Insight Editions · No. 289
From Sensing to Sentience
How Feeling Emerges from the Brain
Todd E. Feinberg
Todd Feinberg argues that sentience, the feeling side of consciousness, is a weakly emergent system property of living, sufficiently complex nervous systems.
A study guide to From Sensing to Sentience by Todd E. Feinberg. 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
From Sensing to Sentience is neurologist Todd Feinberg's case that feeling is an ordinary product of biological emergence, built up over 3.5 billion years from the sensing of single cells to the brains of fish, octopuses, bees and crabs. His theory, Neurobiological Emergentism (NBE), joins an evolutionary account of when and where sentience appeared to a philosophical account of why it is personal. Feinberg's central claim is that there is no explanatory gap between brain and experience, only an experiential gap between knowing about a feeling and having it, and both sides are explained by emergence.
Sentience is a weakly emergent system property of living, sufficiently complex and hierarchical nervous systems; because it emerges from an individual organism's embodied life, it is necessarily personal.
Evolutionary neurobiology from E. coli to vertebrates, cephalopods, arthropods and velvet worms, plus philosophy of mind (Broad, Levine, Chalmers, Searle, Jackson, panpsychism, IIT).
3 stages (ES1 to ES3); 4 complexity variables; 8 Birch criteria; 4 lines to sentience; ES3 at 560 to 520 million years ago.
Readers who want a naturalistic, evolution-first account of feeling and a clear map of which animals are sentient and why, including anyone weighing claims about machine sentience.
The target is sentience, broadly defined. Feinberg means every experience with "something it is like" to it, both exteroceptive (sights, sounds, smells) and interoceptive-affective (pain, pleasure, emotion). The mystery, he argues, lies in its personal nature, captured by Broad's archangel, Levine's explanatory gap, Chalmers's hard problem and Jackson's Mary.
Emergence is the explanatory tool. All biological emergence shares four features (aggregate novelty, unity, process, hierarchy), and hierarchies with reciprocal connections across multiple scales multiply emergent properties. Life itself is the first emergent feature, and it carries sensing with it.
Sentience is inferred by criteria. Exteroceptive sentience requires higher-order mapped representations that create mental images; interoceptive-affective sentience requires valence circuits. Behaviourally, the eight Birch-Crump criteria and the master criterion of nonreflexivity separate sentience from sensing.
Three emergent stages. Four variables (neuron number, specialization, hierarchical levels, level interaction) define ES1 single cells (sensing, 3.5 billion years ago), ES2 "presentient" early nervous systems (about 570 million years ago: C. elegans, amphioxus, sea slugs), and ES3 complex brains (560 to 520 million years ago). The step from ES2 to ES3 took only 10 to 50 million years.
Four independent lines reached ES3. Vertebrates, coleoid cephalopods, arthropods (insects and crabs) and onychophorans each evolved image-forming senses, topographic maps, affective centres and flexible behaviour, a striking convergence that suggests what sentience requires.
There is no "top" where feeling appears. Feinberg rejects Sperry's emergence at the top, Chalmers's strong emergence and Kim's layer cake. Sentient brains combine serial maps, nested hierarchies, hierarchical modular networks and nested oscillations, linked by upward, downward and circular causality: sentience is a system feature, not a level feature.
Rivals are too generous or too abstract. Panpsychism and biopsychism (Reber's emergentist's dilemma) push consciousness down to matter or life because they doubt emergence; NBE offers an emergentist's solution. IIT's Φ is a disguised measure of emergence, but its bar is far too low, its scale continuous, and it ignores life and specific substrates.
The gap is experiential. Because sentience and its personal nature coemerged and coevolved, a complete physical account still cannot substitute for being the organism; that is why objectifying experience is futile, and why the remaining gap poses no threat to physicalism.
| Ch | Title | One-line payload |
|---|---|---|
| P | Preface | A bottom-up, weak-emergence theory of feeling, with an experiential gap in place of an explanatory one. |
| 1 | Introduction: Neurobiological Emergentism | Sentience is all felt experience; its personal nature drives the gap arguments that Searle and NN left unresolved. |
| 2 | General Features of Biological Emergence | Four general features plus reciprocal multi-level hierarchies; sentient animals differ by degree of neural hierarchy. |
| 3 | The Neurobiological and Behavioral Criteria for Sentience | Mental-image maps, valence, the eight Birch-Crump criteria and nonreflexivity separate feeling from sensing. |
| 4 | The Stages of the Emergence of Sentience | Four complexity variables define ES1, ES2 and ES3, with the big leap 560 to 520 million years ago. |
| 5 | Emergent Stage 1: Single-Celled Organisms | E. coli and Stentor sense with sophistication but reflexively; Reber's cellular consciousness is rejected. |
| 6 | Emergent Stage 2: Early Nervous Systems | C. elegans, amphioxus and sea slugs integrate and assign valence but lack maps and affect: presentient. |
| 7 | Emergent Stage 3: Sentient Animals | Vertebrates, cephalopods, insects, crabs and velvet worms converge on maps, affect and flexible behaviour. |
| 8 | Neurobiology, Emergent "Levels," and Sentience | No top, no layer cake: diverse hierarchies and circular causality make sentience a system feature. |
| 9 | Panpsychism, Biopsychism, the Emergentist Dilemma and IIT | NBE dissolves Reber's dilemma; IIT measures emergence but sets the bar too low and ignores life. |
| 10 | Demystifying the Personal Nature of Sentience | Sentience is personal as life is personal; Mary shows an experiential gap compatible with physicalism. |
| 11 | Putting It All Together | Five postulates and a two-part solution: specific substrates plus personal embodiment. |
| E | Epilogue: Emergence, Process, and Sentience | With Crick: emergence is understandable in principle; sentience is a process, and the self is flux. |
Preface
Feinberg sets out to explain sentience, the "feeling" aspect of consciousness, as a product of ordinary biological and neurobiological emergence. He works bottom-up, from the sensing of a bacterium to the brains of frogs, crabs, bees and octopuses, and he introduces a new theory, Neurobiological Emergentism (NBE), that claims to close the supposed explanatory gap without any new physics.
The question and the method
The book is about the origins and nature of sentience, which Feinberg treats as the part of consciousness that involves feeling: there is something it is like for the creature, whether that something is a colour, a smell or a pain. His bet is that the concept that does the most work in explaining it is emergence: a complex system can have novel properties that arise from its parts and their interactions but that are absent in the parts considered alone. A "higher-level" system property is more than the simple sum of the "lower-level" parts.
Most theories of consciousness, especially philosophical ones, run top-down. They begin with fully realized human experience and ask what could produce it. Feinberg reverses the direction. He looks for the origins of sensing in all living things and asks what must be added, step by step, before feeling appears. The journey covers billions of years of evolution, starting from the bacterium Escherichia coli and ending with frogs, crabs, bumble bees, octopuses, velvet worms and humans.
Feinberg finds no "magic bullet" that creates sentience and no single moment when it switched on. Sentience appears at the end of a long and winding road of diverse, progressive emergent levels, and it appears more than once.
What NBE is meant to do
Feinberg chose emergence as the organising idea because he thinks it can unify the biological, neurobiological, evolutionary and philosophical problems of sentience under one set of principles. He gives NBE three jobs.
- General mechanism. Explain how neurobiological emergent processes operate in the creation of sentience in any animal that has it.
- Evolutionary parallel. Show that biological and neurobiological emergence run parallel with the evolutionary progression from sensing to sentience, stage by stage.
- Philosophical payoff. Use the emergent character of sentience to explain both its personal nature and the apparent explanatory gap between brain and experience.
The core biological claim is that sentience is a feature of sufficiently neurobiologically evolved and complex brains. It descends from living single-celled organisms, but it took as much as 3 billion years for it to emerge. The middle of the book (chapters 5 to 7) walks through a three-stage model, with a timeline, that separates simple sensing from sentience proper.
Weak versus strong emergence
The neural complexity of sentience and the personal, unique character of feeling have led some thinkers to say that no standard physical science could ever explain consciousness. The most extreme version is "strong" or "radical" emergence: sentience must result from some novel "fundamental" physical property or process, perhaps one that appears "at the top" or at the highest hierarchical level of the brain.
NBE is explicitly a "weak" emergence theory. Here "weak" means that standard, natural biological emergence is all that is required, exactly as every complex physical system displays emergent system properties. Feinberg's strategy against the strong view is to examine the many diverse neurobiological hierarchies found in sentient animals, at micro and macro scales, across different neural architectures, and to show that each produces unique but standard emergent system properties. The aggregate of these properties is sentience.
Strong (radical) emergence. Sentience needs a new fundamental property of physics, appearing at the top of the neural hierarchy; ordinary science cannot reach it.
Weak emergence (NBE). Sentience is an ordinary system property of a sufficiently complex, evolved nervous system; standard biology explains it.
The rival theories and the gap
Chapter 9 confronts the views closest to NBE: panpsychism, biopsychism with the "emergentist dilemma", and Integrated Information Theory (IIT). Feinberg's position is that sentience is an emergent feature of life, yet life and basal sensing do not meet the criteria for sentience. IIT differs from NBE in many ways, but he argues that it actually supports several emergentist postulates of his theory.
The philosophical heart of the book is chapter 10. Feinberg argues that life is a personal emergent aggregate feature of an embodied organism, and so is sentience. If sentience is the system property of one particular living organism, then it can only be subjective and personal to that organism. That reframes the famous problem.
Feinberg's central philosophical claim is that there is no "explanatory gap" but an experiential gap, whose two sides, objective neural processes and subjective experience, are both explained by standard emergent principles.
The corollary is blunt: trying to fully "objectify" the subjective aspects of sentience is futile. The gap between scientific description of brain processes and knowledge by "direct acquaintance" is real, but it is no obstacle to a full scientific explanation.
The two-emergences test. Imagine growing, in a dish, an exact copy of the neural circuitry of a frog's tectum, perfectly wired but cut off from any body, metabolism or life history. Then imagine the same circuitry inside a living frog hunting a fly. Feinberg's view predicts a difference: sentience needs both emergences, one from evolved neural complexity and one from the individual life of the organism. The thought experiment shows why he refuses to treat feeling as a property of circuitry alone.
Two emergences
Chapter 11 pulls the postulates together. The key idea is that two emergences jointly create sentience and its personal character:
- Neural emergence. Sentience is an emergent feature of neurobiologically evolved nervous systems, as chapters 5 to 7 trace.
- Life emergence. Sentience is an emergent feature of the individual life of the organism, which is what makes it belong to someone.
- Co-emergence. Put the two together, as evolution did, and sentience and its personal nature naturally coevolve and coemerge from life and complex nervous systems.
| Ch | Topic | Role in the argument |
|---|---|---|
| 1 | What sentience is; the explanatory gap | Sets the problem |
| 2 | General features of biological emergence | Builds the toolkit |
| 3 | Criteria for sentience | Defines the test |
| 4 | Three-stage model and timeline | Organises the evidence |
| 5-7 | ES1 cells, ES2 early brains, ES3 complex brains | Traces the evolution |
| 8 | Neural hierarchies and "levels" | Rebuts strong emergence |
| 9 | Panpsychism, biopsychism, IIT | Positions NBE among rivals |
| 10 | Personal nature, experiential gap | Philosophical resolution |
| 11 | Putting it all together | Summary of postulates |
- Jon Mallatt: Feinberg's main collaborator for a decade on the neurobiology and evolution of consciousness; co-author of The Ancient Origins of Consciousness (2016) and Consciousness Demystified (2018).
- Reviewers and advisers: Jonathan Birch, Lars Chittka, Andrew Crump, Robert Elwood, Robert Endres, Rhanor Gillette, Stevan Harnad, Claus Hilgetag, Jeremy Gunawardena, Thurston Lacalli, Christine Martin, Georg Mayer, Athanasios Metaxakis, Paul Nunez, Nicholas Strausfeld and Nektarios Tavernarakis, a list that signals the book's reach into animal sentience, insect cognition, bacterial sensing and invertebrate neuroanatomy.
- Jill K. Gregory: certified medical illustrator at the Icahn School of Medicine at Mount Sinai, who drew the book's figures.
For anyone building or governing AI systems, the preface already draws a line worth keeping: sensing and responding are not feeling. A bacterium, a thermostat and a fraud-detection model all sense and react; Feinberg's question is what additional architecture and life-like embodiment would be needed before "feeling" is even a candidate description.
Introduction: Neurobiological Emergentism
Feinberg defines his target, sentience, as every kind of experience that has "something it is like" to it, and identifies its personal nature as the real source of the mystery. He reviews the classic statements of the explanatory gap (Broad, Levine, Chalmers, Carruthers and Schier), explains why Searle's Biological Naturalism and his own earlier Neurobiological Naturalism left that gap open, and introduces Neurobiological Emergentism (NBE) as the fix.
Sentience: something it is like to be
"Consciousness" is an umbrella term covering everything from simple wakefulness up to high cognition and self-awareness. Feinberg narrows the target to the subjective, feeling aspect, which others call sensory consciousness, primary consciousness or phenomenal consciousness. His preferred term is sentience, from the Latin sentire, "to feel": being sentient means being capable of feeling. This is the aspect Thomas Nagel (1974) pointed to when he said an organism is conscious if and only if there is something it is like to be that organism.
Some writers restrict "sentience" to valenced experience, states that are positive or negative such as pain and pleasure. Feinberg deliberately uses the broader, Nagelian sense. Sentience covers both:
- Exteroceptive experience. What is seen, heard, tasted, smelled and touched from the outside world.
- Interoceptive-affective experience. Internally generated feelings from the body: pain, pleasure, emotions and similar states.
Because every one of these involves "something it is like", all experiences count as sentient. When Feinberg discusses other writers he may say "consciousness", but what NBE explains is sentience.
A definitional choice with consequences. Many in the animal-welfare literature (for example Jonathan Birch and the UK sentience review) treat sentience as the capacity for valenced experience, because that is what matters morally. Feinberg's broader usage makes "sentient" equivalent to phenomenally conscious. Readers comparing his conclusions with welfare-policy work should keep the two meanings apart: an animal could, in principle, have sensory experience with little affect, or the reverse.
Why sentience is hard: its personal nature
Feinberg's diagnosis is that the key element of the mystery is the personal nature of sentience. There seems to be an "explanatory gap" between the objective biology of the brain and the inherent subjectivity of experience. He lays out four classic formulations of that gap.
Broad's mathematical archangel (1925)
C. D. Broad (Charles Dunbar Broad, 1887 to 1971), in The Mind and Its Place in Nature (1925), offered one of the most famous thought experiments in the mind-body literature.
The mathematical archangel. An omniscient being knows the complete chemistry of ammonia and the entire neurobiology of the smell pathway, from olfactory nerve to brain. Broad says the archangel could predict the changes in the mucous membrane and olfactory nerves, but could not predict that these changes would be accompanied by a smell at all, still less the peculiar smell of ammonia, unless someone told him or he smelled it himself. The point: no objective knowledge substitutes for the personal experience of what it is like.
Levine's explanatory gap (1983)
Joseph Levine, in "Materialism and Qualia: The Explanatory Gap" (1983), coined the phrase. His example is pain and C-fiber firing. Even if pain is identical to C-fiber firing, nothing about the firing makes it naturally "fit" the way pain feels rather than some other set of phenomenal properties. The identity leaves how pain feels as "merely brute fact".
Chalmers's hard problem (1995, 1996)
David Chalmers made Levine's gap central to the "hard problem" of consciousness. The "easy problems" are explaining the objective computational and brain processes behind discrimination, report and control. The hard problem asks why all that processing does not go on "in the dark", free of any inner feel. Why does a retina discriminating electromagnetic waveforms yield a vivid sensation of red? A mere account of functions stays on one side of the gap; the bridge must be built from something else.
Chalmers adds the problem of the character of experience (1996): why is seeing red like this rather than like that, why not the experience one has looking at blue, or even the sound of a trumpet? This is the closest relative of Levine's point.
The "something left over" argument
Peter Carruthers and Elizabeth Schier (2017) name the pattern the "something left over argument": no matter how many easy phenomena are explained, something is always left over, namely why experience is the way it is. The underlying assumption is that experiences are the wrong kind of thing to be explained by the structure and function of mental states.
- Broad (1925), the archangel. Total physical knowledge cannot predict the smell of ammonia.
- Levine (1983), the explanatory gap. C-fiber firing does not explain why pain feels as it does; the link is brute fact.
- Chalmers (1995, 1996), the hard problem and the character of experience. Function does not explain why processing is felt, or why red feels like red.
- Carruthers and Schier (2017), something left over. Every functional explanation leaves the hard phenomenon unexplained.
Feinberg's answer, previewed here, is that the processes of biological emergence hold the key to explaining the gap.
From Biological Naturalism to Neurobiological Naturalism
John Searle's Biological Naturalism holds that all of consciousness, including feeling, is a natural biological feature of the brain requiring no new scientific laws or principles. "Biological" because the right level of explanation is biology; "naturalism" because consciousness belongs to nature like photosynthesis, digestion or mitosis. Searle also insists that consciousness has a first-person or subjective ontology: every conscious state is someone's state, and it cannot be reduced to third-person ontology any more than neuron firings can be reduced to feelings, because in either direction you leave something out. For Searle this poses no obstacle to a complete science of consciousness.
Feinberg agreed with the subjective ontology but saw one clear limitation. Searle rightly said consciousness was a natural biological feature, yet offered no biological explanation of either the emergence of consciousness or its uniquely personal nature. Photosynthesis and mitosis have clear scientific explanations; sentience resists that kind of standard biological theorizing. Searle had, in effect, no scientific explanation for the explanatory gap itself.
Over the previous decade, with evolutionary biologist Jon Mallatt, Feinberg built Neurobiological Naturalism (NN), named after Searle's theory. NN made two claims and proposed a model:
- Natural but special mechanisms. Consciousness exploits complex biological mechanisms unique to neurobiology, but these are fully natural and need no new laws.
- Life as an ingredient. Life itself is an important ingredient in creating consciousness and its subjectivity.
- Three-stage model. Stage one, nonconscious single-celled organisms; stage two, animals with neurobiologically more complex reflexes but no consciousness; stage three, a diverse group of animals with more evolved brains and consciousness, reached during the Cambrian period, over 520 million years ago.
| Theory | Who | Core claim | What it leaves open |
|---|---|---|---|
| Biological Naturalism | Searle | Consciousness is natural biology with first-person ontology | How it emerges; why it is personal |
| Neurobiological Naturalism | Feinberg and Mallatt | Special but natural neural mechanisms; life matters; three stages | The gap still lacks a mechanism |
| Neurobiological Emergentism | Feinberg | Emergence explains both sentience and its personal nature | Claims to leave nothing open |
Neurobiological Emergentism
NBE is presented as an outgrowth of and advance upon NN. Feinberg concedes up front that the idea that consciousness is an emergent brain process is not new; what he claims is new is the integration of biology, neurobiology, nervous-system evolution and the philosophy of sentience under emergence.
