Things that question being/consciousness

  • What does anaesthesia do to your consciousness? Where does the state of being temporarily get abolished to?
  • Where does subjective universe come into play?

Real problem of consciousness

Explain, predict, and control the phenomenological properties of conscious experience. It’s different from hard problem because it’s not trying to place consciousness in relation to the universe. It doesn’t need a magic like property for consciousness.

Historical Ideas

Study of consciousness and trying to know it almost feels impossible, but we felt the same about phenomenon of being alive quiet some time ago. Theories like vitalism tried explaining it similarly to how dualism tried explaining consciousness. With modern medical science, we came to understand what being alive meant. So, there is a good argument that we will figure the consciousness part as well. The only major hurdle is being alive is objective but consciousness is subjective.

What is it like to be a bat?

Proposed that there would be something to be a bat, experienced by a bat, that even humans can’t get a grasp on.

Consciousness as certain outward action like language

Previous to phenomenology, being conscious had been confused with having language, being intelligent, or exhibiting certain behavior. This contradicts with people dreaming/ people with bodily paralysis.

Phenomenology

Idea that consciousness is about subjective experience.

Higher order Thought Theory

  • When there is a higher-level cognitive process that is somehow oriented towards our mental content, it renders them conscious.
  • Emphasize on frontal brain regions too for consciousness.

Hard and Easy Problems

  • While the hard problems focus on big questions that are hard to answer, easy problems cover what we can do that can be specified in terms of function/behavior.

Dualism

Rene Descartes

Described the composition of a being from dwo duals, therby introducing the idea of dualism.

  • Res Cogitans ( Mind Stuff )
  • Res Extensa ( Matter Stuff )

Physicalism

Universe is made of the physical stuff, and that conscious states are either identical to, or somehow emerge from the arrangement of physical stuff.

Functionalism

Consciousness doesn’t depend on what a system is made of, only on what the system does, on the functions it performs, and on how it transforms inputs into outputs. So, mind and consciousness are forms of information processing implemented by brains. So, by making input-output relation a certain way, consciousness shall arise.

Idealism

The consciousness is the ultimate source of reality. So, matter emerges from the mind.

Panpsychism

Many believe that this exists to give us an easy way out to the hard problem by taking it to an empirical dead end.

Mysterianism

Physical understanding of consciousness exists but it’s elusive just like our knowledge is elusive to animals.

Neural Correlates of Consciousness

NCC is defined as the minimal neuronal mechanisms jointly sufficient for any one specific conscious percept. It proposes that there is some specific pattern of neural activity that is responsible for every experience e.g. seeing red. So, if the activity is present, experience of redness will happen. Disentangling a true “NCC” is very difficult, perhaps impossible.

Binocular Rivalry

The first limitation to NCC was explained by this experiment. Showing two different images to each eye. In this case, the conscious perception switches back and forth between two objects shown on two different eyes. So, this means that conscious perception changes even when the inputs remain constant.

Informative and Integrated Theory [ IIT ]

Information is any reduction of uncertainty. Every conscious experience delivers a massive reduction of uncertainty. But this isn’t enough. It needs to be integrated as well. We don’t experience colours separately from the shapes, nor objects independently from their background. One of the key move was a proposal that if every conscious experience is IIT at the level of phenomenology, then the neural mechanisms underlying conscious experiences should also exhibit both these properties. In virtue of expressing both these properties in neural mechanisms, they account for core phenomenological features of every conscious experience.

A random and independent system, like molecules in a gas has maximum information, but shows no integration at all since all elements are independent from one another. LZW complexity would score very high on such systems. Each neuron is as if it’s disconnected from every other neuron. At another extreme is system where all elements do exactly the same thing, i.e. we can determine state of every other element using a single element. e.g. A crystal lattice. This has maximum integration, but no information. This kind of behavior is close to global epileptic seizures. Algorithmic Complexity here would be very low. In the middle lies systems that have varied degree of these two elements in which there is some degree of co-ordination where system tends to behave like a whole. So, this kind of system is what’s referred to as being “complex”.

IIT says information is as real as mass or energy. At the core of IIT is single measure \(\phi\) which is measure of how much a system is more than sum of its parts in terms of information. e.g. Flock of birds. A system is conscious to the extent that its whole generates more information than its parts.

To have high value of φ: Firstly, system must rule out a large number of alternative possible global states. Secondly, the information when considering the system as a whole must be greater than considering all individual neurons. e.g. If you remove a part of brain, it loses some amount of consciousness. Or if you divide it into two cortical hemispheres, in some cases it can harbor two independent consciousness. IIT also explains why cerebellum doesn’t seem involved in consciousness. It has vast number of semi-independent neurons unlike cerebral cortex which has dense interconnected wiring. Then, question becomes if wiring doesn’t change, how come we aren’t conscious in sleep or anaesthetic state? The answer is that the ability of cortical neurons to interact with each other is highly compromised in such states. According to IIT, many systems can thus impart consciousness.

The major problem is measure of information is relative to observer as measured by Claude Shannon’s theory of entropy. e.g. If you already are 100% certain that it’s going to rain, it isn’t information to you. If the information measured is observer relative, that would mean the consciousness too is observer relative.

