reality tests (secretly swapping ground truth) check whether a system notices its inputs no longer match the world.
- Base rates and precision in financial-crime screening. Sanctions, PEP and fraud alerts are mammogram arithmetic: with rare true matches, most hits are false. Improve priors (risk segmentation), likelihoods (richer features) and precision weighting (verified identifiers outweigh fuzzy names), and let each posterior become the next prior.
- Good regulators and agent pipelines. Risk controls should be System B, modelling how the portfolio responds to interventions; agentic platforms should separate Haggard's what, when and whether, with a hard veto gate before irreversible actions, and Ashby's requisite variety warns that a controller simpler than its environment fails on unrepresented cases.
- Exploration over comfort. The dark room problem is a design warning: agents rewarded only for predictability avoid informative edge cases, so objectives need explicit information-gain terms.
- Design for the Garland test. Customer-facing assistants should not simulate feelings, should disclose machine status, and should treat seeming-consciousness as a conduct and customer-vulnerability risk separate from capability risk; fluent self-reports of experience are not evidence.