- Sensing emerges from life. Just as life is an emergent feature of inanimate matter, the basic sensing capabilities present in all living things are emergent features of the organism's life.
- Sentience emerges from specific neural features. The progressive emergence of sentience from sensing can be traced to specific features of neurobiologically complex and evolved brains.
- Not different in kind. The emergence of sentience is not different "in kind" from emergence in general or in biology; it is the greater degree of neurobiological emergence in sufficiently evolved nervous systems that produces it.
- No gaps. NBE explains the route from the objectivity of material brains to the subjectivity of experience without scientific or philosophical gaps.
The fourth principle rests on Feinberg's view that sentience and its personal nature co-occurred and coevolved pari passu, so a single emergent story explains both, leaving an experiential gap rather than an explanatory one.
That experiential gap is nonmysterious and naturally occurring, and it poses no scientific barrier. The theory claims a seamless explanatory pathway from the physical brain to personal experience, with a caveat Feinberg repeats throughout: attempting to fully "objectify" experience is futile. NBE, he says, neither supports dualism nor argues against physicalism, and it is meant to settle some long-running debates.
Feinberg's move is to locate the mystery in personhood, not in physics. If what makes feeling seem inexplicable is that it always belongs to one particular living system, then an account of how a particular system's properties emerge can explain why the feeling is accessible only from inside.
Is the gap dissolved or renamed? Critics of this family of moves (often called "phenomenal concept" or "acquaintance" replies) say that showing why subjective facts are only accessible from one perspective does not yet explain why there is any perspective at all. Feinberg's later chapters must carry that weight; this chapter only asserts it.
The Searle limitation is familiar from architecture work: saying a capability is "a natural feature of the platform" is not a design. A system-level property needs an account of which components, interactions and hierarchy levels produce it. NBE is Feinberg's attempt to supply that missing design for sentience.
General Features of Biological Emergence
Before tracing sentience through evolution, Feinberg fixes the toolkit: four general features of biological emergence (aggregate novelty, unity, process, hierarchy) and four features of emergence in hierarchical systems. The decisive point is that all living things share the general features; what sets sentient animals apart is specialized neurons arranged in many reciprocally connected neural hierarchies.
What emergence means
The word comes from the Latin emergo, "to arise or come forth". Its first scientific use is usually credited to G. H. Lewes in Problems of Life and Mind (1875). Since then its meaning has been stable: novel features of a system arise from the parts and their interactions, and those features are absent from the parts considered in isolation. An emergent "higher-level" property is more than the sum of the "lower-level" parts. Ernst Mayr (1982) noted that emergence has long been invoked for the hardest phenomena: life, mind and consciousness.
Feinberg's interest is biological. The principles below are consistent with emergence in general, but when he says "emergence" he means emergence in living systems, from any plant or animal down to single-celled organisms.
The four general biological emergent features
- Aggregate system feature and novelty. An emergent feature is a property of the whole system that is not present in the parts, so it is novel relative to them.
- Unity. For emergence to occur, the parts must be physically united, integrated, or at a minimum interacting in some fashion.
- Process. Emergent features are processes created by the dynamic interaction of the system's parts.
- Hierarchy. Hierarchical systems are critical to emergence and increase emergent properties.
Aggregate system functions and novelty
First and foremost, emergent features are "higher-level" properties created by the collective functions of the parts, which no part possesses alone. Hence an emergent property is a novel system feature that is not "reducible" to its parts.
- Water. The standard inorganic example: liquidity, or wetness, is a property of water, yet no single water molecule is wet.
- Life. The example that matters most for sentience: an atom or a DNA molecule inside an organism is not alive, but a single-celled organism, or a body made of single cells, is.
Philosopher Elly Vintiadis (2013) gives a compact definition: a property is emergent if it is a novel property of a system that arises when the system reaches a certain level of complexity, and that, although it exists only insofar as the system exists, is distinct from the properties of the parts. The definition is uncontroversial. The interesting question for sentience is what type of novelty it is, which reopens the weak versus strong emergence debate. Feinberg declares for the weak view now and defends it in chapters 8 to 10.
Weak emergence. The novel property is new at the system level but follows from the parts, their interactions and standard laws. Liquidity and life are the models.
Strong emergence. The novel property involves new fundamental laws, powers or properties not derivable even in principle from the parts. Some say consciousness must be like this.
Unity
The parts of an emergent system must be connected, integrated or at least interacting. The required degree of connection varies, depending partly on whether the emergent feature arises within a hierarchical level or across levels. Even outside the nervous system this holds: in the liver, the functions performed by individual hepatocytes depend on much closer physical and temporal continuity between cells than the emergent functions of the larger lobule they belong to.
Unity matters most for sentience. The nervous system, analysed objectively, consists of many billions of individual neurons, yet sentience is experienced as a largely unified field. That kind of integration would be impossible if the parts were not connected.
Process
Emergent features are processes, not things. Neuroanatomical "parts and wholes" are the easiest aspects of a nervous system to define, but sentience is as much a process as digestion, photosynthesis or mitosis, Searle's examples. This is one reason sentience cannot be strictly localized the way an anatomical structure can.
Sentience is something a nervous system does, not a place in it. Because emergent features are dynamic processes of interacting parts, asking "where" sentience is located is like asking where digestion is located inside a single enzyme.
Hierarchies promote emergence
Feinberg calls hierarchical arrangements, especially neurohierarchical connections and processes, absolutely essential for the emergence of sentience. In neural hierarchies lower levels influence higher levels, which in turn influence lower levels, and structures within the same level influence each other through extensive reciprocal connectivity. Those reciprocal connections increase exponentially in advanced brains that have sentience.
- Hierarchy matters everywhere in biology, and especially in neurohierarchical systems.
- Reciprocal connections among parts within and between levels greatly enhance the emergence of novel properties.
- Multiple scales at once. Emergent properties may occur simultaneously at multiple spatial and scalar levels and across diverse temporal frequencies.
- More levels, more novelty. Novel properties emerge in the system as a whole as additional (typically "higher") levels are added.
Nested and nonnested hierarchies
Neural hierarchies come in diverse architectures, and all of them add to the overall emergence of the system. Chapter 8 develops the distinction at length.
| Architecture | How levels relate | Biological example |
|---|---|---|
| Nested | Higher levels are composed of lower levels | Molecules within cells within tissues within organs |
| Nonnested | Lower levels project activation and "information" upward without being physically contained | Retina projecting to tectum or visual cortex |
Specialization as levels are added
As a hierarchy evolves and more levels are added, its parts and levels become more specialized in structure and function. Feinberg's example is vision: in fish the visual system is centred in the optic tectum, but over vertebrate evolution visual processing moves rostrally toward the cortex in mammals.
The rostral-shift story is a simplification. The mammalian superior colliculus (homologue of the tectum) still does heavy visual work, especially orienting and attention, and in fish and amphibians the pallium also contributes to vision. Feinberg's point that specialization increases with added levels stands, but vision did not simply relocate from midbrain to cortex.
Why these features matter for the rest of the book
All animals, sentient or not, possess most of the general features of biological emergence; sentient animals are distinguished by specialized neurons and multiple, hierarchically interconnected neural hierarchies that make new emergent properties possible.
This framing does two jobs at once:
- Continuity. Because sensing in a bacterium and sentience in a bee obey the same emergence principles, there is no new law to be found.
- Discontinuity. Because degree of hierarchy, reciprocity and specialization differs enormously, there is a principled line between sensing and sentience, which chapter 3 turns into explicit criteria.
The stadium wave. A crowd produces a wave that travels around a stadium: an aggregate, unified, dynamic process that no individual spectator possesses, and that cannot be "found" in any seat. Now add levels: sections coordinate with section leaders who watch the scoreboard, which in turn watches the sections. The wave becomes able to reverse, split and respond to the game. Feinberg's claim is that feeling is like the second wave, not the first: same principles, many more reciprocally linked levels.
The four features read like an architecture checklist for any system-level capability: novel at the whole, requires integration, is a running process, and grows with layered reciprocal structure. In AI platforms, emergent behaviours (say, multi-agent coordination) likewise depend on connection topology and feedback between layers, not on any single component.
The Neurobiological and Behavioral Criteria for Sentience
Since nobody can experience another creature's mind, sentience must be inferred from criteria. Feinberg splits sentience into two subtypes, exteroceptive (marked by mapped neural representations and sensory mental images) and interoceptive-affective (marked by valence), and adopts the eight Birch et al. and Crump et al. criteria for pain sentience. The thread tying the behavioural criteria together is nonreflexivity: sensing is reflexive, sentience is not.
The other minds problem
Feinberg opens with a concession: there are no absolute, universally agreed criteria for objectively determining sentience, especially in other animals. This is a version of the philosopher's "other minds problem". We cannot experience what it is like to be Nagel's bat, still less whether there is anything it is like to be an Escherichia coli bacterium. So judgments must rest on inference from a set of general criteria and the evidence in each case.
He divides the task into three questions:
- Subtypes. What are the major subtypes of sentience?
- Infrastructure. What neurobiological infrastructure does each subtype require?
- Behaviour. What behavioural evidence supports the presence of each subtype?
Two subtypes of sentience
| Feature | Exteroceptive sentience | Interoceptive-affective sentience |
|---|---|---|
| Source | The outside world | The body and internal states |
| Examples | Sight, sound, touch, smell, taste | Pain, hunger, thirst, drives, emotions |
| Neural signature | Higher-order mapped (isomorphic) representations | Less tightly mapped; different structures |
| Valence | Not in itself positive or negative | Valenced (pleasant or unpleasant) |
| Marker | Sensory mental images | Nonreflexive motivated behaviour |
Exteroceptive sentience: maps and mental images
Exteroceptive sentience is the capacity to have "feelings" associated with sensing the environment, through photosensation (vision), mechanosensation (touch and proprioception), chemosensation (taste and smell) and hearing. The feature that separates it from mere sensing is whether the sensory processes create sensory mental images.
Feinberg leans on two precedents. Gerald Edelman made mental images central to "primary consciousness", defined as being mentally aware of things in the world, having mental images in the present. Antonio Damasio describes how "mapped neural patterns" yield the mental images of his "core consciousness", a close relative of Edelman's notion.
The necessary neural condition for a unified sensory image is a higher-order mapped neural representation of the sensory domain. Maps come in several kinds:
- Topographic maps. Spatial ordering is preserved from the receptor sheet up through every level of the central hierarchy.
- Somatotopic map. Touch, position sense and nociception mapped point by point onto the body, distorted so that face and hand get the largest areas.
- Retinotopic map. In mammals, in the occipital lobe, preserving the spatial layout of the retina.
- Tonotopic map. Auditory cortex organised by sound frequency, not by location in space.
- Chemosensory maps. Taste and smell organised mainly by chemical receptor properties.
A mapped representation is necessary but not sufficient. The spinal cord has somatotopic maps at low levels of the touch pathway, yet these do not create unified, sentient mental images by themselves. Sentience needs successively hierarchical mapped representations. Lower-level maps are still useful as neurobiological markers of the transition from simple sensing toward sentient images, which is why the emergence of exteroceptive sentience can be traced alongside the growth of neurohierarchical organization.
Interoceptive-affective sentience: valence
The second subtype is internally derived. Interoceptive feelings respond to physiological changes in the body (pain, hunger, thirst, motivational drives), while affective awareness covers all emotional, valenced experience. It does not require tight mapping and relies partly on different anatomical structures.
The defining property is valence. Ralph Adolphs and David Anderson (2018) define it as the pleasant/unpleasant dimension, or, in stimulus-response terms, appetitive versus aversive. Kent Berridge (2019) splits affective valence into two parts:
- Hedonic aspect. Positive hedonic impact (pleasure) or negative hedonic impact (displeasure, pain), the affective kernels of reward and punishment.
- Motivational aspect. Functions that promote pursuit of rewards (incentive salience, declarative goals) and functions that mediate threat avoidance (fearful salience, passive avoidance).
Berridge's split matters because approach and withdrawal exist in every living organism, needed for feeding and avoiding harm. Feinberg argues these simple sensing behaviours are precursors of sentience but can occur without it, so on their own they are not evidence of sentience.
The Birch and Crump criteria
For interoceptive-affective sentience, evidence can be neuroanatomical (the infrastructure in figure 3.2) or behavioural. Feinberg adopts the criteria of Jonathan Birch et al. (2021) and Andrew Crump et al. (2022), devised to assess pain sentience in cephalopod molluscs and decapod crustaceans but applicable across species. He finds them succinct, convincing and comprehensive and consistent with his own work with Mallatt.
- Nociception. The animal has receptors sensitive to noxious, damaging stimuli (nociceptors).
- Sensory integration. The animal has brain regions that integrate information from different sensory sources.
- Integrated nociception. Neural pathways connect the nociceptors to the integrative brain regions.
- Analgesia. Behavioural responses to noxious stimuli are modulated by chemicals acting on the nervous system, either an endogenous pain or distress modulating system or drugs such as local anaesthetics, anxiolytics or antidepressants.
- Motivational trade-offs. The negative value of a noxious stimulus is weighed against the positive value of a reward, giving flexible decisions.
- Flexible self-protection. Wound-tending, guarding, grooming or rubbing that likely involves representing where on the body the noxious stimulus was.
- Associative learning. Noxious stimuli become associated with neutral stimuli, or new ways to avoid them are learned through reinforcement, going beyond simple classical conditioning; instrumental, reversal and trace conditioning count as stronger evidence.
- Analgesia preference. When injured, the animal learns to self-administer analgesics, prefers a location where they are available, or prioritizes them over food.
Feinberg sorts the list by type of evidence:
- Neuroanatomical (criteria 1 to 3). Nociception, sensory integration, integrated nociception.
- Behavioural (criteria 5 to 7). Motivational trade-offs, flexible self-protection, associative learning.
- Mixed neurochemical and behavioural (criteria 4 and 8). Analgesia and analgesia preference.
- Working rule. For behaviour he applies criteria 4 to 8.
The criteria became policy. The Birch team's 2021 report for the UK government (the LSE review of sentience in cephalopod molluscs and decapod crustaceans) used these eight criteria and led to the Animal Welfare (Sentience) Act 2022, which extends legal recognition of sentience to all cephalopods and decapods. In April 2024 the New York Declaration on Animal Consciousness stated strong scientific support for consciousness in mammals and birds and "at least a realistic possibility" of conscious experience in all vertebrates and many invertebrates (including cephalopods, decapods and insects). Birch's The Edge of Sentience (2024) develops a precautionary framework built on the same evidence base.
Nonreflexivity: the common thread
Criteria 4 to 8 share one property: the behaviours are nonreflexive. Nociception (from the Latin nocere, "to injure") needs only detection of a potentially damaging stimulus plus a reflexive withdrawal. Pain requires nonreflexive conscious awareness. Robert Elwood (2019) gives the working definition: a reflex is a simple, short-term response; any complex, long-term response shows an extended alteration of motivational state and cannot be a reflex.
- Reflexes are faster, hard-wired, automatic, stereotyped, more often innate, need no learning, are less neurohierarchical, and rarely need structures above the spinal cord and brainstem.
- Nonreflexive behaviours are longer-lasting, flexible, more cognitive, involve trade-offs between options, and reflect enduring, motivated states built on higher-order brain structures.
- Rule of thumb. The more neurologically complex and neurohierarchical a behaviour, the less likely it is reflexive and the more likely it is sentient.
Two examples show how elaborate a reflex can be while needing no sentience:
Pupillary light reflex. Light enters the pupil, reaches the retina, retinal ganglion cells, optic nerve and midbrain, and hard-wired outflow constricts the pupil. It takes milliseconds, is involuntary and innate, and works in sleep or even coma.
Knee jerk. In humans it uses just two neurons, one sensory and one motor, joined by a single synapse. No feeling is required.
This matters because single cells such as E. coli and the ciliate Stentor roeselii show surprisingly complicated food-seeking and avoidance. Some researchers read these as intentional and sentient; Feinberg holds they are reflexive sensing without evidence of sentience.
How each behavioural criterion shows nonreflexivity
- Criterion 5, trade-offs. Crump et al. say trade-off behaviour shows that the nociceptive response goes beyond a reflex and involves centralized, integrative processing.
- Criterion 6, flexible self-protection. The animal should target its response to where on the body the stimulus was applied, varying the response as if trying different solutions.
- Criterion 8, analgesia preference. At minimum involves motivated choice of relief and trade-offs between pain and food.
- Criterion 7, learning. The richest domain, set out below.
The learning hierarchy
| Type | What happens | Reflexivity |
|---|---|---|
| Habituation (nonassociative) | Response declines to a repeated stimulus | Most reflexive |
| Sensitization (nonassociative) | Response increases after a noxious stimulus | Most reflexive |
| Classical (Pavlovian) conditioning | A conditioned stimulus (bell) comes to predict an unconditioned stimulus (food) | Less reflexive |
| Operant (instrumental) conditioning | Animal links its own behaviour with a reinforcer | Reasonable indicator of sentience |
| Unlimited associative learning | Compound, multimodal, trace and second-order learning | Very strong indicator |
Eva Jablonka and Simona Ginsburg argue that open-ended (unlimited) associative learning, which needs multimodal discrimination, motivational trade-offs, instrumental goal-directed conditioning, trace conditioning and second-order learning, is a very strong indicator of sentience, always accompanied by integrative brain areas. In humans such learning occurs only with conscious awareness. Operant learning is not itself a "feeling", but it indicates learned valences and memory of enduring, motivated affective states.
Lynne Sneddon et al. (2014) add that increasing complexity of responses to noxious stimuli, beyond detection and reflex, implies some form of experience, and that pain, being multidimensional, needs a multimodal assessment.
Nonreflexivity is the bridge from behaviour to feeling. Every behavioural criterion Feinberg accepts is a way of showing that the animal is not just reacting but is weighing, remembering, locating and choosing, which requires centralized, hierarchical processing.
Other motivated behaviours
Beyond the pain criteria, Feinberg also counts naturally occurring, nonconditioned, goal-directed, affectively motivated behaviours, such as flexible escape behaviours, which arise without any specific learning procedure in the way flexible self-protection does.