Solution to this: Firstly, perturb system in all possible ways and see what happens. Secondly, infer the maximum entropy distribution from an exhaustive, complete knowledge of system’s physical mechanisms. By doing this, we can infer possible states even if we haven’t observed them.

Again, the problem is we have access to what system does, but rarely into what system is capable of doing. This is true for our brains. Also, the problem is number of states increase exponentially with each possible degree of freedom we impart for the system.

The complexity of IIT tends to challenge a breakthrough in physics too demanding that information exists just like mass or charge exists. It leads to some degree of panpsychism.

Now, scientists needed a measure to measure both of these properties at the same time.

Using observer relative information for φ

  • Don’t care what the system could do but never does.

Features of consciousness

Levels [ How conscious we are ]

From complete absence like coma or brain death to complete vivid presence of awareness in day to day life.

Contents [ What are we conscious of ]

Sight, Sound, Smell, Emotion, Mood, Thought, Belief etc. Almost all come from some form of sensory input’s interpretation.

Self [ What is the specific experience of being you ]

Can be called a subset of content of consciousness specifically consisting of experiences that give us a FPP of being us: our unique memories, moods, will etc.

Measuring consciousness

Bispectral Index Monitor

Can it be measured by measuring level of awareness and wakefulness?
In vegetative state where person cycles between sleep and wakefulness doesn’t show any level of conscious awareness.
Can it be measured by counting the number of neurons involved?
Cerebellum has 4x neurons as rest of brain put together, but seems barely involved in consciousness.
Can it be measured by overall degree of neuronal activity?
Because brain doesn’t shut down even in unconscious states, it’s not true.

Can it be measured by specific communication pathways of the brain?

Stimulating one part of the brain and tracking what’s spread over space and time in the brain. TMS ( Transcranial Magnetic Simulation ) injects a sharp pulse into brain through skull. EEG measures that respone to zapping. One problem identified was that our conscious experience didn’t change when zapped. In unconscious states like dreamless sleep and general anaesthesia, echoes of zapping are very simple. They die out quickly. However, during conscious states, response echoes widely over cortical surfaces, disappearing and appearing in complex patterns. Mostly the thalamocortical system.

Zap and Zip Algorithm
Tries to give a level to the complexity of pattern. Lempel-Ziv-Welch Complexity.
PCI
Perturbational Complexity Index
  • This even worked so good that it drew a dividing line between states indicative of consciousness and states indicating the absence. This was done independently from neurologist’s analyses.
  • Locked in patients had indistinguishable PCI values from normal ones.
  • In healthy people, brain regions engaged by imagining fluent movements like playing tennis are highly distinct from those activated by imagining navigating through spaces.

Measuring complexity of spontaneous cortical activity - Seth is working on measuring algorithmic complexity of spontaneous brain activity instead of TMS induced activity. - Discovered that it drops in early sleep and anaesthesia. - Complexity during REM sleep is much the same as conscious wakefulness, overlapping with the fact that we are conscious of our dreams.

Story of how PCI helped

Marcello Massimini recently told me about one case where measuring PCI made all the difference. A young man with severe head injuries had been admitted to hospital in Milan. He remained unresponsive to simple questions and commands, suggesting a diagnosis of vegetative state. But his PCI was as high as a healthy, fully conscious person – a perplexing observation, especially since he was not locked-in. The clinical team eventually tracked down a relative, an uncle who travelled to Italy from North Africa, where the young man’s family still lived. When this uncle started engaging with his nephew in Arabic there was an instant response: smiling at jokes, even giving a thumbs-up when watching a movie. He had been conscious all along – just unresponsive in Italian. Why this should be is hard to say. Massimini believes it might be a strange case of ‘cultural neglect’, as if the Italian world simply did not matter to him any more. Either way, this young man’s story could have ended very differently without the telltale electrical echoes measured by the PCI.

Psychedelics

Influence the brain’s serotonin system Dissolves the ego boundary – a distinction between self and others They increased the measure of conscious level relative to waking rest level. Even in placebo controlled experiments, the results were pretty clear. The question introduced with this result is “Can there be next level of consciousness that can be unlocked by psychedelics? " Complexity is different from randomness. The algorithmic complexity is more a measure of randomness. Complexity that we want is not extreme point of disorder but extreme point of order before disorder.

Perception

  • As a brain, it has to go into electrical signals indirectly related to contents of consciousness. These signals don’t come with labels. They don’t even distinguish themselves as visual or auditory or any other modality signal. Following the idea that brain is a prediction machine, the feel is best guess of what our brain thinks of the cause of sensory inputs.
  • Signals flowing “top-down” or “inside-out” serve only to refine all important “bottom-up” flow of sensory information.

Perception defined by various people

  • Helmholtz proposed the idea of perception as a process of ‘unconscious inference’.
  • Arab scholar Ibn al Haytham proposed perception depends on processes of ‘judgement and inference'.
  • Immanuel Kant used term noumenon referring to ’things in themserlves’ and Ding an sich - a mind independent reality that will always be inaccessible to human perception, hidden behind a sensory veil.
    • Various experiments were done to prove this. One interesting experiment showed that poor families overestimated the size of coins. But studies like this done without much standards are hard to trust.
  • In Gregory’s terms, perception is determined by the brain’s best-supported hypotheses.
  • Predictive Coding or Predictive Processing – all theories that share this idea believe that perception depends on brain-based inference of some kind.
  • Prediction necessarily isn’t limited to future references. It’s referring to model giving output that’s beyond the data that it has.

Bottom Up View of Perception

  • Different perceptual modalities associated with different regions: visual cortex, auditory cortex, and so on.
  • Organised hierarchically within each region for processing
    • e.g. Lower level visual cortex close to sensory inputs while higher levels such as inferotemporal cortex are several stages of processing further away.
    • Signals are pooled together in level above, so they can respond to features.
  • He says just like we perceive sun going around earth but in fact earth is going around the sun, our perception are not being interpreted by the mind, but the mind is made from these perception. Perception, as he believes, is not process of “reading out” sensory data.

Controlled Hallucination View

  • Brain constantly makes predictions for causes of sensory signals. e.g. On seeing any object, visual cortex tries formulating predictions for what the object is.
  • Sensory signals, however, travel bottom up to serve as prediction errors between what is predicted by brain and what brain gets at every level of processing. So, brain is trying to adjust top-down predictions based on bottom-up sensory signal to suppress bottom-up prediction errors.
  • So, Continuous process of minimizing prediction errors is how perception happens in our brains.
  • Perceptual experience of seeing any object is determined by content of top-down predictions and not by the bottom-up sensory signals. This is why the name controlled hallucination.
    • We never experience the object itself but only the interpretations of it.
  • In such scenario, the world exists because our mind predicts it. Instead of perception being a process of bottom-up feature detection, it is top-down neuronal fantasy that takes feedback from reality but might not be transparent to what reality actually is.
  • The word controlled since it is somewhat still tied to causes in the world unlike normal hallucinations which have no real basis in the world i.e. entirely created by the mind. Hallucination, thus, can be defined as uncontrolled perception.
  • We experience this in our daily lives like our perception for colors. There is no one-one mapping for our color sensitive retina cells and color experience. It depends on how our brain makes best guesses for what they should be.
  • Redness thus depends on both the color red present in the world and the properties of brain that experiences it.
  • In terms of color, we see brain trying to make best guess and get confused every now and then. There was a famous dress that people couldn’t figure out single color for. Based on how ambient light is, it was perceived by people to be two different colors. This also explains that perception is more a generative act.
  • Controlled Hallucination theory takes predictive processing as a theory of how brains work and develops to account for nature of conscious experiences.

Objections to controlled Hallucination

  • Now, the question becomes – What are the absolute primary qualities of an object that exist independently of an observer?
    • Locke proposed these to be occupying space, having solidity, and moving. e.g. A moving train however you perceive it looks mostly the same to everyone.
    • Secondary qualities are those whose existence depend on an observer. Those which produce sensations or ideas in our brain. e.g. Color is a good example of secondary quality.
  • Another objection is that though you haven’t seen a gorilla anywhere till date, how can you already expect it. You see it similarly scary as one who has seen it already in a random encounter.
    • But, the brain can do lot of different break downs and build ups to tell you what you saw based on what you have already seen.
  • When knowledge fails to affect perception, we call that perception cognitively impenetrable.

Abductive Reasoning

  • Type of predictive processing too.
  • Seeking the best explanation. Reasoning backward, from observed effects to their most likely causes, rather than from causes to effects as it happens in deduction and induction.
  • Bayes Rule is mathematics for going from what we already know to what we should believe in next based on what we are learning now.
  • Perceptual Priors can be encoded at several levels of abstraction and flexibility. Likelihoods in the brain encode mappings from potential causes to sensory signals.
  • Both this and controlled hallucination try to minimise prediction errors.

Precision of signals

  • Precision of sensory signals is also not something that goes into the brain as an input. So, brain estimates it.
  • figure out likely causes of sensory inputs but also while figuring out how reliable the relevant sensory inputs are.

Inattentional Blindness Phenomenon

  • In some cases, unattended sensory data may have no influence at all.
  • The gorilla experiment tells us that people by focusing on something as a part of a scene can completely miss out something that they saw but didn’t attend to because they were focused on something else.
  • In this case, the brain records the signal as low estimated precision, and so has no influence on updating our perceptual best guesses.

Action

  • From Inattentional Blindness Phenomenon, we can ask to redefine which comes first. We generally believe Sense. Think. Act. is the right order, which brings another important point to the picture. i.e. Action.
  • This also explains that action is critical to perception. Every action alters perception by changing incoming sensory data, and every perception is the way it is in order to help guide action.
  • It can go even further that action comes first. We can think of brain as actors trying to generate actions, and continuously calibrating these actions using sensory signals to meet an organism’s goals.
  • Prediction errors can also be provided by performing actions in order to change sensory data, so that new sensory data matches an existing prediction. This process of minimising prediction error through action is called Active Inference. This was proposed by Karl Friston.