The nonreflexive line is fuzzy. Feinberg admits it is not always easy to draw a strict line between reflexive and nonreflexive behaviour. Critics of behavioural criteria point out that complex, flexible behaviour can be produced by systems without experience (from plant signalling to machine-learning agents), and that some reflexes are modulated by learning. The criteria work as probabilistic evidence, which is how Birch frames them, not as proof.
Two subtypes, one emergence
Feinberg's summary claim is that exteroceptive images and interoceptive-affective awareness tend to co-occur in the same organisms, emerge at the same neurobiological and evolutionary stage, and therefore likely coevolved. Two facts support this:
- Shared criterion. Sensory integration (criterion 2) is required for both subtypes.
- Shared infrastructure. Both rely on unique, advanced neural architectures in the central nervous system.
Sentience in both its forms is, for Feinberg, a novel emergent feature of sufficiently complex, variably neurohierarchical central nervous systems, and nonreflexive behaviour is how we detect it from outside.
The trading crab. A hermit crab receives a mild shock inside its shell. If it simply bolts every time, that is compatible with a reflex. Now give it a choice: the shocked shell is a high-quality one, and only a poorer shell is available. If the crab tolerates more shock for the better shell and leaves sooner when the alternative improves, it is weighing a cost against a benefit. Elwood's experiments with hermit crabs did roughly this, and it is the kind of evidence criterion 5 is designed to capture.
For model-risk teams the structure is instructive: no single test of "understanding" or "intent" in an AI system is decisive; Birch's approach is a graded evidence checklist combining architecture (criteria 1 to 3) and behaviour (criteria 4 to 8). The same pattern suits AI capability and welfare assessments, where architectural and behavioural indicators are scored separately and weighed together.
The Stages of the Emergence of Sentience: General Principles
Feinberg reduces "neurobiological complexity" to four measurable variables (number of neurons, neural specialization, number of hierarchical levels, interaction between levels) and uses them to define three emergent stages: ES1 single cells that sense, ES2 "presentient" animals with early nervous systems, and ES3 animals with complex brains and sentience. The big leap comes between ES2 and ES3, roughly 560 to 520 million years ago.
Sensing is not sentience
The chapter's anchor claim restates the book's thesis in its sharpest form. Basic sensing is an evolutionarily early emergent feature of all organisms, but sensing does not constitute sentience. Sentience is a neurobiological feature of sufficiently evolved and complex brains; it evolved from single-celled life, yet took as much as 3 billion years to emerge. Feinberg still calls sentience an emergent feature of both sensing and life, a theme he returns to in chapters 10 and 11. To trace the road, he needs markers.
Complexity, defined narrowly
NBE proposes that the magnitude of four objective neurobiological and evolutionary variables determines the emergence of sentience.
- Number of neurons. An increasing number of neurons in the nervous system.
- Neural specialization. An increasing degree of specialized neural functions.
- Hierarchical depth. An increasing number of neurohierarchical levels.
- Level interaction. An increasing degree of interaction between those levels.
When Feinberg says a brain is "more neurobiologically complex" or "more evolved", he means only that these four magnitudes are larger, and he notes that the four tend to increase together in nervous-system evolution. The list is shorthand: the detailed features appear in tables 5.1, 6.1 and 7.1 in the following chapters.
Complexity has a precise meaning in NBE. More neurons, more specialization, more levels and more cross-level interaction produce more novel emergent features, and past a point those features add up to sentience. The claim is quantitative in spirit even though no thresholds are given.
Support from complexity science
Feinberg aligns this with the general view that emergence is maximized in systems with many interacting parts: the more interacting parts, the greater the chance of novel aggregate features. He also recalls Sneddon et al.'s link between behavioural complexity and nonreflexivity in pain.
- James Ladyman and Karoline Wiesner (2020): "there is no conception of complexity or complex systems that does not involve emergence."
- Melanie Mitchell (2009): systems in which organized behaviour arises without an internal or external controller are self-organizing; because simple rules produce complex, hard-to-predict behaviour, the macroscopic behaviour is called emergent. She offers an alternative definition of a complex system as one that exhibits nontrivial emergent and self-organizing behaviours.
The conclusion Feinberg draws is that system complexity and emergence are interrelated, and that the relation is especially magnified in neurohierarchical systems, one of the foundations of NBE.
Four variables, no metric. The variables are plausible but unweighted and unquantified: Feinberg gives no threshold of neurons or levels at which sentience appears, and bees with about a million neurons end up sentient while some animals with far more neurons in diffuse nets do not. The model works as an ordering heuristic, and chapter 7 shows that architecture (maps, hierarchy, integration) matters more than raw counts.
The three-stage model
The model, built on Feinberg and Mallatt's work from 2016 to 2020 but now centred on emergence, proposes three broadly drawn stages.
- ES1: single-celled organisms. First appearance 3.5 to 3.4 billion years ago. The lowest level of neurobiological evolution: no neurons, no nervous system, basic reflexive sensing. Nonsentient.
- ES2: neurons, nervous systems and evolutionarily early brains. First appearance about 570 million years ago. Mid-level values of the emergent factors; a big advance in sensing over ES1 but short of full sentience. Feinberg calls them "presentient"; their status is more ambiguous than ES1, and some ES2 animals are further along than others.
- ES3: animals with more neurobiologically complex nervous systems and brains. First appearance about 560 to 520 million years ago. The highest levels of the emergent factors and the clearest evidence of sentience.
The evolutionary progression from sensing to sentience makes a substantial leap between ES2 and ES3, and Feinberg treats the personal nature of sentience as a natural result of the same progression.
| Stage | Who | When first | Nervous system | Status |
|---|---|---|---|---|
| ES1 | Bacteria, protists (E. coli, Stentor, Paramecium) | 3.5 to 3.4 billion years ago | None | Sensing, nonsentient |
| ES2 | Early neural animals (cnidarians, simple worms, early bilaterians) | About 570 million years ago | Nerve nets, simple brains | Presentient, ambiguous |
| ES3 | Vertebrates, arthropods, cephalopods and other complex-brained animals | 560 to 520 million years ago | Complex brains, maps, hierarchies | Sentient |
Three features of the model deserve emphasis because the rest of the book leans on them:
- Stages, not switches. The boundaries are "broadly drawn": ES2 in particular is a gradient, with some animals much closer to ES3 than others, so the model predicts borderline cases rather than a sharp line.
- Long gaps between stages. About 3 billion years separate the first cells from the first neurons, while ES2 to ES3 took perhaps only 10 to 50 million years, a sign that once neurons existed, hierarchical elaboration accelerated.
- Cumulative emergence. Each stage keeps all the emergent features of the stage below (life, metabolism, homeostasis, sensing) and adds new ones, so sentience is built on top of life, never apart from it.
The timing places ES3 in the late Ediacaran to Cambrian, the period of the Cambrian explosion, when animal body plans, eyes and predator-prey arms races diversified rapidly. Feinberg and Mallatt have long argued that this is when image-forming vision and complex brains, and hence sentience, first appeared.
The Cambrian-origin view is widely cited but not settled. Since 2023 the "realistic possibility" stance of the 2024 New York Declaration has broadened attention to insects and other invertebrates, consistent with multiple ES3 origins. At the same time, work on cellular basal cognition (for example Michael Levin's programme on bioelectric signalling and problem-solving in cells and tissues) has strengthened claims that single cells and aneural organisms learn and decide, which pushes on the ES1 and ES2 boundaries Feinberg draws. Neither line of work has produced agreed evidence of sentience in aneural organisms.
Turning the dial. Picture a single control panel with four dials: neurons, specialization, levels and interactions. Start a bacterium at zero on all four. Turn up neurons alone and you get a large but flat nerve net, like a jellyfish: more sensing, no images. Turn up levels and interactions as well and you get maps feeding maps that talk back. Feinberg's model claims that somewhere along this second route, and only there, sensing turns into feeling. The thought experiment shows why he refuses to equate sentience with neuron count.
The four variables are a maturity model, the kind architects use for capabilities: size, specialization, layering and integration. Its limits are those of every maturity model: it orders cases well but does not set the threshold, so judgments at the boundary (ES2 animals, or advanced AI systems) still need the case-by-case criteria of chapter 3.
Emergent Stage 1: Single-Celled Organisms and the Emergence of Sensing
Single cells already show every general feature of biological emergence: life itself is an emergent aggregate property of molecules enclosed by a membrane, and sensing comes with it. Feinberg examines E. coli's run-and-tumble chemotaxis and the century-old controversy over Stentor roeselii's "avoidance hierarchy", and rejects Arthur Reber's Cellular Basis of Consciousness: single cells sense, but they do not feel.
Life as the first emergence
All living systems, however simple, are full of emergent features, and even the earliest single cells run processes far more complex than their nonliving ingredients. ES1 organisms meet every feature in table 2.1: many parts, multilevel aggregate functions, and an embodied cell that integrates those parts. So the life of a single cell is an emergent aggregate feature of its atoms, molecules, proteins, membranes and ribosomes and their interactions.
- Who. Single-celled organisms such as the prokaryote Escherichia coli and the eukaryotes Amoeba and Stentor.
- When. Protocells about 3.7 billion years ago; prokaryotes about 3.5 to 3.4 billion years ago.
- Novel structures. Macromolecules (proteins, nucleic acids, sugars, lipids), organelles, cells; embodiment by a semipermeable membrane that concentrates the chemistry of life.
- General novel emergent processes. Life; metabolism converting food to energy (ATP) and new materials, with efficient energy use that slows entropy; homeostasis, keeping the internal environment constant against external change.
- Novel but nonsentient processes related to sentience. Embodiment yields self-organization into a single aggregate cell with most features of table 2.1, and the capacity to sense and respond to the environment.
The feature that matters most for the later story is the capacity to sense the environment, respond to internal homeostatic states, and adapt behaviour to both. Elizabeth Haswell et al. (2011) note that all organisms, from bacteria to plants and animals, must sense mechanical force externally (shear, gravity, touch) and internally (osmotic pressure, membrane deformation). Since only organisms that avoid destructive forces survive, basic sensing must have been present in the earliest protocells, about 3.7 billion years ago.
- 3.7 billion years ago: first protocells.
- 3.5 to 3.4 billion years ago: first prokaryote fossils (no nucleus or internal membranes).
- About 1.5 billion years ago: oldest generally accepted eukaryote fossils (with a nucleus).
- About 2 billion years: time taken for the prokaryote-to-eukaryote transition, a reminder of how slowly major biological novelties arise.
Eukaryote dates are moving. The 1.5 billion year figure is conservative. Fossils interpreted as eukaryotic go back to about 1.6 to 1.75 billion years (for example in China and northern Australia), and molecular clocks often place the last eukaryotic common ancestor somewhere around 1.8 to 2 billion years ago, though estimates vary widely. The 2 billion year gap is an order of magnitude, not a measurement.
E. coli: sophisticated, but reflexive
Bacterial chemotaxis biases movement toward higher concentrations of beneficial chemicals and away from harmful ones. Despite having no neurons, E. coli responds adaptively to nutrients and toxins, oxygen, pH, osmolarity, temperature, and the intensity and wavelength of light. Its signalling pathways are among the best understood in biology: extracellular ligands bind transmembrane chemoreceptors (MCPs), which act through the kinase CheA, the response regulators CheY and CheB, the methylation enzyme CheR and the phosphatase CheZ to set the probability of "runs" versus "tumbles" (after Micali and Endres, 2016).
The motor side is equally elegant. E. coli has 5 to 10 flagella randomly distributed over its surface, each driven by a biological rotary motor:
- Run. In favourable conditions all motors turn counterclockwise, the flagella bundle, and the cell swims smoothly forward.
- Tumble. Heading toward worse conditions, the motors switch to clockwise, the flagella fly apart, and the cell tumbles and changes course at random.
The result is a biased random walk. The bacterium never "chooses" a direction; hard-wired coupling between sensors and motor ensures that it probabilistically ends up in better conditions. Feinberg's verdict: the behaviour is adaptive, as photosynthesis and metabolism are adaptive, but reflexive by the chapter 3 criteria.
Stentor and the "trying" debate
The ciliate Stentor roeselii became famous in arguments over single-cell sentience. Stentor are protozoans (an informal term for single-celled eukaryotes, including amoebas and paramecia), 500 to 1,200 micrometres long, with a trumpet-shaped, funnel-like front, covered in cilia for feeding and swimming. Ciliates as a group have a full array of chemosensing and mechanosensing abilities. S. roeselii can swim freely or sit sessile, anchored by its "foot" inside a tube built of debris and its own mucus.
Jennings, 1902 to 1906. Zoologist Herbert Spencer Jennings squirted carmine powder (a red dye) by pipette in front of the ciliate's mouth and reported a sequenced avoidance hierarchy:
- Bending away from the dye.
- Ciliary alteration, reversing the cilia to change water currents and reduce particles reaching the mouth.
- Contraction into the tube.
- Detachment from the tube and swimming away.
Jennings (1906) admitted steps were sometimes omitted or reordered, but concluded that the organism was "trying" successively different reactions "till one is found that affords relief".
The replication history is mixed:
- 1967 failure. A study published in the behaviourist era could not replicate Jennings, but it used a different Stentor species.
- Dexter et al. (2019). Using S. roeselii, the team could not elicit avoidance with modern carmine, but pulses of polystyrene beads in a sodium azide (NaN3) suspension worked. They observed all four of Jennings's behaviours and, with sophisticated statistics, found that a hierarchy was to some extent present.
- The caveats. They reported "substantial heterogeneity", few full sequences and many partial ones in varying orders, and found the choice between contraction and detachment consistent with "a fair coin toss".
Feinberg reads the Dexter results as showing a statistical probability of adaptive behaviours rather than a strict, nonreflexive hierarchy, which is what sensing without sentience should look like.
Reber's Cellular Basis of Consciousness
Cognitive psychologist Arthur Reber is the strongest advocate of unicellular sentience. His Cellular Basis of Consciousness (CBC) model says single cells, Stentor among his favourites, have not just sensing but consciousness. A bacterium, Reber writes (2019), does not merely sense a sugar gradient: it perceives its positive valence and moves toward it; it interprets an acidic molecule as aversive and retreats in a measured way. Feinberg notes the echo of Jennings's "trying".
| Criterion (chapter 3) | ES1 organisms |
|---|---|
| Basic sensing | Yes, across many modalities |
| Neurohierarchical maps, mental images | No; no neural anatomy at all |
| Valence-appropriate approach and avoidance | Yes, adaptive |
| Infrastructure for interoceptive-affective feeling | No |
| Nonreflexive behaviour | No; fully explained by reflexive biochemistry |
| Verdict | Sensing, not sentience |
Reber (CBC). Sentience is coextensive with life; a bacterium perceives and interprets valence; the cell is the basic unit of consciousness.
Feinberg (NBE). Single cells sense and respond with valence-appropriate behaviour, but lack maps, affective infrastructure and nonreflexive behaviour; their behaviour is fully explained by standard biochemistry.
Basal cognition has grown since 2023. Research on learning and decision-making in aneural systems (slime moulds such as Physarum, bacterial biofilm electrical signalling, habituation in single cells such as Stentor coeruleus and ciliates more generally, and Levin's work on bioelectric problem-solving in tissues) has expanded. Much of this work shows cognition-like information processing without neurons; most of its proponents frame it as basal cognition rather than sentience, which is compatible with Feinberg's line between sensing and feeling.
The thermostat with a history. Give a thermostat a memory: it learns that the house cools fast on windy days and starts heating earlier. Its behaviour is now adaptive, probabilistic and context-sensitive. Few would call it sentient. Feinberg's claim is that E. coli and Stentor sit on the thermostat's side of the line, however impressive their biochemistry: complexity of response alone is not evidence of feeling unless it rides on integrative, hierarchical structures.
Conclusion of ES1
Sensing in single cells has billions of years to go before the degree of emergence needed for sentience is reached. Sentient animals show "a veritable explosion" of novel neural and emergent features absent in ES1. Feinberg nonetheless applauds the link between life and sentience that Reber and others draw: he proposes an uninterrupted, roughly stepwise progression from bacterial sensing to animal sentience over billions of years, with no fundamental gaps.
E. coli is a model of robust control without understanding: a probabilistic policy over two actions, tuned by gradient sensing, produces goal-reaching behaviour. Many production AI agents work the same way, and the chapter is a useful guard against reading intention into adaptive behaviour when explaining systems to regulators or users.
Emergent Stage 2: Neurons, Nervous Systems, and Evolutionarily Early Brains
With neurons came fast signalling, nerve nets, then central nervous systems and simple brains. Feinberg examines four ES2 animals from three lineages, the worm C. elegans, the protochordate amphioxus, and the sea slugs Aplysia and Pleurobranchaea, and finds real neurohierarchical integration and valence circuits, but no image-forming maps and no affective infrastructure. They are "presentient": a "royal road" to sentience, not yet the destination.
From sponges to nerve nets to brains
The first multicellular animals appeared around 700 to 600 million years ago, most likely sponges. Modern sponges have just four cell types, are immobile, and have no nervous system. The next step for sentience was animals with more cell and tissue types, including neurons.
The first nervous systems were probably nerve nets in early jellyfish-like animals and simple marine worms: distributed over the body and largely decentralized, but already a major advantage for feeding, defence and mating. The cnidarians (modern jellyfish and relatives) retain them. In "true jellyfish" there is a diffuse nerve net plus nerve rings; the most elaborate sense organs, the rhopalia, carry pigmented photosensitive ocelli (simple eyes) and statocysts for gravity and orientation, and the net's nerve endings detect touch.
Bilaterians and the first brains
The first worms, early bilaterians (animals with matching left and right sides), had evolved by 570 million years ago. Over the next 30 to 50 million years they gave rise to today's bilaterians, both invertebrates and vertebrates. In one proposed scenario, parts of the nerve net were enlarged and centralized in the head, the end that meets the environment first, while longitudinal nerve cords carried sensory information and motor commands. These were the first brains and nerve cords, the start of central nervous systems. Roundworms, earthworms, flatworms, sea slugs and amphioxus still reflect this stage.
- When and who. First appearance about 570 million years ago; most invertebrates, for example most worms.
- Novel neural structures. A multicellular body with neurons, reflex arcs, sensory receptors and motor effectors (muscles, glands); neurons transmit much faster signals; nerve nets, then central and peripheral nervous systems; some animals have a simple brain with movement-patterning circuits; receptors are mechano-, chemo- and photoreceptor cells.
- Control and connectivity. Rapid signalling lets the animal control a multicellular body; reflex arcs and neuron networks coordinate all its parts.