Active Inference

  • Similar to other predictive processing, this too depends on generative modeling.
  • It relies on generative model to predict sensory consequences of actions. e.g. “If I look over there, what sensory data I am likely to encounter?”
    • Conditional Predictions
  • e.g. When I am looking for keys, I tidy all items on my desk to find it. Here, choosing the relevant action relies on having a generative model able to predict how sensory data would change as result of my action of tidying up.
  • Action itself can be thought of as a formm of self-fulfilling perceptual prediction.
    • In active inference, we can see action as self-fulfilling proprioceptive predictions.
  • Proprioception keeps track of where body is and how it is moving, by registering sensory signals that flow through muscles and bones.

Relationship between action and perception

  • Rather than seeing action and perception as two sides of coin, we can see them both as forms of brain-based prediction.
  • So, our body is getting a lot of self-fulfilling proprioceptions and sensory signals from the world, from which our brain creates new sensory flow.
  • Incoming sensory signals is met through a cascade of top-down predictions, with prediction error signals streaming upward to simulate ever better predictions and elicit new actions.
    • This rolling process gives rise to an approximation to Bayesian inference, a good enough system, in which the brain setttles and resettles its evolving best guesses about the causes of sensory environment. In that a vivid perceptual world – a controlled hallucination - is brought to being.

Beholder’s share

  • Beholder’s share is part of perceptual experience that is contributed by the preceiver and which isn’t found in the artwaork or the world itself.
  • One cannot perceieve what one cannot classify.
  • Explained by Continual Flash Suppression Experiment
    • Images presented to left and right eyes. One eye is presented with a picture and another with rapidly changing patterns of overlapping oblongs. On this process, brain tries to fuse two images into a single scene, it fails, and changing oblongs since they tend to dominate, that’s what the person sees. Conscious perception of the picture is suppressed by continuous flashing shapes.
  • Perceptual Expectation Effect
    • They cued participants with the word before each experiment trial and showed that cue word 70% of the time. By measuring how long it took for each image to emerge from flash suppression, they determined how quickly people consciously saw a particular image – when it was expected vs when it wasn’t.
    • Result was people were faster and more accurate at seeing houses when they were given house as cue word. Same for face cue. So, valid perceptual expectations indeed lead to more rapid and more accurate conscious perceptions.
    • Single letters like “U” are easier to identify when part of a word rather than part of some jumbled letters “RXUEV”. Word Superiority Effect.
      • They even managed to record fMRI and see that neural signature of a letter in visual cortex was sharper in word situation.