- Novel processes related to sentience. Significant increases in neurohierarchical exteroceptive processing and approach-avoidance behaviour; basic motor programs and central pattern generators (CPGs) for rhythmic locomotion and feeding; behaviour becomes less reflexive and more complex than at ES1.
Judging sentience at ES2 is harder than at ES1 because the leap in neurobiological features is so large. Feinberg chose "presentient" because the stage spans a broad range, from animals near ES1 to animals close to ES3.
Four ES2 animals
| Animal | Lineage | Neurons | Best sensory feature | Valence circuit |
|---|---|---|---|---|
| C. elegans | Nematode worms | 302 (hermaphrodite) | Amphid sensilla, polymodal neurons | Attractant/repellent interneuron layers |
| Amphioxus | Protochordates, on the vertebrate line | Up to ~20,000 adult; 300 to 500 larva | Four photoreceptor types; frontal eye | Escape via post-infundibular neuropile |
| Aplysia | Gastropod molluscs, on the coleoid line | About 10,000 | Rhinophore olfaction; oral veil touch | Incentive integrator in cerebral ganglion |
| Pleurobranchaea | Gastropod molluscs | Comparable | Oral veil chemo- and touch | Feeding and turn networks |
Caenorhabditis elegans
C. elegans is a transparent nematode about 1 mm long living in soil, where it eats bacteria, the ES1 organisms of the previous chapter. Nematodes are one of the largest phyla: William Schafer notes 20,000 to 30,000 described species, with the true number estimated at 100,000 to 10 million. C. elegans individuals are mostly hermaphrodites, with some males. Its flexible cylindrical cuticle gives shape and protection.
Nervous system. The hermaphrodite has just 959 cells, 302 of them neurons. The centre of the system is a circumoral nerve ring (the brain), with cell bodies in ganglia before and behind it; a ventral nerve cord runs head to tail and connects reciprocally to the ring; a dorsal nerve cord and numerous circumferential commissures complete the layout.
Senses. Chemo- and mechanosensory organs surround the mouth; the main organ is a pair of amphid sensilla, the worm's olfactory and taste organs, each with 12 sensory neurons, which also contain touch and temperature neurons. The worm has interoceptive and nociceptive sensitivity to thermal and mechanical harm.
A layered worm brain. Brittin et al. (2021) mapped the spatial embedding of the C. elegans connectome and found a multilayered, multiscale, modular and nested architecture with sensory computation, integration, sensorimotor convergence and brain-wide coordination. The known feed-forward loop motif among neuron triplets becomes an architectural motif reminiscent of layered cortex and of residual networks in AI. Schafer (2016) similarly speaks of scale-invariant conservation of organisation between the worm and the human brain across many orders of magnitude.
Approach and avoidance. Athanasios Metaxakis et al. (2018) describe integrated, hierarchical processing of positive and negative signals. Distinct sensory neurons (AWC, AWA, AWB, ASH) respond to water-soluble attractants and repellents; layers of interneurons (AIA, AIY, AIB, AIZ) feed motor neurons that produce forward (approach) or backward (avoidance) movement. Many sensory neurons are polymodal, so the worm can cross-modulate several senses and fold in its interoceptive state.
Amphioxus: on the vertebrate line
Amphioxus (Branchiostoma), the lancelet, is a fish-shaped marine animal 4 to 6 cm long that burrows tail-first into sand in warm seas. It belongs to the protochordates, with two invertebrate subphyla: cephalochordates (amphioxus) and urochordates or tunicates (sea squirts). Protochordates have a notochord and a dorsal hollow nerve cord, which in vertebrates become the spine and central nervous system; in amphioxus both persist for life.
Nervous system. The adult brain may have upward of 20,000 neurons; larvae perhaps 300 to 500. Amphioxus is considered the closest living proxy for the ancestral chordate, with likely progenitors or homologues of telencephalic, diencephalic, midbrain and hindbrain structures, and an unpaired frontal eye homologous to vertebrates' paired eyes.
Senses and escape. Thurston Lacalli estimates that amphioxus has more epidermal sensory cell types than C. elegans, including olfactory ones, and four photoreceptor types; the frontal eye is too small and simple to form images. In Lacalli's model of larval escape, the integrative post-infundibular neuropile receives input from the frontal eye, the lamellar body (analogue of the pineal), and epithelial sensory cells entering via the rostral and anterodorsal nerves. Escape is triggered most easily by touching the rostrum, whose fibres converge on the primary synaptic zone and, with frontal-eye input, regulate the response.
Sea slugs: on the coleoid line
Molluscs range from bivalves (scallops, clams, oysters), through gastropods (snails and slugs, 80 percent of mollusc species), to the cephalopods (squid, nautilus, cuttlefish, octopus); the octopus is an ES3 animal. Aplysia ("sea hares") comprise 37 species, from a few centimetres to over 60 cm. They are famous in neuroscience through Eric Kandel, who shared the 2000 Nobel Prize for work on learning and memory, much of it in Aplysia californica.
Nervous system. About 10,000 central neurons in nine major ganglia: paired buccal, cerebral, pleural and pedal ganglia plus a single abdominal ganglion, joined by commissures and connectives. The abdominal ganglion runs heartbeat and breathing; the buccal ganglia run feeding. Sensory pathways are chains of only one to three neurons, much shorter than in cephalopods, though the cerebral ganglion receives input from all others and could provide some central control.
Senses. Tiny pinpoint eyes too small to form images, limited hearing, but strong chemosensation and touch. The posterior tentacle, the rhinophore, is the main olfactory organ. The oral veil shows some peripheral somatotopic encoding, but no central somatotopic map has been found.
Valence circuit. Rhanor Gillette and Jeffrey Brown (2015) describe, in Pleurobranchaea californica, an "integrator circuit for incentives" in the cerebral ganglion that assigns positive or negative valence to incoming signals. It drives nearby central pattern generators, which in turn command the downstream pedal ganglia to produce feeding approach or avoidance turns, with switch neurons and corollary discharge neurons coordinating the choice.
Sensing or sentient?
- C. elegans: presentient. No image-forming eyes or hearing, so no distance senses and no neural basis for mental images. It has nociceptors and responds to opioids such as morphine, but nociceptors are sparse and it lacks multisynaptic central affective infrastructure. Behaviour is tied to current stimuli.
- Amphioxus: a more evolved presentient brain. Its sensory neurons do not derive from neural crest or placodes, so its senses are not homologous to vertebrate ones; no image-forming eyes or hearing, hence no isomorphic maps. Nicholas Holland and Jr-Kai Yu found hints of touch topography in the larval cord, but unlike vertebrate head-to-tail mapping; most pathways have one or two neurons to a single centre. Only "glimmers" of midbrain, dien-mesencephalon and telencephalon.
- Sea slugs: on a par with or beyond amphioxus, still ES2. Few neurons and senses, small cerebral ganglia, short sensory hierarchies, no full affective infrastructure. Yet their approach-avoidance is beyond ordinary reflexes.
Andrew Barron and Colin Klein (2016) give the sharpest behavioural argument against nematode sentience. Nematodes do not search for food beyond the immediate environment; their behaviour is organised by reference to primary sensory input. In mammals and insects, homeostatic drives send the animal to where resources are expected to be, guided by an integrated, spatial simulation of the environment. Nematodes have memory, but no evidence that it is spatial or builds a model of the world. So they lack the egocentric modelling of the environment required for subjective experience.
C. elegans (Barron and Klein). Behaviour indexed to current sensory input and internal state; memory with no spatial component; no egocentric world model.
Insects and mammals. Behaviour indexed to an integrated spatial simulation; animals travel to where food is expected, not merely where it is sensed.
For sea slugs, Hirayama and Gillette (2012) offer the most generous reading: the interaction of the mollusc's goal-directed feeding network with its turn network forms "a simplest decision module for approach/avoidance", acting at a "precognitive level" in this solitary, cannibal predator, and potentially the core circuitry around which valuation and comparison were elaborated in vertebrates.
The ES2 line is drawn by anatomy, and the evidence is moving. C. elegans shows state-dependent decisions, learned avoidance and responses to analgesics, and some researchers argue for minimal valenced states in it. Gastropods show operant learning (Aplysia feeding studies). On the Birch criteria, several ES2 animals meet more than one criterion. Feinberg's verdict rests heavily on the absence of maps and long hierarchies, which some would weigh less than behaviour.
The 2024 New York Declaration names a "realistic possibility" of consciousness for all vertebrates and many invertebrates, including insects, decapod crustaceans and cephalopods, and calls for research on animals such as nematodes, among others, without asserting it. In 2024, the full connectome of the adult fruit fly brain (about 140,000 neurons, FlyWire consortium) was published, a resource that sharpens comparisons between ES2-style circuits and ES3 insect brains.
Conclusion: the royal road
ES2 animals sense more richly than any single cell, but they lack sensory mental images and the affective infrastructure for interoceptive-affective feeling; their behaviour is less reflexive, which makes them a "royal road" to sentience rather than sentient.
Two contributions make ES2 indispensable:
- Speed and connectivity. Neurons supply the processing speed and connectivity that higher hierarchical levels need.
- Differentiated responsiveness. ES2 sets up nonsentient but increasingly complex, differentiated exteroceptive responsiveness that comes to fruition at ES3.
Getting home without a map. Two delivery robots must reach charging docks. The first follows the strongest signal it currently detects and backs away from obstacles. The second keeps an internal map, remembers where docks were last time and heads there even when no signal is present. Both reach docks most days. Barron and Klein's argument is that only the second has the structure that subjective experience requires: a model of a world it is located in. C. elegans, on this view, is the first robot.
The Brittin finding that a 302-neuron brain shows residual-network-like layered motifs is a caution against equating architectural resemblance with capability. The same motif appears in a worm and in deep networks; what it computes depends on the depth, maps and integration around it, which is Feinberg's whole point.
Emergent Stage 3: Animals with More Neurobiologically Evolved Nervous Systems
ES3 is where sensing becomes sentience, between about 560 and 520 million years ago, and it happened at least four times: in vertebrates, coleoid cephalopods, arthropods and, Feinberg's new addition, onychophorans (velvet worms). In each line he finds the same package: centralized topographic maps that make sensory images, an affective infrastructure for valenced feeling, and nonreflexive, flexible behaviour. Sentience is the clearest example of biological emergence: a novel aggregate process of a complex, hierarchical system of living neural elements, none of which is sentient on its own.
From presentient to sentient
The road so far:
- Life first. The pathway starts with life, and the emergent features of life keep contributing at every later stage.
- ES1. Single cells add reflexive sensing and approach-avoidance, but are nonsentient.
- ES2. Presentient animals have more evolved brains and richer sensing, but no infrastructure for sensory images, limited hierarchical integration, poorly specialized affective circuits, and mostly reflexive behaviour.
- ES3. Brains evolved enough to make a truly monumental transition in how animals experience their world and themselves.
At ES3 the animal has clearly delineated neurohierarchical pathways building centralized topographic maps from several senses (vision, touch, hearing), which create exterosensory mental images; it shows integrated interoceptive-affective sentience in its neural infrastructure and behaviour; and both come with nonreflexive behaviour.
- When and who. First appearance 560 to 520 million years ago: all vertebrates, coleoids (octopus, squid, cuttlefish), all arthropods including insects and decapods (crabs), and onychophorans (velvet worms).
- Novel neural structures and processes. A brain with more neurons (approximately more than 100,000); many differentiated neuron subtypes; elaborated sense organs including image-forming eyes and receptor organs for touch, hearing and smell; a more fully developed neural infrastructure for affect and pain; expanded neural hierarchies with extensive interactions.
- Novel emergent features related to sentience. Centralized topographic maps create exteroceptive "sensory images"; centralized positive and negative valenced affects, including pain beyond nociception; increasingly nonreflexive (volitional) actions and globally directed behaviour.
The 100,000-neuron marker does not fit insects. Table 7.1 lists "approximately more than 100,000" neurons as an ES3 feature, yet Feinberg later gives insect brains as "less than 100,000 to 1 million" neurons and counts all insects as sentient. Small insects (some parasitoid wasps and tiny beetles) have brains with only several thousand to tens of thousands of neurons. The figure is best read as a rough indicator, consistent with chapter 4's point that architecture matters more than counts.
How ES3 fits the general emergence principles
Aggregate system feature of interacting parts
Sentience is built from an ever-increasing number of individually nonsentient parts. What makes the difference is neurohierarchical organization in complex brains. ES1 and ES2 organisms lack both the parts and the interactions needed. Simple sensing occurs in cells with no neurons, in worms with few intervening neurons (C. elegans, amphioxus) and in sea slugs with more hierarchy but simple brains; the evidence for sentience only appears with the complexity of ES3.
ES3 animals also have the most neuron types, and differentiation of neurons and their interactions exponentially increases the potential for novel emergent features. Their sense organs (image-forming eyes, sophisticated hierarchies for hearing, taste and smell) and their expanded affective infrastructure are matched by an explosion in the specialization of brain regions.
Process
Both life and sentience are embodied processes. Ernst Mayr (1982) warned that "life" and "mind" are reifications of activities: "mind" refers to mental activity, life to the processes of living, and treating them as separate entities risks smuggling in something like a soul. William James (1904) made the same point about consciousness: he denied that the word stands for an entity but insisted that it stands for a function, the function of knowing.
Sentience, like life, is something particular living bodies do. Every general biological feature and every special feature of sentient brains is a functional feature of a particular living embodiment, and Feinberg says this is what will later explain why sentience is personal.
Hierarchy and novelty
All sentient animals have neural hierarchies in all sensory domains: lower levels receiving input influence higher levels that direct motor output, far more extensively than in reflex-dominated nervous systems. And sentience itself is the best example of emergent novelty: a novel aggregate system process derived from a neurobiologically complex hierarchical system of living nervous elements, its novelty reached by adding neural features. The late features that create it (huge numbers and differentiation of neuron subtypes, many tightly and reciprocally connected levels) are ideally suited to increase emergent novelty.
Four phylogenetic lines to sentience
The earlier Feinberg-Mallatt model named at least three lines: all vertebrates, coleoids and all arthropods. Feinberg now adds a fourth: onychophorans.
| Line | Example animals | Neurons | Image-forming eyes and maps | Affective infrastructure | Key behavioural evidence |
|---|---|---|---|---|---|
| Vertebrates | Fish, frog, birds, mammals | Millions to billions | Camera eye; tectal and cortical maps | Solid in all classes | Trade-offs, frustration, self-analgesia, place preference, operant learning; túngara frog escape |
| Coleoids | Octopus, squid, cuttlefish | 170 to 500 million (octopus) | Retinotopy in octopus and squid | Inferred: vertical lobe; arm-to-brain nociception | Crook's lidocaine place preference; camouflage; coconut carrying |
| Arthropods: insects | Fruit fly, honeybee, bumblebee | Under 100,000 to 1 million | Compound eyes; fifth-level retinotopy | Mushroom bodies, central complex | Cross-modal recognition, judgment bias, heat trade-off, observational learning |
| Arthropods: crabs | True crabs, hermit crabs | Not given | Compound eyes; multisynaptic pathways | Hemiellipsoid body = mushroom body | Flexible escape; shell choice; vacancy chains |
| Onychophorans | Velvet worms | Not given | Small, low-resolution ocelli with lens | By homology with arthropods | Female-dominated social hierarchy |
Vertebrates
Amphioxus, the closest living proxy for the vertebrate line, lacks the complexity and nonreflexive responding of ES3 animals. Vertebrates have it all.
Exteroceptive sentience: maps and images
Even the most basal vertebrates possess the infrastructure for mapped sensory representations; it is a feature of all vertebrate brains. In mammals the main maps are in the cerebral cortex, in birds in the enlarged corresponding parts of the cerebrum. In fish and amphibians, the midbrain optic tectum holds a finely detailed, point-by-point map of the environment, mostly visual but also from hearing, touch and balance. Feinberg and Mallatt propose that the tectum created image-based consciousness in early vertebrates and that the site shifted from tectum to cerebral cortex during mammalian evolution. Because the tectum gets no direct smell input, smell perception is cerebral in all vertebrates.
The leap from amphioxus to fish traces to the vertebrate "camera eye", which forms a photograph-like retinotopic image that ES3 connectivity can integrate with touch and water-vibration senses. The camera eye needs a focusing lens, which develops from ectodermal placodes, an embryonic tissue unique to vertebrates. Together with the vertebrate-only neural crest, placodes build all the special, sentience-associated sense organs that create mapped representations. These innovations mark a critical turning point in exteroceptive consciousness, over 520 million years ago.
Affective infrastructure
The vertebrate affective system is more diffuse than the tight somatotopy of the exterosensory systems, with innumerable cross-communicating parts, but both subtypes share hierarchical pathways, reciprocal connections between centres and levels, and specialization. Feinberg and Mallatt's comprehensive review found nearly solid or complete evidence for the affective structures (valenced affects including pain and fear-like responses) in bony fish, amphibians, reptiles, birds and mammals.
Behavioural evidence
Feinberg and Mallatt's earlier behavioural criteria closely match the Birch and Crump list.
- Behavioural trade-offs, value-based cost-benefit decisions (for example food versus predation risk near the food). Solid in all vertebrate groups.
- Frustration behaviour, such as aggression after a reward is denied, showing enduring affect. Shown in fish, birds and mammals.
- Self-delivery of pain relievers or rewards. All vertebrate groups.
- Approaching reinforcing drugs (amphetamines, ethanol) or conditioned place preference. All vertebrate groups.
- Operant conditioning with positive or negative outcomes. Fish, amphibians, reptiles, birds and mammals.
All five are nonreflexive in the chapter 3 sense. A sharper test is to compare escape behaviour, which is intrinsically valenced, in an ES2 relative and an ES3 descendant.
The túngara frog (Bulbert et al., 2015). The ground-dwelling túngara frog, Engystomops pustulosus, faces two predators: snakes from the ground and bats from the air. Presented with models, frogs fled away from the snake but moved toward the bat, effectively undercutting the bat's flight path and making the attack less effective. The authors concluded the frog has adaptive flexibility of escape strategies. Compared with amphioxus's escape circuit, this is plainly nonreflexive.
Verdict. All vertebrates, even basal amphibians such as the frog, show the full ES3 package: a full array of senses, an affective infrastructure, and nonreflexive sensory and affective behaviour.
Coleoid cephalopods
The coleoids are the octopuses, squids and cuttlefish. The octopus has received the most attention, and the clear emerging consensus is that it is sentient.
The octopus nervous system
- 170 to 500 million: neurons in the octopus nervous system (estimates vary).
- About 50 million: neurons in the central brain, one of the largest among invertebrates.
- 3 divisions: the brain, the optic lobes, and the arm nervous system.