Brain as organ of experience

  • Experiencing hallucinations produced through LSD give us an insight that brain is the organ of experience, than a representative of real world.
  • Faces have a strong prior in our brains.
  • Hallucination is when perceptual prior is so strong that it overwhelms the sensory data so that brain’s grip on the causes in the world starts perishing.
  • Hallucinatory experiences can be created in laboratories by programming the hallucination machine. This exercise helps to understand brain in terms of computational phenonmenology, using computational models to build explantory bridges from mechanisms to properties of perceptual experience.
  • Objecthood is property of how visual conscious contents generally appear for us. e.g. Picasso’s cubist portfolio investigates that to find what exactly is the most minimal that we can go upto still feeling the same object. Artists try to use this idea to portray and make art pieces more meaningful.
  • Objecthood is mostly explored by “sensorimotor contingency theory” which says this: Experience depends on Practical Mastery of how actions change sensory inputs. So, based on actions and how sensations change, our brain perceives.
  • According to this theory, we become perceptually aware of back of the tomato even though we cannot directly see it, because of this implicit knowledge.
  • Situation where this objecthood is absent is “Grapheme-colour synaesthesia” where synaesthesia means mixing of the senses.
  • By understanding people having experiences of colour when seeing letters: like letter A eliciting redness, regardless of color on the page. They happen consistently and automatically. They don’t even get confused and these are perceived automatically by these brains.
  • Such kind of red that comes from synaesthesia is more discrete compared to the actual color perception which changes even when ambient lighting changes. In such cases, there would be no phenomenology of objecthood because it’s not something discrete.
  • Many experiments have shown that physical change is neither necessary nor sufficient for the perception of change. e.g. Perception of movement on a static picture and so on.
  • In “change blindness”, the opposite happens. Like us perceiving some changes only when a lot of things change at once. Like people won’t perceive change in color on one half of the video when it changes over a long period like a span of 40 seconds or so.
  • If experience of change are perceptual inferences, so too are experiences of time.
  • We experience seconds, minutes, hours, months etc. but don’t have a “time sensor” or such in our brains.
  • Experience of time is controlled hallucination as well. Except the circadian rhythm, there is no time experiencing logic in our brains. And circadian rhythm is exactly that: understanding time based on some other perceptual contents.
  • People underestimated duration of long videos and overestimate short videos in an experiment in library of videos of various duration. Busy scenes were estimated to be longer than quiet scenes, even for videos that were of same duration.
  • When using neural net for estimation using pixel changes as well, its estimates were virtually identical. So, this also shows that perception of time can emerge from “best guess” about rate of change of sensory signals, which seems to be what our brains are doing.
  • Then in another test, they tried to estimate duration from fMRI of people looking at same set of videos. And they arlso found the same thing. In visual cortex, it was predictive. In other regions, it was not so. So, that also proved to the fact that we were guessing time based on visual changes.
  • David Eagleman set out to test common intuition that subjective time slows down in moments of high drama, like bungee jumping or before car crash. So, he wanted to see if perception is actually sped up in such scenes. They designed a special digital watch which displayed a series of numbers that flickered quickly so that they were impossible to read in normal conditions. The assumption was that if the brain was perceiving faster in such scenes, the numbers should have been more visible during freefall. They couldn’t. So, it also showed that there was indeed no internal clock to perceive that faster.
  • David Hume in 18th century argued that we project causality out into the world on the basis of repeated perception of things happening close in time.
    • We don’t experience the models of our perception as models. We perceive with and through our generative models, and in doing so out of mere mechanism a structured world is brought forth. Read this multiple times as it is quite significant in terms of explaining how humans perceive reality.
  • Starting from the principle that brain must infer the hidden causes of its sensory inputs, our generative models do perceptual inference to give us the idea of what “seems to be” out there as reality.
  • A chair has a mind-independent existence; chairness does not.
  • Delirium – a word surfacing from the depths of the sixteenth century – comes from the Latin delirare, to deviate, to be deranged. Unlike dementia, which is a chronic and degenerative condition, delirium is usually temporary.
  • He explains his mother’s delirium as brain trying to find closure, trying to make its best guess. The brain is trying to find a good explanation of all the inputs that it has received, and worked its way up to explain it.
  • What is “self”? Where does it go when a person gets deluded once and then returns back to former self?
  • Self seems to be another controlled hallucination just like time.
  • Teletransportation Paradox
    • If we exactly replicate the molecules, as far as we understand the Science of it, there’s no doubt that it will be conscious in the same way as original.
    • Personal identity is psychological continuity, not physical continuity. So, once they are separate, they become different.
    • But both are as real as the other.
  • Kant argued that concept of self as “simple substance” is wrong. Hume talked about self as bundle of perceptions. Metzineger wrote that there’s no singular self. Buddhists say that there is no permanent self and one can reach an entirely selfless states of consciousness. DMT also strip away people’s sense of self.
  • Craniopagus twins share even some brain structures: then what is individual self to them?
  • Just like we perceive the differences in time between our own selves, the philosophical zombies of us will be different.
  • Embodied selfhood also relates to our body. There is a strong identification that happens. Emotions and moods are also aspects of embodied selfhood. Then comes FPV - a subjective POV for perceptual experience, which seems to reside somewhere between the eyes, and slightly behind our forehead. Then, experience of volition, intention, agency are also central to selfhood. This is what we term as “free-will”. All these before we link it to any kind of identity. Then comes our name, history, future etc.
  • Even the feeling of unified self can be brought down in cases like dementia, and in severe cases of amnesia. Even alien hand syndrome can cause this to be warped. Outer body experiences are more extreme examples of that. So, self is not an immutable entity that is there. It’s as if this collection that leads to arising of selfhood is neural encodings that are trying to keep your body alive.
  • What similarities do rubber hand syndrome and self-hood have over ideologies in Buddhist philosophy which too agrees that this self that you are compiling in your heads is not the real self at all?
  • And synchronicity of the action in rubber hand syndrome is quite important as well for the brain to get confused and perceive that rubber hand as its own. And this experiment has been expanded for the whole body as well.
  • There have been experiments like from Canadian neurologist Wilder Penfield whose patient would feel like she isn’t there when stimulated electrically in right superior temporal gyrus. And for Olaf Blanke when his patient said he saw himself lying in bed, from above, but only their legs when stimulated in angular gyrus. Its as if brain is trying its best guess for location of FPV as well.
  • A VR experiment done in 2008 for body swap Illusion by Swedish researcher Henrik Ehrsson. And in all these experiments, it depended on how susceptible the person is or how hypnotizable one is.
  • Non-human animals and infants also experience this embodied selfhood without being related to personal identity at all.
  • The story from Ghazini is fom a man who after suffering from herpes encephalitis got massive damages to hippocampus in both cerebral hemispheres. Thus, got both retrograde amnesia as well as anterograde amnesia. So, seemed to permanently exist in a permanent present of 7 mins and 30 seconds. So, he had to write this “I”/selfhood every time.
  • In case of Clive, love for his wife remains undiminished even though sometimes he cannot remember meeting her. And when Clive plays piano, or sings, or conducts, the music falls freely from him with a fluency that makes him seem whole again.
  • And us being social creatures, a lot of this is very necessary for accomodating oneself in the society as well. And social perception is complicated af by itself.
  • With expressions we are also trying to insert our thoughts into another’s mind.
  • The best model of anyone’s mental states will include a model of how to model someone else’s mental states as well.
  • We seem to feel like we haven’t changed in appearance. We feel like our self is all continuous when on deeper introspection, we even ourselves realize that it wasn’t. This is like the “change blindness” where slow moving things aren’t perceptible to us. And this fosters our idea that self is immutable. And the reason for perception is for control. To stay alive.
  • Rene Descartes first introduced this idea of mind and matter.
  • Descartes claim was that non-human animals was that they lacked rational, spiritual, and conscious attirbues that came along with having a soul.
  • Julien Offray de La Mettrie, then extended to this idea of beast-machine model that Descartes provided, and argued that humans were machines too (l’homme machine ).
  • From Descartes came Cartesian view of mind and probably the term.
  • Varieties of sensation and perception are collectively called exteroception. Perception of body from within is called interoception. Interoceptive signals come from internal organs. These signals cme from brainstem and thalamus before getting to cortex for introspective processing in the insular cortex.
  • Previously, emotions were seen as perceptions of changes in bodily states and not that it is a change in the mental state. We feel sorry because we cry, angry because we strike and so on.
  • Appraisal theories have solved the problem of emotional range because each specific emotion now no longer needs a bodily state to occur. We can see it very clearly from the hormonal changes as well.
  • e.g. Listlessness and Ennui might be based on same bodily state but completely different.
  • The female interviewer received more calls from men who had been crossing the rickety bridge than from those who had been crossing the sturdy bridge. In an experiment by Donald Dutton and Arthur Aron in 1974. Often these physical changes which look similar get confused by brain. Fear and Anxiety for arousal. They called this “Misattribution of arousal”. And when a male questioned men in same setting, there was no effect of type of bridge on the number of follow up calls. It illustrates that emotional experiences depend on how physiological changes are evaluated by higher level cognitive processes. Author calls this idea “Introspective Inference”.
  • So, this idea solves for emotional range shown without a distrinction between non-cognitive and cognitive state in the brain.
  • They even found that people sensed greater degree of ownership in a virtual reality rubber hand when it flashed their heartbeat matched in time than when it flashed out of time - suggesting that body ownership depends on the integation of both exteroceptive and interoceptive signals. Also depends on how aware a person is about their heartbeat though, so there are quite some interesting signals about what we can’t even measure accurately.
  • In terms of beast machine theory, one of the implications is that emotions and moods etc. come from inside out, and not the outside in. Now, we need to ask why our internal physiological condition be built into our conscious lives from the ground up.
  • Then comes the idea of cybernetics when thinking about machines - as potentially having purpose.
  • Good Regulator Theorem 1970 by William Ross Ashby and Roger had the title : “Every good regulator of a system must be a model of that system. "
    • When systems are not imperfect in terms of the data that they are being fed, then it can actually regulate it quite well.
    • In case of noisy sensor itself, then inference has to happen as best guess about the sensory signals. Similar to how this regulator has to somehow understand a little about how this noise is modeled without understanding the causality of noise, same happens with our perception. Its not for figuring out what’s there, its for control and regulation.
  • Control oriented perception is thus a form of active inference.
  • Idea of cybernetics is that, for some systems, control comes first.
  • To answer why perceptions of emotion and mood are needed, we can take the idea from cybernetics - that of an essential variable.
  • So, essential variables are physiological quantities like body temperature, sugar levels, oxygen levels etc. in case for a living organism to keep it alive. So, emotions and moods are control-oriented perceptions which regulate the body’s essential variables.
  • Predictions geared towards finding out things will have a very different phenomenology from those geared towards controlling things. e.g. catching a cricket ball is done by using optic acceleration cancellation where you just try to try to move such that ball “looks the same” in a particular way. Instead of trying to figure out complex mechanics and rules of where other fielders are or run as fast by figuring out where the ball would land.
  • So, the actions and brains predictions about sensory consequences are not about finding out where the ball is. They are geared towards controlling how the ball perceptually appears. Perception in this situation is a controlling hallucination just as much as it is a controlled hallucination.
  • James Gibson and idea of “affordances” as answer to how we perceive the world.
    • Affordance is an opportunity for action - a door for opening, a ball for catching etc. rather than an action-independent representation of the way things are.
  • From William Power’s Perceptual Control Theory, we don’t perceive to behave in a particular way. We behave so that we end up perceiving things in a particular way.
  • There is no phenomenology of objecthood, as when looking at a coffee cup on the table, nor is there anything like movement in a spatial frame, as when catching a cricket ball.
  • The control-oriented perceptions about emotions are basically trying to put body’s essential variables where they belong. So, that’s why instead of experiencing emotions as objects, we experience how well or badly our overall situation is going, and likely to go.
  • Beyond self of emotions and moods, there is this difficult experience of simply being alive.
  • Beast Machine Theory: Conscious experience of world around and inside happens because of our living bodies.
    • Primary goal for organism is to be alive. So, imperative that comes from evolution.
    • Brains exist to maintain good physiological integrity in the face of danger and opportunity.
    • Evolution gave brains not for any complex task but to help it stay alive, so that the organism’s essential physiologically essential variables are within compatible ranges for survival.
    • That maintenance of variables is cause of interoceptive signal.
    • Brain has no direct access to physiological states of the body, so these states have to be inferred through Bayesian best guessing.
    • Best guessing is done through prediction error minimisation. This is introspective inference a.k.a controlled hallucination.
    • So, it exemplifies active inference, errors are minimised by acting to fulfill top-down predictions for the most part.
    • This predictive control supports anticipatory responses, through predictions about future bodily states, and their dependence on this-or-that action.
    • Running away from a bear before being eaten is an example of external anticipatory regulation. The transient increase in blood pressure needed to run away effectively, or even to stand up from your desk after you’ve been working for a while, is an internal anticipatory response.
    • Achieving stability through change, as compared to homeostasis, which means tendency towards state of equilibrium.
    • So, introspective inference is allostatic regulation of physiological condition of the body.
    • The affective experiences are controlling hallucinations just as much as they are controlled hallucinations.
    • e.g. For most part, fear has an external reference. In deepest levels of experienced selfhood where one utterly feels alone for being doesn’t have an external reference. So, that’s the ground state of conscious selfhood according to the author. Formless, shapeless, control oriented perceptual prediction about present and future physiological condition of the body itself. In his own words, “This is where being you begins.”