- More than 30: differentiated, fused lobes making up the "brain".
- Up to 65 million: neurons in each optic lobe, the largest of the fused ganglia, which also serves motor control and memory.
- Over half a million: neurons in the chromatophore lobes of Octopus vulgaris.
Exteroceptive sentience
The octopus has the largest population of sensory receptors of any cephalopod. Pungor et al. reported retinotopic visual processing in the octopus optic lobe, comparable in some respects to other species, and Chung et al. reported clear-cut retinotopic organization in the squid. Evidence for visual mental images is solid; evidence for somatotopy in other senses is less clear.
Affective infrastructure
Behavioural evidence is strong but direct anatomical evidence is thinner. Nociceptors are well documented; the higher regions supporting affect must be inferred. The prime candidate is the "frontal-vertical lobe" (vertical lobe) system. Shigeno et al. (2018) compare it to the vertebrate pallium (cortex), hippocampus and amygdaloid complex, calling the vertical lobe the best candidate for a pallium analogue in molluscs, whether by common descent or convergence.
Robyn Crook's octopus experiments (2021). Crook recorded from the brachial connectives linking the arm nerve cords to the brain. Injecting acetic acid into an arm produced prolonged, sustained activity at several sites in the connective; injecting lidocaine at the same site abolished it completely, evidence that nociceptors connect to integrative brain regions (Birch criterion 3).
Conditioned place preference. Crook first found which of three chambers each octopus preferred. Pairing the preferred chamber with acetic acid injection made octopuses avoid it, compared with saline controls. Then, when animals received lidocaine over the injection site, they preferred the lidocaine-paired chamber only if they had previously received acetic acid. The pattern of avoiding pain and seeking relief shows goal-directed behaviour motivated by pain, supporting criteria 4, 7 and 8.
Birch et al. found high or very high confidence for criteria 4 and 6 to 8 in octopods, and a clear consensus among cephalopod researchers that octopods and cuttlefish have associative learning.
Nonreflexive behaviour
Compared with the sea slug's approach-avoidance circuit, cephalopod behaviour shows a dramatic increase in complexity and nonreflexivity. Schnell et al. (2021) credit their cognitive suite with behavioural flexibility that lets them innovatively modify behaviour in foraging, anti-predator and mating contexts.
- Camouflage. Chromatophore organs in the skin change colour under a hierarchically organized set of brain lobes. John Messenger (2001): at the top, the optic lobes select motor programmes (body patterns) from visual information; at the bottom, motoneurons in the chromatophore lobes execute them; receptors for all classical neurotransmitters are present; and the system apparently operates without visual or proprioceptive feedback.
- Texture and mimicry. Cephalopods also control skin papillae to change texture, matching rocks or floating algae; some octopuses mimic sponges, several fish species and even sea snakes.
- Coconut carrying. Julian Finn et al. watched more than 20 veined octopuses (Amphioctopus marginatus) over nine years. They lived in coconut shell halves, gastropod shells or human refuse, reoccupied them when flushed out, and carried stacked coconut halves up to 20 metres for later use, which the authors call tool use. However it is labelled, the multistep sequence is directed at a future goal and cannot be reflexive.
Verdict. Cephalopods are sentient ES3 animals with well-developed senses (at least retinotopy in vision), good but not absolute evidence for pain infrastructure, and a range of nonreflexive, purposeful behaviour.
Insects
Insect brains have under 100,000 to 1 million neurons, far fewer than vertebrates' millions, yet they show many of the features needed for sentience.
Exteroceptive sentience
The insect brain is well differentiated, with distance senses including high-resolution vision, olfaction and hearing, plus taste and touch, many with somatotopic organization. All insect senses have multilevel serial pathways to the brain, some reaching a fourth or fifth synaptic level before the forebrain. Johannes Seelig and Vivek Jayaraman showed that the fruit fly's retinotopic pathway reaches the fifth level of the visual hierarchy in the central complex, a region for higher multisensory processing and behavioural decisions, comparable to the at least fifth-order visual pathways in mammals. Insect hierarchies also have reciprocal interactions between levels.
Cross-modal bumblebees (Solvi et al., 2020). Bumblebees trained to tell cubes from spheres by touch in the dark could then recognise them by sight alone, and vice versa. The authors could not say whether bees stored whole shapes or local features (curved or flat edges), but recognition across temporally and physically distinct inputs requires a modality-independent internal representation: insects, like large-brained animals, build "a complete, globally accessible, gestalt perception" of the world.
Affective infrastructure
Two regions are implicated:
- Mushroom bodies. Central to sensory integration, learning and memory, and to assigning positive or negative valence to sensory representations such as odours in Drosophila.
- Central complex. Midline structures integrating sensorimotor, memory and learning functions, including the fan-shaped body (FSB). Hu et al. found FSB neurons activated by leg shock; inhibiting a subpopulation reduced avoidance of shock and noxious heat, and optogenetic activation increased conditioned avoidance, for both learned and innate responses. The central complex has the stronger evidence for pain processing.
- Gibbons et al. reviewed the Birch criteria and gave very high confidence that adult Diptera (flies, mosquitoes) and Blattodea (cockroaches, termites) have integrated nociception (criterion 3).
Behavioural evidence
Pessimistic honeybees (Bateson et al., 2011). Honeybees vigorously shaken to simulate a predator attack were more likely to treat ambiguous odours as predicting punishment, a negative cognitive bias consistent with a negative emotional state.
Optimistic bumblebees (Solvi et al., 2016). Bees learned blue card = reward, green = none. Half then got an unexpected sucrose treat; faced with an ambiguous blue-green card placed midway, treated bees entered faster, a positive judgment bias. The drug fluphenazine, a dopamine blocker, eliminated the bias.
No pain, no gain (Gibbons et al., 2022). Bumblebees chose between high-quality feeders (40 percent sucrose) and lower-quality ones on coloured heating pads. When high-sucrose feeders were heated to noxious temperatures, bees avoided them while the unheated alternatives were also rich, but increasingly fed from heated feeders as the unheated sucrose concentration fell. The trade-off rested on memory of feeder values, not on immediate sensing.
Lars Chittka and Clint Perry (2019) attack the view that small brains confine insects to live "in the moment" with hard-wired routines triggered by precise stimuli. Especially in Hymenoptera (bees, ants, wasps) they document problem-solving, flexibility, foresight, planning and even tool use beyond fixed routines.
Learning by watching (Loukola et al., 2017). Bumblebees had to roll a ball to a central target for sucrose. Bees that watched a live demonstrator learned better than those that saw a "ghost" demonstration (ball moved by magnet) or none. Better still, observers did not copy exactly: they rolled the ball closest to the target, even if it was a different colour from the demonstrated one, a sign of cognitive flexibility.
The bee evidence has grown. Since the book, Dona et al. (2022) reported object play in bumblebees (rolling balls with no reward), Chittka's The Mind of a Bee (2022) summarised the case for bee sentience, and the 2024 New York Declaration on Animal Consciousness explicitly listed insects among animals for which conscious experience is a realistic possibility. Gibbons, Crump, Chittka and colleagues' review of insect pain (2022) is the source of the Diptera and Blattodea ratings.
Crabs
The decapod crustaceans (crabs, lobsters, shrimp) number up to 15,000 species, about half of them crabs. "True crabs" are the infraorder Brachyura; Anomura, the sister group, includes hermit crabs.
Senses. Brachyurans have complex multisynaptic pathways in several domains, including high-grade compound-eye vision, and hierarchical olfactory, chemosensory and mechanosensory systems. Brachyuran and anomuran nervous systems are largely similar in organization and complexity.
Affective infrastructure. Nicholas Strausfeld has argued convincingly that the crustacean "hemiellipsoid body" is the homologue of the insect mushroom body, so Feinberg treats it as the decapod mushroom body. Given that homology plus the behaviour, the evidence for affective sentience in both insects and true crabs is strong. Birch et al. and Crump et al. had high or very high confidence for Brachyura and high confidence for Anomura.
Escape in Neohelice granulata. Daniel Tomsic's lab at the University of Buenos Aires studies this semiterrestrial burrowing crab of the Southwest Atlantic intertidal zone (formerly Chasmagnathus granulatus). Facing a "looming stimulus", the crab first freezes; if the threat approaches it makes a "home run" to its burrow and enters if danger escalates; with no burrow, it freezes, runs, then raises its claws. Crabs constantly adjust speed and direction to the changing visual flow, either changing course without turning or rotating to fixate the predator with one eye's lateral pole and running sideways. Though largely innate, the behaviour is modified by time of day, season, predator abundance, and learning and memory.
Hermit crabs. Robert Elwood of Queen's University Belfast has studied them for decades. With soft abdomens, they live in empty gastropod shells, and their shell behaviour is rich: seeking larger shells as they grow, judging fit and weight, and fighting multivariable "contests" for a better shell. Elwood documents motivational trade-offs between avoiding noxious stimuli and other needs, attending to injured sites, and negative motivational change toward shells where they were shocked.
Vacancy chains (Rotjan et al., 2010). In a vacancy chain, the lead individual takes a new resource unit and vacates its old one, which the next individual takes, and so on (in humans: jobs, cars, housing). In hermit crabs a crab inspects a vacant shell; if it is too big, it keeps its own and a line forms behind it. Rotjan et al. describe "queuing", size-ordered lines where the largest crab holds the empty shell and each smaller crab grips the shell in front, and "piggybacking", lines not in size order and with no vacant shell. When the lead crab moves in, everyone trades up. Feinberg reads this as goal-directed, motivated, nonreflexive and positively valenced social behaviour.
Verdict. Both Brachyura and Anomura meet the criteria: exteroceptive and interoceptive-affective sentience and nonreflexive, including social, behaviour.
Small errors in the source. The book spells the crab "Neohelice granulate"; the species is Neohelice granulata. The 2011 shaken-honeybee study is by Melissa Bateson and colleagues (the book has "Batesen" in one place). Neither affects the argument.
Onychophorans: velvet worms
Velvet worms (phylum Onychophora) are by most estimates the closest living relatives of arthropods, having diverged in the Cambrian, over 540 million years ago. The proposed Panarthropoda groups Onychophora, Arthropoda and Tardigrada ("water bears").
- Brain. At least a bipartite (two-segmented) brain; Christine Martin et al. report that onychophoran mushroom bodies and central bodies are homologous with those of arthropods.
- Touch. Body and feet covered with papillae, each tipped with a bristle connected to touch-sensitive nerve cells.
- Smell and taste. Sensilla around the lips and two antennae; antennal nerves feed bilateral olfactory lobes of about 80 olfactory glomeruli plus large macroglomeruli.
- Vision. Simple but image-forming eyes (ocelli) behind each antenna, each with lens, cornea and retina and an optic nerve to the brain, but small and very low resolution.
- Affect. Not specifically investigated; inferred by homology with insect and crab brains.
A female-dominated society (Reinhard and Rowell, 2005). Euperipatoides rowelli lives in groups of up to 15 females, males and young. Hierarchy forms through aggressive-dominant and passive-subordinate behaviour; animals from other logs meet intense aggression, perhaps from kin recognition. They hunt collectively, but the first to feed is always a female, who eats alone for 45 to 60 minutes, in bouts of up to 10 minutes, attacking, biting, kicking and pushing away other females waiting near the prey. Only then do the other females feed, followed by males and young.
Verdict. Primitive but image-forming eyes, an affective infrastructure likely by homology, and complex, nonreflexive social behaviour: Feinberg proposes that velvet worms are sentient.
The weakest of the four lines. Velvet-worm sentience rests on homology plus social behaviour, with no direct tests of the chapter 3 criteria for pain or affect. Social hierarchies also occur in animals whose sentience few would claim on that basis alone, and dominance and kin discrimination can be mediated by chemical cues and fairly fixed routines. Feinberg's own phylogeny allows that sentience may have arisen once in the panarthropod ancestor rather than separately in velvet worms and arthropods, in which case "four lines" may really be three origins.
Summary: what ES3 shows
Feinberg draws four conclusions.
- The essence of NBE. The increase in general emergent features in complex biological systems corresponds strikingly to the increase in emergent features in sentient brains. Sentience, despite its unique qualities, has all the properties of an emergent process in table 2.1: a novel process of a complex, neurohierarchical system of living neural elements, reached by adding system features absent in the parts.
- Retention as well as addition. Evolution keeps basal emergent features (life, process) while adding new ones, so no single biological, neural or evolutionary factor explains sentience; many variables must be combined.
- Acceleration. About 3 billion years separated ES1 from ES2, but only 10 to 50 million years separated ES2 (Precambrian) from ES3 (Cambrian). Once the neural factors began at ES2, their evolution sped up the emergence of sentience along with hierarchically advanced brains.
- Nothing mysterious. There is nothing unusual about the physics of neurons and nothing unexplainable about how sentience fits standard biological emergence.
Convergence is the strongest argument in the chapter. Vertebrates, cephalopods and arthropods evolved their brains independently, yet each arrived at maps, memory centres, premotor centres and affective circuits in similar functional arrangements (figure 7.2). Independent arrival at the same architecture suggests it is what sentience requires, not an accident of one lineage.
Policy has caught up with the science Feinberg reviews. The UK Animal Welfare (Sentience) Act 2022 covers all vertebrates, cephalopods and decapod crustaceans after the Birch report. The 2024 New York Declaration on Animal Consciousness extended a "realistic possibility" of consciousness to all vertebrates and many invertebrates, including insects. Feinberg's ES3 roster (vertebrates, cephalopods, arthropods) largely matches that consensus; velvet worms remain outside most formal assessments.
The chapter ends with a pointer: to explain how sentience can be a natural, personal (subjective) system feature of certain brains, Feinberg must clarify what "emergent levels" are, what is happening at the micro-macro level of brains. That is chapter 8.
Swap the brain, keep the feeling. Imagine you could watch three animals hunt in the same rock pool: a goby, an octopus and a crab. Each tracks prey visually, weighs risk against reward, learns from injury, and adjusts on the fly. Their brains share no common sentient ancestor; their neural parts are arranged differently. If all three feel, then feeling is multiply realizable across nervous-system designs, tied to organisation (maps, hierarchy, valence, flexibility) rather than to any particular tissue. This is the strongest empirical support for NBE and, notably, the main premise of functionalist arguments about machine consciousness too.
For AI evaluation, ES3's lesson is convergent architecture plus behavioural flexibility. Assessments of advanced AI systems (for example, recent work applying indicator properties drawn from neuroscientific theories) follow the same logic: look for architectural indicators and nonreflexive, trade-off-sensitive behaviour, and weigh them together. Feinberg would add that his list also requires life and embodiment, which current AI lacks.
Neurobiology, Emergent "Levels," and Sentience
Where in the brain does sentience emerge? Feinberg's answer is nowhere in particular: it is a system feature, not a level feature. He rejects "emergence at the top" (Sperry) and strong emergence (Chalmers) as routes to dualism, rejects the "layer cake" picture of levels (Kim), and replaces them with a diversity of hierarchies (serial nonnested maps, scalar nested hierarchies, hierarchical modular networks, nested oscillations) knitted together by upward, downward and circular causality.
The question of where
The ultimate aim of the book is a scientific scenario that runs seamlessly from biology to sentience, including its personal nature. That requires knowing how and where, spatially and temporally, sentience emerges in the brain, which Feinberg calls one of the hardest and most consequential parts of the whole problem.
Sperry and "emergence at the top"
Because sentience needs a complex hierarchical nervous system, it is tempting to think that as levels are added to the neuraxis, consciousness appears "at the top" or at the "highest level." Feinberg calls these "emergence at the top" theories.
The prototype is Nobel laureate Roger Sperry's emergent interactionism. Sperry (1990) held that consciousness is a dynamic emergent of brain activity, neither identical with nor reducible to neural events, and not an epiphenomenon: it sits "in a position of top command at the highest levels in the hierarchy" and exerts control over biophysical and chemical activity at subordinate levels. He described it as putting "conscious mind back into the brain of objective science."
Feinberg suspects the appeal comes from the strictly nonnested sensory pathways (vision, touch), which give the impression that a complete "neural somatotopic model" culminates in cortex. He gives three objections:
- An illusion even for exteroception. Sensory pathways do not end in a single summit where experience happens; their maps work as parts of a wider system.
- Wrong for affect. Interoceptive-affective feelings such as emotion are not strictly hierarchical in the same way, so "top" has little meaning for them.
- Dualism. Although Sperry denied being a dualist, a mind that controls subordinate physical levels looks like Descartes's res cogitans acting on res extensa.
Sentience needs successive addition of structures and levels, but evolution also retains basal emergent factors (life, process) active at every level. Like life, all the "lower" and "higher" parts contribute, but they do not function as if "layered" one upon another. Sensing, present at all stages, is "a beginning, but not the end."
Strong emergence and downward causation
Feinberg's main complaint about emergence at the top is that it makes sentience look like an immaterial feature that "pops out" of a physical brain, which does not explain the explanatory gap but deepens it.
Antti Revonsuo (2010) summarises the strong emergentist intuition: subjective, qualitative experience exists only from the first-person point of view, physical mechanisms have nothing subjective about them, and getting one from the other is "like trying to squeeze wine out of pure water." Feinberg extracts three puzzles from this:
- How does sentience emerge?
- How does a material brain create something with ontologically subjective existence?
- How can mechanisms with nothing subjective in them produce qualitative "feels"?
Chalmers's strong emergence
David Chalmers (2006) distinguishes the two kinds of emergence sharply. Strongly emergent phenomena are not deducible from the complete distribution of particles and fields plus the laws of physics; their existence would require new fundamental laws of nature. Chalmers thinks there is exactly one clear case: consciousness. He does not say what the new laws are or at what level of physics they act. His position elsewhere is "naturalistic dualism": conscious states "naturally supervene" on physical systems but are ontologically distinct and irreducible.
Feinberg calls this "having your cake and eating it too": the brain creates mental states, yet their subjective aspects are in some ontological sense independent of it. The same dualist dilemma Sperry faced.
Downward causation, strong and weak
Sperry (1991) framed his view as a shift from conventional microdeterminism to "macromental determinism" with top-down emergent control, under labels such as emergent interaction, downward causation, macro or holistic determinism, and called it "the key to the consciousness revolution." This is a solution to the problem of mental causation, which Revonsuo defines as the mind's having causal powers to change purely material processes in the brain.
Strong downward causation (Chalmers 2006). The causal impact of a high-level phenomenon on low-level processes is not deducible even in principle from initial conditions and low-level laws. Chalmers finds it coherent but does not know of any actual case.