    • Subjective stability of selfhood emerges naturally from the beast machine theory.
    • To effectively regulate the body’s physiological condition, priors on interoceptive signals need to have high precision, so that they tend to become self-fulfilling. e.g. Body temperature is predicted to be rather constant over time, which according to active inference, is why it turned this way. The need comes from the need for precise priors - strong predictions - about stable bodily states for good physiological regulation.
    • To put it another way: for as long as we live, the brain will never update its prior belief of expecting to be alive.
    • Brain might attenuate prior expectations related to perceiving changes in the condition of body, in order to further ensure that physiological essential variables stay where they ought to be. So, there we have explanation for “self-change blindness”. So, we are guided in a loop that strengthens the idea that no change is happening the more it happens and gets less and less surprising. So, as if good regulators think need to overestimate their stability than the true fact.
    • And in the same way that our perceptions of the world can sometimes lack the phenomenology of being real, the self too can lose its reality. During illness, this subjective stability can be severly attenuated, or even abolished in conditions where person loses identity.
    • He goes to this idea that “soul” is now the perceptual expression of deep continuity between mind and life.
    • Very deeply spiritual as well : For something to exist, it needs to have a boundary between this and that thing. That boundary must persist over time as well. e.g. Dropping a drop of oil on water stays as is, dropping a drop of ink on water doesn’t. So, the oil drop continues to exist because it has not dispersed itself evenly throughout the water. But neither oil drops nor rocks actively maintain their boundaries, they just get dispersed slowly enough for us to notice them as existing while this happens.
  • So, living systems, by virtue of existing, actively resist the dispersion of internal states.
  • Though by second law, its supposed to be improbable and unstable, living beings temporarily fend off this law.
  • According to Free Energy Principle, for a living system to resist the pull of second law, it must occupy the states which it expects to be in. Does it mean it predicts the disorder, that’s why there doesn’t seem any disorder at all?
  • To minimise a quantity, system has to be able to measure it. The problem here is that sensory entropy cannot be detected or measured.
  • Free Energy as a quantity which approximates sensory entropy. This is something organism can sense.
  • Free Energy from the perspective of organism seems to be basically the same thing as sensory prediction error.
  • FEP also uses the idea that living systems have or are models of their environment.
  • The ideas of predictive perception and controlled hallucination follow seamlessly from Friston’s FEP. In this view, organisms gather and model sensory information to maximise sensory evidence of their own existence. “I predict myself, therefore I am. "
  • In FEP, decreasing sensory prediction error can mean that organism can go stare at blank wall to reduce errors, but as bodily state changes like glucose levels, the signals start to deviate.
  • So, these systems because of the process by they are alive need to allow some things to change in order for other things to stay the same.
  • One major problem of FEP is that its not falsifiable. The theory of predictive processing, for example, can be falsified if it turns out that the brain doesn’t use sensory prediction errors in the process of perceiving.
  • Free Energy Principle by Karl Friston
  • FEP helps beast machine theory as well in these ways:
    • FEP gorunds this theory in the bedrock of Physics. The imperative to stay alive becomes more Physics based.
    • We start with things exist and proceed to the body and world on this theory compared to all the things done to go out to in in terms of the BMT.
    • The third benefit of the FEP lies in the rich mathematical toolbox that it brings to the table.
  • The FEP starts from the simple statement that ‘things exist’ and derives from this the whole of neuroscience and biology, but not consciousness. IIT starts from the simple statement ‘consciousness exists’ and from there launches a direct assault on the hard problem.
  • So, BMT is a middle ground. With FEP it has a theoretical grounding. With IIT, it shares phenomenological properties of consciousness.
  • The aspect that we seem to cling to most tightly is being in control of our actions and author of our narratives. This feels like I am taking the action and could have chosen otherwise. So, this experience of volition. But we have understood that these are constructed according to Bayesian best guessing.
  • Readiness Potential and experiment in UCSF by Libet in 1980s told us that we could predict the action before being aware of it in our motor cortex. So, free will has a complicated scene. 100s of milliseconds before the conscious intention to move became awareness in a person’s brain, readiness potential was already ramping up.
  • In 2012, Aaron Schurger made an experiment to show that readiness potential might not be signature of brain initiating an action, but might instead be the artefacts of the way they are measured. So, he asked this question whether readiness potential wasn’t looked at other times when there was no voluntary action. By modifying Libet’s experiment, he found out that it was spiking even when there wasn’t any voluntary action. So, basically disconnected it from the voluntary action. He used beep to trigger this.
  • One cannot choose the wants. “Man can do what he wills, but he cannot will what he wills. " said Schopenhauer.
  • Compared to reflex action, we have a sense that we could have done otherwise. I could have made coffee instead of tea and so on. So, we see action as freely willed when we infer that its causes came from within aligned with one’s beliefs and wills, we feel we had the choice to take the alternative action, and detached from alternative causes in either the body or the world.
  • So, in Patrick Haggard’s terms, its about “what” speficifying which action to make, “when” determining timing, and late-breaking “whether” specifying the choice.
    • “What” comes from organised sets of beliefs, goals, values etc. together with perception of environment, in order to specify a single action out of many possibilities.
    • “When” is all about timing. Localisable to supplementary motor area. Gentle electrical brain simulation of supplementary motor area can generate urge to move, even in absence of any movement.
    • “Whether” is more about decisioning. So, in the end if we have decided on action, its more about inhibit or not. Localisable to more frontal parts of brain.
  • And this loop into the body, world, and back again is what gives us sense of volition. Its more like self-fulfilling perceptual prediction. They are useful for guiding future behavior, just as much as for guiding present behavior.
  • The usefulness of feeling ‘I could have done otherwise’ is that, next time, you might.
  • Now, when we get to “state of flow”, it get complicated. The phenomenology of volition might be completely absent.
  • People with anarchic hand syndrome make voluntary actions which they do not experience as being theirs. Ones with akinetic mutism are unable to make any voluntary actions at all. Texas Tower Sniper is a story where a brain tumor can transform an engineering student into a mass school shooter in case of Charles Whitman. Or make a teacher pedophile which disappeared when tumour was removed, and returned when it grew back.
  • A conscious intention is as real as a visual experience of color.
  • Heraclitus: ‘No man ever steps in the same river twice, for it’s not the same river and he’s not the same man.’
  • From 9th to 18th century, animals were held criminally responsible and executed even. Pigs were the most common offenders in E.P. Evan’s 1906 documentation of history of animal criminal persecutions. Crimes from eating children to consuming consecrated biscuits.
  • Consciousness in animals cannot be inferred just from similarity to human consciousness. Fault of anthropomorphism and anthoropocentrism.
  • Consciousness and intelligence should not be associated too closely as well.
  • Of course its easier to say one is conscious if it can communicate in more intelligent forms like language and metacognition with us.
  • BMT since applies to the idea of being alive as primary seems to infer that consciousness might be more widespread.
  • Mammalian brains are very similar across species.