Weak downward causation. The high-level impact is deducible in principle but unexpected. This is the only kind NBE needs, and it is ubiquitous in biology as constraint.
The NBE reply
Feinberg denies any need for "radical" or "strong" downward causation, just as he denies strong upward emergence. Emergence is universal in complex biological systems; neural hierarchies greatly maximize its degree. The apparent "interaction" between subjective experience and objective brain is an illusion.
The brain and its subjective features are aspects of the same emergent aggregate system, and they are not dissociable. Sentience has causal powers only within that system, and those powers are personal to the individual animal. There is therefore no "gap" and no causal interaction between brain and mind as if they were two things.
Two features, Feinberg argues, make dualism unnecessary and set NBE apart from other emergentisms:
- Standard mechanisms. Sentience's emergence is explained by standard biological and evolutionary principles that apply to any emergent process, so no special kind of emergence is needed.
- A traced progression. Applying the same principles across nonsentient (ES1), presentient (ES2) and sentient (ES3) stages yields a progression that is step-wise but neurobiologically and emergently seamless; sentience "doesn't suddenly pop out."
Against Chalmers, Feinberg holds that no new physical principles are needed: the emergence of sentience is a difference of degree in standard emergence, not of kind, and the massive quantitative increase in emergent processes gives the false impression of a qualitative explanatory gap.
Does "degree, not kind" answer Chalmers? Chalmers's claim is epistemic and modal: facts about experience are not deducible from physical facts, however many. Showing that neural emergence is ordinary and graded does not by itself show that the phenomenal facts are deducible. Feinberg's real answer comes in chapter 10, where he argues that the non-deducibility is a consequence of experience being personal, not of missing physics. Readers should judge the argument there, not here.
What are "levels"?
The concept of levels of organization is contested. Daniel Brooks, James DiFrisco (spelled "DiFresco" in the book) and William Wimsatt, editors of Levels of Organization in the Biological Sciences, note that theories of levels differ methodologically, epistemologically and ontologically, and "levels" and "hierarchies" are often used interchangeably without clarity. Brooks (2017) doubts any universal usage will emerge: "levels" expresses knowledge contextualized within a particular system, its methods and questions.
The layer cake
The older picture, which Brooks describes, is a vertically stratified world in which each horizontal slice "makes up" the next. Jaegwon Kim (1992) endorsed a version: when material processes reach a certain complexity, genuinely novel and unpredictable properties emerge, cumulatively, yielding "a layered structure," each level emergent from and dependent on the one below. Feinberg accepts the complexity-to-novelty part but rejects the layering for nervous systems.
The many kinds of neural hierarchy
If sentience results from numerous diverse hierarchical levels that are mutually nondissociable, how do they relate to it? Feinberg's answer: sentience emerges in successive stages of increasing neurohierarchical complexity and novelty at multiple levels, through an explosive increase in emergent processes (table 4.1), but these are system features that do not emerge at any given level.
Claus Hilgetag and Alexandros Goulas (2020) show how loosely "hierarchy" is used even within one paper: laminar-specific projections, topological projection sequences, position along the posterior-anterior axis, ordered functional complexity, gradients of cortical thickness, laminar projection patterns. These are very different matters; failing to separate them "is bound to result in confusion."
Feinberg groups neural hierarchies into two partially overlapping families, often present together.
- Serial (nonnested) isomorphic pathways. Higher levels are not made of lower ones; lower levels project to higher ones. Levels are defined by synaptic order (primary, secondary, tertiary), usually caudal to rostral, "upstream" to "downstream". Examples: retinotopic vision, somatotopic touch (adjacent body parts mapped by adjacent neurons), tonotopic hearing (tones ordered by position on the cochlea). Essential for exteroceptive sentience.
- Scalar (compositional) nested hierarchies. "Higher" levels have more parts and are physically composed of lower levels, which are nested within them; neuronal differentiation increases with scale.
- Hierarchical modular networks (HMNs). "Modules within modules" at ever larger scales; some modules have central hub nodes; lower modules may also connect to higher ones serially, so HMNs combine nested and nonnested arrangements.
- Temporal scales. Hierarchies also operate at different time scales: Wolf Singer describes a "fantastic intertwining" of processes in which higher frequencies are more local, spanning under 0.1 Hz to 30 Hz and more.
- Caudal-rostral organization and synchronized oscillations. All sentient species have, at minimum, caudal-rostral neurohierarchy, serial hierarchies, scalar micro-macro levels, HMNs and synchronized oscillations.
Olaf Sporns (2011) makes the distinction precise. In hierarchical modularity, levels are levels of modularity, not stages of processing: the brain's grouping of neurons into local populations, regions and large-scale systems resembles a network that can be recursively subdivided into ever smaller and denser modules, with important consequences for the temporal structure of brain dynamics.
| Hierarchy type | How levels are defined | Example | Main role in sentience |
|---|---|---|---|
| Serial nonnested | Synaptic order along a pathway | Retina to LGN to V1 | Exteroceptive maps and images |
| Scalar nested | Number of parts; composition | Neurons in columns in areas | Micro to macro integration |
| HMN | Modules within modules, hubs | Cortical networks | Both, plus global integration |
| Temporal | Oscillation frequency and locality | Under 0.1 Hz to over 30 Hz | Binding across scales |
Feinberg's key observation: in none of these patterns is there a "top" where the parts "come together." Even if neurons are the "bottom", lower elements both create and interact with higher aggregate levels in a circular pattern, so a clean story of part-to-whole causation is hard if not impossible to discern.
Upward, downward and circular causality
Instead of layered levels, Feinberg looks at directions of causation. Kim (2000) lists three: same-level causation, downward causation (a higher property causes a lower one) and upward causation (a lower property causes a higher one).
- Upward ("bottom-up"). From lower to higher levels; the clearest case is the nonnested serial sensory pathways converging rostrally, creating exteroceptive feelings.
- Downward ("top-down"). From higher to lower levels; critical too, though less discussed.
- Same-level. Interactions within a level.
These directions do not mean that novel features emerge at a higher or lower level; causal interactions run every way, so sentience is a system feature, not a level feature.
Downward causation as constraint
Biological systems theory usually calls top-down effects constraints: higher levels impose control over lower ones (distinct from the evolutionary sense of constraint as bias on selection). H. H. Pattee (1970) put it at the centre of theoretical biology: life is defined by hierarchical constraints that "harness matter to perform coherent functions", at every level from molecule to brain: in the origin of life, in development, in evolution, and in the brain, "where there appears to be an unlimited possibility for new hierarchical levels of description."
- Anatomical descending pathways. Rostral ("higher") to caudal ("lower") projections, some reciprocal or synchronous.
- Macro-to-micro scalar causality. Within nested hierarchies, the whole shapes its parts; HMNs combine routes 1 and 2.
- Constraint. Higher levels and emergent processes restrict the degrees of freedom of lower ones, harnessing them to coherent function, as in all emergent biological systems.
Circular causality
Paul Nunez (2016): brains are complex systems in the sense of complexity science. Small-scale systems act bottom-up to generate larger emergent systems with novel properties, which then act top-down on the smaller systems, completing a loop of circular causality.
So hierarchical systems acquire novel properties both via upward emergence and via downward constraint, and the greater degree of upward, downward and circular interaction in sentient brains geometrically magnifies the emergent features of the whole. The directions are so intertwined and mutually causal that "direction of causation" and "levels of emergence" become at best ambiguous and at worst a heuristic abstraction.
Find the top of a whirlpool. A whirlpool in a river is made of water molecules, yet it also forces each molecule into a spiral path. Asking at which "level" the whirlpool exists, or which molecule is "at the top", gets no grip: it is a pattern of the whole that its parts both create and obey. Feinberg's claim is that sentience is like this, but with many interlocking whirlpools at different scales and speeds (maps, modules, oscillations), so there is no summit for a Sperry-style mind to sit on.
Theories of consciousness have tested "where" directly. The 2023 adversarial collaboration between Integrated Information Theory (IIT) and Global Neuronal Workspace (GNW) theory (Cogitate consortium, results announced in June 2023 and published in Nature in 2025) contrasted posterior "hot zone" predictions with prefrontal ignition predictions; the results challenged key predictions of both theories rather than crowning either. That outcome fits Feinberg's view that no single "top" region generates experience, though neither theory is NBE.
Summary
Neurobiological emergence comes in many forms:
- Hierarchical architecture. Serial nonnested isomorphic pathways, scalar nested hierarchies, and HMNs combining both.
- Temporal integration. Mechanisms built on wave frequencies at different scales.
- Causal direction. Same-level, bottom-up, top-down (constraint), and circular causality.
Therefore sentience does not emerge at the "top" of the brain; it is an aggregate emergent system property. Chapter 9 compares NBE with rival theories; chapter 10 returns to its personal nature.
Architects will recognise the move. Asking which layer of a platform "has" resilience or security is a category error: those are system properties produced by interacting layers, feedback and constraints. Feinberg's rejection of "emergence at the top" is the same argument applied to brains, and a useful corrective when people ask which component of an AI stack "understands".
Neurobiological Emergentism: Panpsychism, Biopsychism, the "Emergentist Dilemma," and IIT
Feinberg tests NBE against three neighbours. Panpsychism and biopsychism both push consciousness down (to all matter, or to all life) because they think emergence cannot produce it; Feinberg argues the barrier they react to is imaginary, and that Reber's "emergentist's dilemma" becomes an emergentist's solution. Integrated Information Theory (IIT) turns out to be an emergence theory in disguise, yet it differs from NBE on how low the bar is, on continuity versus stages, and on whether the substrate matters.
Panpsychism
Panpsychism is a family of views sharing the claim that consciousness is a fundamental feature of the world. Philip Goff et al. (2021) define it in its contemporary form as the view that mentality is fundamental and ubiquitous in the natural world. Combined with the usual assumption that fundamental things exist at the micro level, it entails that some micro-level entities have mentality and are found throughout the material universe. It need not say that a rock has a mind, only that the rock's fundamental parts do.
In emergence terms, panpsychism is the opposite of "emergence at the top": the basic elements of consciousness exist even below the level of life. John Heil (2019) puts it vividly: conscious qualities are not emergent, not epiphenomenal, not add-ons; they "go all the way down." Electrons may lack minds, but just as they have charge and mass, they have experiential qualities; the brain has a "Technicolor phenomenology because its parts do."
The panpsychist anti-emergence argument. Goff et al. argue that science need not assume consciousness must be explained by non-conscious processes such as facts about non-conscious neurons. The panpsychist alternative is to explain human and animal consciousness in terms of more basic forms of consciousness, postulated as properties of the fundamental constituents of matter, perhaps quarks and electrons.
Feinberg's reply is that the tension between materialism and sentience is unnecessary. The apparent split between "physical" and "mental" comes from failing to appreciate the progressive biological and evolutionary increase in standard emergent features as brains advance toward sentience, and how monumental that increase is. Two things obscure this:
- Untraceable complexity. The exponential increase in emergent properties makes it virtually impossible to trace all the mechanisms in full.
- Personal nature. Sentience is personal and subjective, which chapter 10 addresses.
Panpsychists would say Feinberg misses their target. Their argument is not that neural emergence is too complex to trace but that no amount of structure and dynamics can yield intrinsic, qualitative character (the "intrinsic nature" or Russellian argument). Feinberg's appeal to degree of emergence does not engage that claim directly; his chapter 10 argument about personal, first-person access is the relevant response. Panpsychism in turn faces its own unsolved combination problem: how micro-experiences combine into one unified mind.
Biopsychism and Reber's "emergentist's dilemma"
Biopsychism, a term coined by Ernst Haeckel in 1892 (spelled "Haekel" in the book), now usually means that all living organisms are sentient. It is panpsychism's cousin, pushing consciousness down from all matter to all living matter.
Arthur Reber's case for his Cellular Basis of Consciousness rests on what he calls the Emergentist's Dilemma, perhaps the biggest fulcrum of his argument.
- Premise. If sentience emerged at some point in evolution, there was a species in which consciousness "burst into existence."
- Premise. One "cosmic moment" earlier, all less complex species were "mere mechanical entities" with not a glimmer of awareness.
- Premise. No coherent account explains such an abrupt appearance of a mind-like thing.
- Conclusion. So sentience did not emerge between bacteria and animals; it must have been present in bacteria from the start.
Feinberg dismantles it in two steps.
First, ES1 organisms are not sentient. Michael Woodruff (2016) summarises why CBC fails. The genes, sensory and motor substrates and biochemistry of bacterial chemotaxis are known in detail and are sufficient to explain it without sentience. So CBC fails on two assumptions: there is no empirical reason to set aside Occam's razor and add sentience as an explanatory variable, and no good evidence of evolutionary continuity in the genetic substrates of bio-sensitivity between bacteria and animals with nervous systems. A nervous system remains necessary for sentience.
Second, life still matters. Feinberg does not endorse "all life is sentient", but life processes play a basal role. His step-wise account starts with single cells as necessary first steps in a 3.5 billion year emergence.
NBE faces no emergentist's dilemma but offers an emergentist's solution: the same principles that explain sensing in single cells explain its stepwise elaboration into sentience, so sensing is an early emergent feature of life, and sentience is an emergent feature of both sensing and life.
Reber (biopsychism). No account can explain sentience switching on, so every living cell is sentient; the bacterium perceives and interprets valence.
Feinberg (NBE). Sentience never switched on in one burst; it emerged through graded stages over billions of years. Life is necessary for sentience but not sufficient.
Integrated Information Theory
IIT was proposed by Giulio Tononi and developed with colleagues. Its full apparatus of axioms, postulates and mathematics is beyond Feinberg's scope, but its central hypothesis is simple enough to compare. Tononi (2015): IIT identifies the essential properties of consciousness (axioms) and infers the properties physical systems need to account for them (postulates); from these it can in principle derive, for any system of elements in a state, whether it is conscious, how much, and which experience it has.
The core quantity is Φ ("big phi"), integrated information:
- Unity requirement. The cause-effect structure a system specifies must be intrinsically irreducible to that of independent subsystems obtained by unidirectional partition.
- Measure. Φ (a non-negative number) quantifies how much the cause-effect structure changes if the system is cut along its minimum partition (the one that makes the least difference).
- Existence. If a partition makes no difference, the whole is reducible to its parts; a whole with no cause-effect power beyond its parts does not exist in and of itself.
- Mechanisms. A subset of elements contributes a specific aspect of experience only if its combined repertoire is irreducible ("small phi", φ).
What NBE and IIT share
- Phi is a measure of emergence. A system with Φ must be unified, irreducible to its parts, with cause-effect power beyond its parts, which are all basic features of emergence. Feinberg argues Φ is, in effect, a mathematically oriented measure of the emergence of consciousness. Christof Koch (2012) says Φ measures "the extent to which the system is more than the sum of its parts," the holism of the network.
- Recurrent, reciprocal, circular interactions. Both stress recurrent and reciprocal interactions between parts and levels, though this is shared by many neuroscientific theories.
Where NBE and IIT differ
- The bar for consciousness. IIT requires an extremely low degree of complexity: objects without interacting parts are not conscious, nor are parts of a conscious system whose integrated information is below the maximum, but anything with interacting parts can have Φ above 0, including thermostats or an isolated proton with its three interacting quarks. Some read this as panpsychism. NBE sets the bar at ES3 brains.
- Continuous versus punctuated. Φ is continuous, rising with integration. NBE holds that sentience emerges in punctuated stages and does not clearly appear until about 520 million years ago. Since all living things have Φ far above 0, either all of them are sentient (which Feinberg has argued against) or IIT is wrong. For NBE, the degree of emergence, not its mere presence, decides.
- Does the substrate matter? For NBE, animal sentience is an emergent feature of life, and different qualia have different neurobiological substrates. IIT explains the character of experience by identifying every experience with a maximally irreducible "conceptual structure", a "form" in cause-effect space: the quality of experience is the shape of that structure, its quantity is Φmax. Ellia and Chis-Ciure (2022), comparing IIT with Feinberg and Mallatt's NN, note that for IIT "the substrate per se is irrelevant" so long as it specifies the right integrated information.
Feinberg calls the conceptual structure "to say the least a conceptually abstract postulate" that makes no reference to the biology of different experiences. It is simpler, he says, to hold that all qualitative experiences, within categories (red versus blue) and across them (exteroceptive versus affective), share emergent principles from life upward but differ in their neurobiological substrates: from receptors and affective centres to global scalar levels, the parts are materially different and vary with their experiential features. The parts and their interactions are causally related to the qualia they create; otherwise the personal "character of experience" is hard to explain.
| Question | NBE (Feinberg) | IIT (Tononi, Koch) |
|---|---|---|
| What is consciousness? | Emergent system property of living, complex brains | Maximally irreducible cause-effect structure |
| Role of emergence | Explicit, central | Implicit in Φ ("more than the sum") |
| Minimum system | ES3 brains (about 520 million years ago) | Anything with Φ above 0 |
| Graded or staged? | Punctuated stages | Continuous |
| Life required? | Yes | No |
| Quality of experience | Specific neural substrates | Shape of conceptual structure |
| Substrate | Matters | Matters only via causal structure |
"Substrate irrelevant" needs care. IIT is not functionalism. Tononi and Koch argue that consciousness depends on a system's physical cause-effect structure, so a digital computer simulating a brain may have negligible Φ and no experience even if it behaves identically. In that sense IIT says substrate architecture matters, though not biology as such. Feinberg's contrast is sharper on life than on substrate.
IIT's standing has changed since 2023. The Cogitate adversarial collaboration (results announced June 2023, published in Nature in 2025) tested IIT against Global Neuronal Workspace theory; several predictions of both theories failed, including IIT's prediction of sustained synchronization within posterior cortex. In September 2023, 124 researchers signed an open letter (Fleming et al., posted as a preprint) calling IIT "pseudoscience", citing its untested core and panpsychist implications; many others replied that the label was unjustified, and the dispute remains unresolved. IIT 4.0 (Albantakis, Tononi and colleagues, 2023) restated the theory, recasting the conceptual structure as a "Φ-structure" of distinctions and relations. Feinberg's complaint that IIT's bar is too low is close to the critics' panpsychism worry.
The quarrelling thermostats. Under IIT, a simple thermostat with interacting parts has some tiny Φ, so some faint experience. Now wire a thousand thermostats into a building-management grid with feedback loops: its Φ might exceed that of a nematode. Should the grid matter morally more than the worm? Feinberg's NBE gives a clear no: neither is alive in the right way or has maps, affect and nonreflexive behaviour. The thought experiment shows why a measure of integration alone sits uneasily with intuitions about who can feel.
Φ-style integration metrics are sometimes proposed for assessing AI systems. Feinberg's critique is a useful caution for governance teams: integration is a necessary feature of many complex systems, including payment networks and LLM architectures; on its own it is not evidence of sentience, and policies that rely on a single metric will over- or under-attribute.
Demystifying the Personal Nature of Sentience
This is the philosophical payoff. Feinberg argues that sentience is personal for the same reason life is personal: both are emergent system features of one embodied organism. Because sentience and its personal nature coemerged and coevolved, nothing is left unexplained. What remains is an experiential gap, the difference between knowing about an experience and having it, which Jackson's Mary illustrates and which poses no threat to physicalism.
The problem restated
Joseph Levine and like-minded philosophers locate the mystery in the gap between the brain's objective features and the subjective, personal character of experience. That gap is also what makes strong emergence look like a reasonable default. John Searle (1992) states the starting fact: because mental phenomena are tied to consciousness, and consciousness is essentially subjective, the ontology of the mental is a first-person ontology: mental states are always somebody's mental states.
The NBE explanation in three postulates
- Ordinary emergence, extraordinary degree. Sentience needs no radically unique emergence (physically fundamental, strong, "at the top", dualist, panpsychist or biopsychist). The degree of standard biological emergence enabled by complex, hierarchical nervous systems produces a novel system feature.
- Personal because biological. Consciousness is built on the emergence of life in a single organism; life and consciousness are both system features of embodied organisms. So each feeling is a personal system feature of that individual organism, just as its life is. Life provides the natural initial conditions for personal sentience: embodiment allows, and indeed necessitates, the personal nature of sentience.
- Coemergence and coevolution. Putting 1 and 2 together: life brings sensing (ES1); evolution adds neural features (ES2); at sufficient complexity, sentience appears as an aggregate system feature (ES3). Pari passu, its personal nature advances as a feature of the embodied organism. Sentience and its personal nature are both naturally emergent aggregate system features of complex brains.
Feelings belong to someone for the same reason metabolism belongs to someone. Nobody finds it mysterious that my digestion is mine and not yours; Feinberg's claim is that once sentience is seen as a life-process of a particular body, its first-person ownership is no more mysterious than that.
What the gap theorists were really struggling with
Feinberg diagnoses two related problems behind dualist and strong-emergence responses:
- Separation at the top. The assumption that personal qualities emerge from the "highest" neural levels, as if separate from the brain, perhaps nonmaterial, and then acting causally on lower levels. Chapter 8 dealt with this.
- Two perspectives. How to reconcile the brain as observed or explained (third person) with the brain as directly experienced (first person). This chapter deals with it.
The experiential gap
Feinberg finds no scientific explanatory gap between brain and personal experience, but he grants an experiential gap between objective explanations and subjective experience.
- The concession. No amount of knowledge or description of brain functions can equal, capture or substitute for "something it is like to be", any more than describing one's experience can substitute for having it.
- The claim. With a naturalistic theory of both the emergence of sentience and its personal nature, the gap between first- and third-person views can be scientifically reconciled without dualism. NBE is entirely compatible with physicalism, and it makes the experiential gap scientifically unproblematic.
| Explanatory gap (Levine) | Experiential gap (Feinberg) | |
|---|---|---|
| What it says | Science cannot explain why brain states feel as they do | Knowing about an experience is not having it |
| Kind of gap | Gap in explanation | Gap between two ways of knowing |
| Implication | Missing science, maybe new physics or dualism | Nothing missing; a consequence of personal emergence |
| Can it close? | Needs a bridge not yet built | Cannot and need not close: objectifying experience is futile |
And then along comes Mary
Frank Jackson set out the knowledge argument in "Epiphenomenal Qualia" (1982) and "What Mary Didn't Know" (1986).
Mary's room. Mary is a brilliant neuroscientist who knows every physical fact about colour vision, but she has lived her whole life in a black-and-white room and never seen colour. What she lacks is "knowledge by acquaintance." One day she walks out and sees red for the first time. Does she learn something new? If so, it seems physical knowledge was not all the knowledge there is. Feinberg notes that this is Broad's archangel again: both know all the facts yet cannot deduce what colour (or ammonia) is like.
- Premise. Before release, Mary knows all the physical facts about colour vision.
- Premise. On release, Mary learns something new: what seeing red is like.
- Intermediate conclusion. So not all facts are physical facts.
- Conclusion. Physicalism is false.
Jackson's early reading (as he restated it in 2004): for pangs, pains and sensings of red, no aggregation of elements with physical properties can make up the phenomenal side; that side is "an extra" left out by the physical sciences, "or so it seems." Then Jackson reversed himself: what seemed true is, on reflection, false. We must embrace some version of physicalism; we are "nothing over and above a very complex aggregation of purely physical elements", and since he refuses to eliminate phenomenal consciousness, it must exist somewhere in that complex aggregation.
Jackson's later view, that a complex aggregation of interacting physical elements can produce phenomenal consciousness, is exactly NBE's claim, and it accepts that knowing facts about an experience differs from having it without that difference blocking a natural, emergentist explanation.
Feinberg's reading of figure 10.1. "Knowing" comes in two types: by direct acquaintance (experience) and by objective description and explanation. The before-and-after difference is explained by the simple fact that only by being a living neural brain producing the emergent process of experience can knowledge by acquaintance be acquired. No dualism and no nonphysical explanation is needed.
This is one of several standard physicalist replies, and each is contested. Philosophers have answered Mary with the ability hypothesis (Lewis, Nemirow: Mary gains know-how, not new facts), the acquaintance hypothesis (Conee: she gains acquaintance with a known property), and the phenomenal concept strategy (she gains a new concept of an old fact). Feinberg's version resembles the acquaintance reply, grounded in biology. Critics such as Chalmers respond that these replies explain why the gap seems to exist without explaining why physical processes are accompanied by experience at all. Feinberg does not engage this literature in detail.
Conclusions
Feinberg concludes that the remaining philosophical questions do not bear on the scientific question of how the brain creates experience, as Jackson ultimately acknowledged.
- Explained. NBE accounts for the natural emergence of sentience and for its personal nature, which derives partly from living biological embodiment traceable back to the first sensing organisms.
- Compatible with physicalism. The experiential gap does not violate physicalism and does not support strong emergence.
- Rivals answered. NBE argues against sentience being immanent in nearly all matter (panpsychism), in all life including single cells (biopsychism), in all integrated matter (IIT), or arising from novel fundamental forces (naturalistic dualism, strong emergentism).
Gap as mystery (Broad, Levine, Chalmers, early Jackson). The first-person facts are left out of physical science; something extra is needed.
Gap as perspective (Searle, later Jackson, Feinberg). First-person facts are facts about one living system, accessible by being it; science explains why that access is private, and nothing is left out.
The perfect brain scan. Suppose a future scanner reads out, in complete detail, every neural process occurring while you taste coffee, and a scientist studies the printout. She now knows everything about your coffee experience except one thing: she is not having it. To have it she would have to be your living brain, or drink coffee herself and produce a different (her own) experience. Feinberg's point is that this limitation is not a failure of the scan; it follows from experience being a system property of one particular organism.
The experiential-gap framing helps in conversations about AI interpretability. A complete mechanistic account of a model's internals tells you what it computes; it would not, by itself, settle whether there is anything it is like to be that model. Feinberg's answer for animals ties that question to life and embodiment, which is one reason many theorists remain sceptical of machine sentience in disembodied systems.
Neurobiological Emergentism: Putting It All Together
Feinberg condenses NBE into its postulates: sentience is an emergent feature of embodied life; it evolved in punctuated stages; it needs only an increasing degree of standard emergence, not new physics; its mechanisms are diverse, so it is a system property, not a summit; and NBE reconciles the explanatory gap with a two-part solution (specific substrates plus personal embodiment) that leaves only an experiential gap.
The postulates of NBE
- Sentience is ultimately an emergent feature of embodied life. Any theory of animal sentience must begin with life; sensing in ES1 and ES2 organisms and sentience in ES3 animals are both emergent features of embodied life processes.
- Sentience evolved in progressive stages with increasing novel emergent features. Three stages (ES1, ES2, ES3), no single feature responsible, punctuated transitions, and shrinking time frames between stages.
- An increasing degree of standard neurobiological emergence makes sentience possible. No strong or radical emergence or new fundamental physics is required.
- The emergent mechanisms of sentience are diverse. Serial maps, nested hierarchies, HMNs, oscillations, and every direction of causation; so sentience is an aggregate system property, not something at the top.
- NBE naturally explains and reconciles the "explanatory gap." A two-part solution leaves only an experiential gap.
Four postulates or five? Feinberg introduces the chapter as summarising "four core postulates" and then lists five; the fifth (the gap) is arguably the philosophical consequence of the first four. He also dates the emergence of sentience here to "at a minimum about 520 to 540 million years ago," where earlier chapters give 560 to 520 million years ago for ES3. Both are loose Cambrian-era estimates, but the inconsistency is the book's.
Postulate 1: embodied life
The neurobiological mechanisms of sentience are emergent features of the organism's embodied life, so any theory of animal sentience must begin with life. ES1 and ES2 organisms are not sentient, but their sensing, like sentience, is an emergent feature of embodied life processes that serve as basal system features for both sensing and sentience.
Feinberg finds philosophical support in Evan Thompson's Mind in Life (2007). Thompson argues that the standard hard problem is embedded in the Cartesian framework of "mental" versus "physical", which should be replaced by an approach centred on life or living being. The gap does not vanish but changes character: the question is no longer the "contrived" one of deriving a subjectivist concept of consciousness from an objectivist concept of the body, but understanding the emergence of living subjectivity from living being, where living being already has an "interiority" that escapes the objectivist picture.
Life closes part of the gap, neural emergence closes the rest. Thompson admits the gap does not go away just because consciousness belongs to embodied life; "something more" is needed. Feinberg's "something more" is the set of unique emergent neurobiological features of sentient animals in postulates 2 to 4.
Postulate 2: progressive, punctuated stages
There are three biological and neurobiological emergent stages, no single feature that accounts for sentience, and punctuated transitions between stages, with time frames that shrink as evolution approaches sentience (about 3 billion years from ES1 to ES2, only 10 to 50 million from ES2 to ES3). At ES3 there is an "explosion" in the four table 4.1 variables: number of neurons, specialization, number of hierarchical levels and interaction between levels.
The decisive observation is a correspondence. The general features of emergent systems (table 2.1) are novel aggregate properties from the dynamic interaction of physically united, integrated or interacting parts, amplified by hierarchy. The features of nervous systems that create sentience (table 4.1) are exactly these features in greater degree.
| General emergence (table 2.1) | Sentient nervous systems (table 4.1) |
|---|---|
| Novel aggregate system property | Sentience as novel aggregate property |
| Parts must be united or interacting | More neurons, more connectivity |
| Emergent features are processes | Dynamic neural activity, oscillations |
| Hierarchy increases emergence | More levels, more level interaction |
| Specialization as levels are added | More specialized neuron types and regions |
If sentience is an emergent property of complex neural hierarchies, Feinberg says, this striking correspondence is exactly what we should expect.
Postulate 3: degree, not kind
Chalmers (1995) proposed taking experience as fundamental, "alongside mass, charge, and space-time," since everything in physical theory is compatible with the absence of consciousness. Feinberg calls this "not entirely implausible" but hard to prove unless such a fundamental feature is actually found. If the transitions ES1 → ES2 → ES3 are punctuated but natural and continuous, then standard biological emergence is scientifically sufficient, as it is for every complex physical system. An increasing degree, magnitude and variety of emergent processes, enabled partly by hierarchical nervous systems, accounts for sentience; sentience is novel, but its emergence needs no novel physics.
Chalmers (1995). Physics leaves consciousness out, so take experience as a fundamental feature and build laws connecting it to physical states.
Feinberg (NBE). Physics plus biology already contains the resources: a massive increase in ordinary emergence, traced across three stages, yields sentience.
Postulate 4: diverse mechanisms
Recapping chapter 8: emergence comes through serial nonnested isomorphic pathways, scalar nested hierarchies, hierarchical modular networks combining both, temporal integration via wave frequencies, and causation that runs same-level, bottom-up, top-down (constraint) and circular. Hence sentience does not emerge at the "top"; it is an aggregate emergent system property.
Diversity also runs within sentience (exteroceptive versus interoceptive-affective) and across species (chapter 7). So, just as subtypes of sentience feel different, sentience is plausibly experienced differently across species, even though all ES3 animals experience "something it is like to be."
Postulate 5: reconciling the gap
Feinberg lines up the gap's four classic forms: Broad's archangel (no physical knowledge substitutes for smelling ammonia), Levine (nothing about C-fibre firing "fits" pain; the feel is "brute fact"), Mary (no colour science tells her what red feels like), and Chalmers's character of consciousness: when middle C is struck, why is the auditory processing accompanied by that experience, with its characteristic rich tone and timbre?
- Specific substrates explain specific feels. No neuroscientist disputes that the feeling of red and the pain of a pinprick result from specific neural functions. Colour, sound, pain and affect pathways differ enormously, so these sensations should not, and could not, all feel the same. The "why does red feel red rather than like a trumpet" question ignores this diversity of substrates.
- Personal embodiment explains why it is felt by someone. The subjective nature and character of each state cannot be dissociated from its emergent neural substrate and from the personally living, embodied system whose state it is (postulate 1).
So sentience and its personal nature and character are both emergent features of evolved complex brains, and their coemergence did not appear suddenly or mysteriously: it took billions of years from life to sentience.
Broad, Levine, Chalmers and the Mary problem describe not a scientific explanatory gap but an experiential gap, whose two elements are both explained by the emergent nature of sentience, so attempting to fully "objectify" subjective experience is futile.
The weak point of part 1. Pointing out that red and pain have different substrates explains why they differ from each other, which most theorists accept. The hard-problem question is why any substrate is accompanied by experience, and why this substrate goes with this feel rather than an inverted one. Feinberg's answer to that rests on part 2, the personal-embodiment claim; readers who find chapter 10 unpersuasive will find part 1 alone insufficient.
Conclusions
Feinberg identifies three parallel progressive stages in the emergence of sentience:
- Neurobiology. An increase in complex neurobiological features.
- Time. An extended evolutionary timeline of billions of years.
- Experience. The eventual emergence of subjective experience.
He claims NBE is a fully naturalized theory of sentience that unifies biology, neurobiology, evolution and some of the philosophy of consciousness without denying the personal subjective nature of sentience and without trying to fully objectify it. It neither supports dualism nor argues against physicalism.
Thompson's question in a jar. Picture a cell culture of neurons on a chip, firing complex patterns, and next to it a living lamprey. Thompson asks: where is there already an interior from which a perspective could grow? For the lamprey, the answer is the self-maintaining living body, with metabolism and homeostasis that are for the organism. For the culture, it is unclear. NBE's postulate 1 says this difference matters: no life, no starting point for a personal perspective, whatever the neural complexity.
Organoids and "biocomputing" make the jar scenario live. Groups have trained cultured neurons on simple tasks (the "DishBrain" Pong experiment by Cortical Labs, published in 2022), and in 2023 researchers launched an "organoid intelligence" programme built on brain organoids. Ethicists now debate whether such systems could become sentient; Feinberg's criteria (maps, affect, nonreflexive behaviour, embodied life) offer one principled way to argue that current systems fall far short.
NBE's postulates map onto a familiar architecture argument: capabilities emerge from scale, specialization, layering and integration, and ownership of a capability belongs to the system boundary in which it runs. For AI governance, the second part matters: Feinberg ties the "someone" in sentience to a living, self-maintaining system, a criterion that current software agents do not meet.
Epilogue: Emergence, Process, and Sentience
Feinberg closes with Francis Crick, whose 1994 account of emergence anticipated NBE: wholes are more than the simple sum but understandable from parts plus interactions, and the obstacle is "the sheer variety and complexity" of the processes. Feinberg's final emphasis is that sentience is a process, not a structure, which explains why it is personal, and implies that our stable sense of self is an illusion over a constantly unfolding process.
Crick's two meanings of emergence
Francis Crick shared the 1962 Nobel Prize in Physiology or Medicine with James Watson and Maurice Wilkins for the structure of DNA. Later he turned to consciousness, and in the introduction to The Astonishing Hypothesis: The Scientific Search for the Soul (1994) he distinguished two meanings of "emergent":
The mystical meaning. Emergent behaviour cannot in any way, even in principle, be understood as the combined behaviour of the separate parts. Crick found this kind of thinking hard to relate to.
The scientific meaning (Crick's). The whole may not be the simple sum of the parts, but its behaviour can, at least in principle, be understood from the nature and behaviour of the parts plus knowledge of how they interact.
Crick added that there is no obvious reason we cannot obtain that knowledge, of the brain's components and their interactions; "the sheer variety and complexity of the process involved" is what makes progress slow.
Context: Crick's programme
Crick's "astonishing hypothesis" was that "you," your joys, sorrows, memories and sense of identity, are no more than the behaviour of a vast assembly of nerve cells and their molecules. With Christof Koch he launched the search for the neural correlates of consciousness (NCC), focusing on visual awareness in primates and on mechanisms such as synchronized firing in the gamma range. That programme shaped two decades of consciousness science and set the stage for theories such as global workspace and IIT. Feinberg's NBE shares its reductive-but-not-eliminative spirit while adding what Crick's programme largely left out: evolutionary depth across phyla and an account of why experience is personal.
| Crick and Koch (1990s to 2000s) | Feinberg (NBE, 2023) |
|---|---|
| Humans and primates, mostly vision | Vertebrates, cephalopods, arthropods, velvet worms |
| Find the minimal neural correlates | Trace stages of emergence over 3.5 billion years |
| Personal nature left aside | Personal nature explained by embodied life |
Three points of agreement with NBE
Feinberg finds remarkable agreement between Crick's view of almost 30 years earlier and the postulates of NBE.
- Rejection of "in principle" mystery. Crick rejects outright that emergence cannot even in principle explain consciousness, as NBE rejects strong emergence.
- Parts plus interactions. Crick holds consciousness can be explained by the nature and behaviour of the brain's parts and how they interact, NBE's weak emergence.
- The real obstacle is complexity. Crick foresaw that the hurdle would be "the sheer variety and complexity of the processes involved," matching NBE's diversity of emergent mechanisms (chapter 8) and its claim that the exponential increase in emergent processes makes them hard to trace (chapter 9).
The third point matters twice over. First, the diversity and complexity of mechanisms is enormous, as the book has documented. Second, Crick is somewhat unique in stressing not just neural structures but the variety and complexity of the processes they create.
Sentience is a process
Crick's emphasis on process echoes William James (consciousness is a function, not an entity) and Ernst Mayr (life and mind are reifications of activities). However much we think of consciousness as "only" the brain's neurons, sentience is a process, not a structure.
This is not a dismissal of substrate. The physical properties of life and the brain are key. The point is that integrating brain structure with experience requires seeing that emergent processes are processes, spanning all manner of levels and varieties of organization.
Sentience is created as much by what our brains do as by what they are: we should think of it as an embodied and personal emergent process created by the material brain.
| View | What sentience is | Where to look |
|---|---|---|
| Structure view | A property of a brain region or circuit | The "seat" of consciousness |
| Process view (James, Mayr, Crick, Feinberg) | An ongoing activity of a living, embodied brain | Dynamic interactions across levels and time |
Process and the personal nature of sentience
Crick's book solved many problems but did not address the personal nature of consciousness, which Feinberg calls "justifiably mysterious." Why can't the personal nature of red, pain or happiness simply be reduced to the material brain?
His answer is "surprisingly simple":
- Sentience is an emergent feature of the organism's life.
- A life is irreducibly tied to the physical organism that creates it; so are that organism's feelings, to whatever degree they exist.
- Life is an embodied, personal process of a living organism; sentience, as an emergent feature of that life and of its brain's processes, could not be any other way than uniquely personal.
The emergent nature of sentience explains its personal nature. NBE thus does two jobs at once: it accounts for the hidden complexities of how sentience is created and evolved, and it eliminates the "explanatory gap" between brain and sentience.
The self as an unfolding process
Feinberg ends on a reflective note. This view reshapes how we see our own sentience. Our sentience and our "selves," experienced subjectively as a stable entity, are really an illusion: sentience is in constant flux, always emerging, and never truly the same from one moment to the next.
Heraclitus's river, inside. You cannot step into the same river twice, said Heraclitus, because new water is always flowing. Apply that to experience: the "you" reading this sentence is a pattern sustained by neurons that are firing differently from a second ago, in a body whose molecules are steadily replaced. What stays constant is the organisation of the process, not any thing. Feinberg's claim is that the felt solidity of the self is how that process represents itself from inside.
A brief ending to a large claim. The "self as illusion" conclusion has a long pedigree (Hume's bundle theory, Buddhist anatta, and in recent neuroscience Thomas Metzinger and Anil Seth), but Feinberg asserts it in a few lines without engaging that literature or the counterargument that a process can be a real self without being a fixed thing. It reads better as "the self is a process" than as "the self is unreal."
Process views have gained ground. Predictive-processing accounts, popularised in Seth's Being You (2021), treat experience and selfhood as ongoing, embodied regulation of the living body, closely matching Feinberg's emphasis on life, embodiment and process. The main difference is that predictive theories specify a computational mechanism (prediction-error minimisation), whereas NBE stays at the level of emergence principles and evolutionary stages.
"What it does, not just what it is" is a sound rule for evaluating complex systems. In AI and enterprise platforms, behaviour emerges from runtime processes (data flows, feedback loops, agent interactions) rather than from component diagrams alone; monitoring the process tells you more than inspecting the parts.
Master Takeaways
Feinberg's book adds up to a biology-first, physicalist account of feeling: sentience is what happens when ordinary emergence is pushed to an extraordinary degree in living, hierarchical nervous systems, and it is personal because it belongs to one embodied life. Its strength is the comparative, evolutionary evidence; its exposure is the philosophical claim that renaming the gap "experiential" fully answers the hard problem.
The takeaways
- Sentience is weak emergence, not new physics. Feinberg argues that standard biological emergence, in sufficient degree, explains feeling; no fundamental "experience field" (Chalmers) or top-down mind (Sperry) is needed.
- The explanatory gap is really an experiential gap. Knowing everything about a brain is not being that brain; this difference is real but explained by emergence and compatible with physicalism, so objectifying experience is futile.
- Sentience is personal because life is personal. Feeling is an emergent feature of one organism's embodied life, so its first-person ownership is no more mysterious than the ownership of metabolism.
- Sensing is universal; sentience is not. Every cell senses and responds with valence-appropriate behaviour, but feeling needs maps, affective circuits and nonreflexive behaviour, which only ES3 brains provide.
- Sentience evolved independently at least three or four times. Vertebrates, coleoid cephalopods, arthropods and (Feinberg proposes) velvet worms converged on image-forming senses, topographic maps, memory and premotor centres and affective circuits: strong evidence of what sentience requires.
- There is no "top" where consciousness lives. Serial maps, nested hierarchies, modular networks and oscillations, linked by upward, downward and circular causality, make sentience a system feature, not a level feature.
- Four variables order brains toward sentience. Neuron number, neural specialization, number of hierarchical levels and interaction between levels; they rise together in evolution, though Feinberg gives no thresholds.
- Three emergent stages. ES1 single cells (3.5 billion years ago, sensing), ES2 early nervous systems (about 570 million years ago, presentient), ES3 complex brains (560 to 520 million years ago, sentient).
- The pace accelerated. About 3 billion years separated ES1 from ES2, only 10 to 50 million years separated ES2 from ES3, consistent with the Cambrian explosion.
- Exteroceptive sentience means mental images. Its neural signature is successive, higher-order, isomorphic maps (retinotopic, somatotopic, tonotopic); a single low-level map, like the spinal cord's, is not enough.
- Interoceptive-affective sentience means valence. Following Berridge, valence has a hedonic side (pleasure, pain) and a motivational side (pursuit, avoidance); approach and withdrawal alone are not evidence of feeling.
- Nonreflexivity is the behavioural signature. Reflexes are fast, hard-wired, stereotyped and shallow; sentient behaviour is flexible, enduring, trade-off-sensitive and hierarchically based (Elwood's definition).
- The Birch-Crump eight criteria are the working test. Nociception, sensory integration, integrated nociception, analgesia, motivational trade-offs, flexible self-protection, associative learning and analgesia preference; they became UK law in 2022.
- E. coli's chemotaxis is a biased random walk. 5 to 10 flagellar motors switching between counterclockwise runs and clockwise tumbles reach food without ever choosing a direction.
- Stentor's famous "trying" is probabilistic. Dexter et al. (2019) replicated Jennings's four avoidance behaviours but found heterogeneous sequences and a coin-toss choice between contraction and detachment.
- ES2 animals are a royal road, not a destination. C. elegans (302 neurons), amphioxus and sea slugs show layered integration and valence circuits but no image-forming maps or affective infrastructure; nematodes lack an egocentric spatial model (Barron and Klein).
- Insects meet the criteria despite tiny brains. Cross-modal object recognition, dopamine-dependent optimism, heat-versus-sugar trade-offs and observational learning in bees show that architecture beats neuron count.
- Octopuses seek relief from pain. Crook's experiments show lidocaine abolishes arm-to-brain nociceptive signalling and octopuses prefer a lidocaine chamber only after injury.
- Crabs and hermit crabs behave flexibly and socially. Neohelice adjusts escape runs to the visual flow, and hermit crabs form size-ordered vacancy chains to trade up shells.
- Sperry's "emergence at the top" leads to dualism. A mind that commands lower levels reproduces Descartes's interaction problem.
- Chalmers's strong emergence and naturalistic dualism "have the cake and eat it." They make experience brain-dependent yet ontologically separate.
- Downward causation is real but weak. In biology it is constraint (Pattee), and with upward emergence it forms circular causality (Nunez).
- Reber's emergentist's dilemma dissolves. Sentience never "burst into existence"; it emerged by graded stages, so NBE offers an emergentist's solution instead of biopsychism.
- IIT's Φ is a measure of emergence. It quantifies how much a whole is more than the sum of its parts, but its bar (anything with Φ above 0) is far too low, it is continuous rather than staged, and it ignores life and specific substrates.
- Different qualia, different substrates. Red, middle C and pain feel different because they run on different neural machinery; asking why red is not like a trumpet ignores the biology.
- Jackson's Mary supports NBE, on Jackson's later reading. Mary gains knowledge by acquaintance that only being a living brain provides; Jackson himself came to embrace physicalism.
- Sentience is a process, not a structure. With James, Mayr and Crick, Feinberg holds that feeling is what brains do as much as what they are, and the self is a constantly unfolding process.
Frameworks & mental models
| Name | What it says | Use it when |
|---|---|---|
| Neurobiological Emergentism (NBE) | Sentience is a weakly emergent, personal system property of living, complex neural hierarchies | Explaining feeling without dualism or new physics |
| Three aims of NBE | General mechanism; parallel with evolution; explain personal nature and the gap | Summarising what the theory is for |
| Four statements of the gap | Broad's archangel, Levine's gap, Chalmers's hard problem, Carruthers and Schier's "something left over" | Laying out the problem NBE answers |
| Biological Naturalism vs NN vs NBE | Searle: natural and first-person; Feinberg-Mallatt NN: special neural mechanisms, life, three stages; NBE adds emergence | Tracing the theory's lineage |
| General features of biological emergence (table 2.1) | Aggregate novelty; unity; process; hierarchy | Testing whether a property is emergent |
| Emergence in hierarchies (table 2.2) | Hierarchy matters; reciprocal connections; multiple scales and frequencies; more levels, more novelty | Assessing how architecture boosts emergence |
| Two subtypes of sentience | Exteroceptive (mapped mental images) vs interoceptive-affective (valence) | Asking which kind of feeling an animal might have |
| Berridge's valence components | Hedonic (pleasure, displeasure) and motivational (pursuit, avoidance) | Separating feeling from mere approach-avoid |
| Birch-Crump eight criteria | Nociception, integration, integrated nociception, analgesia, trade-offs, flexible self-protection, associative learning, analgesia preference | Grading evidence for pain sentience in any taxon |
| Reflexive vs nonreflexive | Reflexes fast, hard-wired, stereotyped, shallow; nonreflexive flexible, enduring, hierarchical | Judging whether behaviour indicates feeling |
| Learning hierarchy | Habituation, sensitization, classical, operant, unlimited associative learning | Weighing learning evidence (Ginsburg and Jablonka) |
| Four complexity variables (table 4.1) | Neurons, specialization, levels, level interaction | Ordering nervous systems toward sentience |
| Emergent stages ES1-ES3 | Cells sense; early nervous systems are presentient; complex brains are sentient | Placing an organism on the road to feeling |
| Stage feature tables (5.1, 6.1, 7.1) | Novel structures and processes added at each stage | Checking what a stage contributes |
| Four lines to sentience | Vertebrates, coleoids, arthropods, onychophorans | Comparing convergent sentient brains |
| Vertebrate behavioural evidence (table 7.2) | Trade-offs, frustration, self-delivery of relievers, place preference, operant learning | Assessing vertebrate affect |
| Emergence at the top (Sperry) | Mind emerges at the hierarchy's apex and commands lower levels | Spotting hidden dualism |
| Strong vs weak emergence and downward causation | Strong: not deducible even in principle; weak: deducible but unexpected | Deciding whether new physics is claimed |
| Layer cake (Kim) vs system view | Stacked levels vs many interacting hierarchies | Rejecting "which level is conscious" questions |
| Neural hierarchy types | Serial nonnested maps; scalar nested hierarchies; HMNs; temporal scales | Describing brain architecture precisely |
| Upward, downward, circular causality | Emergence up, constraint down, loops between scales (Pattee, Nunez) | Explaining how wholes and parts co-determine |
| Reber's emergentist's dilemma | No account of sentience bursting into existence, so cells must be sentient | Evaluating biopsychism |
| NBE vs IIT | Shared: Φ measures emergence, recurrent interactions; differ: bar, continuity, substrate and life | Comparing NBE with integration theories |
| Experiential gap | Two ways of knowing, by description and by acquaintance, both explained by emergence | Answering knowledge and hard-problem arguments |
| Knowledge argument (Mary) | Complete physical knowledge does not give knowledge by acquaintance | Testing physicalism |
| Five postulates of NBE | Embodied life; staged evolution; degree of standard emergence; diverse mechanisms; gap reconciled | Summarising the theory |
| Two-part solution | Specific substrates explain character; embodiment explains ownership | Answering "why does red feel red" |
| Crick's two meanings of emergence | Mystical (not understandable in principle) vs scientific (parts plus interactions) | Separating respectable from mysterious emergence |
Strengths. The comparative evidence is excellent: the book gathers neuroanatomy and behaviour across phyla in one framework, uses the Birch-Crump criteria that later shaped policy, and makes a powerful case from convergent evolution of sentient brains. Its treatment of levels and hierarchy (chapter 8) is a clear corrective to "seat of consciousness" thinking.
Blind spots. The philosophical resolution is thin relative to its ambition. Showing that experience is private to one living system explains why the gap appears, but critics argue it does not explain why there is any experience at all; Feinberg engages little with the phenomenal concept literature, illusionism, or predictive processing. The four complexity variables have no thresholds (and the "more than 100,000 neurons" marker conflicts with insect counts). The velvet-worm line rests on homology and social behaviour with no direct tests. Life as a requirement is asserted rather than argued in mechanistic detail, which matters for questions about organoids and AI.
Contested or dated. The Cambrian origin of sentience is plausible but debated; ES2 verdicts for C. elegans and gastropods may soften as basal cognition and invertebrate learning work grows. Minor errors: "Neohelice granulate" (granulata), "Haekel" (Haeckel), "DiFresco" (DiFrisco), "four" postulates that are five, and ES3 dated both 560 to 520 and 520 to 540 million years ago.
How the field has moved. The 2023 Cogitate adversarial collaboration (published 2025) found key predictions of both IIT and GNW wanting; the September 2023 "pseudoscience" letter against IIT and the replies to it have polarised the field; IIT 4.0 restated the theory. The 2024 New York Declaration extended a "realistic possibility" of consciousness to all vertebrates and many invertebrates, including insects, broadly matching Feinberg's ES3 roster; the UK Animal Welfare (Sentience) Act 2022 covers cephalopods and decapods. Organoid intelligence and debates on AI consciousness have made NBE's insistence on embodied life newly relevant.
Who's who
| Name | Who | Why they matter here |
|---|---|---|
| Todd E. Feinberg | Neurologist and psychiatrist, author | Proposes NBE |
| Jon Mallatt | Evolutionary biologist | Co-developer of Neurobiological Naturalism and the stage model |
| Thomas Nagel | Philosopher | "Something it is like to be" defines sentience |
| C. D. Broad | Philosopher (1887 to 1971) | The mathematical archangel and the smell of ammonia (1925) |
| Joseph Levine | Philosopher | Coined the "explanatory gap" (1983) |
| David Chalmers | Philosopher | Hard problem, strong emergence, naturalistic dualism |
| Peter Carruthers and Elizabeth Schier | Philosophers | The "something left over" argument |
| John Searle | Philosopher | Biological Naturalism; first-person ontology |
| Ernst Mayr | Evolutionary biologist | Emergence; life and mind as reified processes |
| G. H. Lewes | Philosopher (1875) | First scientific use of "emergence" |
| Elly Vintiadis | Philosopher | Concise definition of emergence |
| Gerald Edelman and Antonio Damasio | Neuroscientists | Mental images in primary and core consciousness |
| Jonathan Birch and Andrew Crump | Philosophers of animal sentience | The eight criteria for pain sentience |
| Kent Berridge | Neuroscientist | Hedonic and motivational components of valence |
| Robert Elwood | Animal behaviour scientist, Queen's Belfast | Reflex definition; hermit crab sentience |
| Eva Jablonka and Simona Ginsburg | Biologists | Unlimited associative learning as a marker |
| Lynne Sneddon | Biologist | Behavioural complexity as evidence of pain |
| Herbert Spencer Jennings | Zoologist (1902 to 1906) | Stentor's avoidance hierarchy |
| Arthur Reber | Psychologist | Cellular Basis of Consciousness; the emergentist's dilemma |
| Michael Woodruff | Neuroscientist | Critique of CBC via Occam's razor |
| Andrew Barron and Colin Klein | Neuroscientist and philosopher | Nematodes lack an egocentric world model |
| Thurston Lacalli | Evolutionary biologist | Amphioxus nervous system and escape circuit |
| Rhanor Gillette | Neuroscientist | Sea slug incentive-integrator valence circuits |
| Eric Kandel | Neuroscientist, Nobel 2000 | Learning and memory in Aplysia |
| Robyn Crook | Neuroscientist | Octopus pain and lidocaine place preference |
| Lars Chittka and colleagues (Solvi, Perry, Loukola, Gibbons) | Bee cognition researchers | Insect cross-modal recognition, optimism, trade-offs, observational learning |
| Nicholas Strausfeld | Neuroanatomist | Crustacean hemiellipsoid body = mushroom body |
| Daniel Tomsic | Neuroscientist, Buenos Aires | Neohelice crab escape behaviour |
| Roger Sperry | Neuroscientist, Nobel 1981 | Emergence at the top; downward causation |
| Antti Revonsuo | Philosopher-neuroscientist | Summary of strong emergentism; mental causation |
| Jaegwon Kim | Philosopher | Layered emergence; upward, downward, same-level causation |
| Claus Hilgetag, Alexandros Goulas, Olaf Sporns | Network neuroscientists | Kinds of neural hierarchy; hierarchical modular networks |
| H. H. Pattee and Paul Nunez | Biologist and physicist | Constraint; circular causality |
| Philip Goff and John Heil | Philosophers | Panpsychism and its anti-emergence argument |
| Giulio Tononi and Christof Koch | Neuroscientists | Integrated Information Theory and Φ |
| Frank Jackson | Philosopher | Mary and the knowledge argument, later physicalism |
| Evan Thompson | Philosopher | Mind in Life: the gap reframed around living being |
| Francis Crick | Molecular biologist, Nobel 1962 | Scientific versus mystical emergence; process |
Connect it
- Sensing is not feeling, in models too. Feinberg's ES1 lesson (rich, adaptive, reflexive behaviour without experience) is a useful guard when customers, staff or regulators attribute intent or feeling to AI assistants such as relationship-manager copilots.
- Graded evidence, not single tests. The Birch-Crump approach of separate architectural and behavioural indicators, weighed together, is a good template for AI capability and welfare assessments and for model-risk frameworks that must avoid over-reliance on one metric (including Φ-like integration scores).
- System properties live in interactions. NBE's "no top" argument mirrors enterprise architecture practice: resilience, security or "understanding" belong to the system of interacting layers and feedback loops, so governance should test runtime behaviour across components, not certify individual parts.
- Embodiment as a boundary condition. Feinberg ties the "someone" in sentience to a self-maintaining living body; whatever one concludes, it gives a principled reason why current disembodied software agents are poor candidates for sentience, a point increasingly asked about in AI policy and ethics reviews.