Comparing measurable attributes in animals for consciousness

2005 Bernard Baars, author and David Edelman came with a list of properties of human consciousness that could be tested on other animals. 17 properties. Finding animal consciousness is more like this movement from anthorpocentrism to a bigger space of possible conscious minds.

Anatomical Similarity

  • Anatomical similarity to human brain.

Sleep Dynamics

  • Changes in brain dynamics as animals fall asleep and wake up to show the dynamics of underlying conscious level. Because in normal waking states, they show irregular, low ampltitude, fast electrical activity. On sleeping states, they switch to more regular, large-amplitude brain dynamics. Anaesthesia has similar effect on mammalian speices too.
  • Behavioral differences like seals and dolphins sleep with their half brain at a time, koalas sleep 22 hours a day, Giraffes sleep about < 4 hrs, and newborn killer whales don’t sleep at all in their first month of life. So, there is too much diversity too. Dolphins don’t have REM, while seals have it only on land.

Contents of consciousness

  • Mice use whiskers, bats use echolocation, mole rats use sense of smell which builds their inner universe differently.
  • Mirror self recognition seems to develop in humans between 18 and 24 months of age. This idea is used to test animals as well.
  • There is also a self-recognition test developed by Gordon Gallup Jr in 1970s where animal is marked with paint or sticker on its body. The animal is allowed to interact with a mirror so that it can see the mark. If it looks for that mark on its own body instead of the image, it passed the test.
  • Few animals like dolphins, killer whales, etc. have passed the test. And there are quite some reasons why animal would fail other than being conscious like avoiding eye contact, not understanding how mirrors work etc.
  • Monkey

    • The author feels that hanging around with monkeys definitely feels like they are conscious.
    • Some monkeys have been seen to do certain things purely for fun.
    • They seem to show some reaction to deliberate unfairness in an experiment by Frans de Waal.
    • Having fun and throwing tantrums are powerful intuition pumps.
    • But monkeys have consistently failed the mirror test.
  • Octopus

    • Octopuses seem to be looking at things for sheer interest in experiments.
    • Very complicated animals indeed.
    • 1/2 billion neurons, roughly 6x of a mouse.
    • Do weird things with their genes.
    • It’s as if the octopus is able to rewrite parts of its own genome on the fly.
    • Have shown complex maze solving, opening jars etc.
    • RNA editing which seems to be related to their nervous system as well.
    • Does camouflage as well.
    • And they are color blind. The central brain might not even know what its skin is doing.
    • Octopus arms behave like semi-autonomous animals: a severed octopus arm can still execute complex action sequences, like grasping pieces of food, for some time after being separated from the body.
    • Body ownership might be quite a odd thing for an octopus because of the degree of freedom and decentralised control of their pods.
    • Octopuses secrete a distinctive chemical throughout their skin. This chemical serves as a signal that can be detected by the suckers, so that they do not reflexively attach. The detached arms would readily grip onto the arms from which the skin had been removed, but would never grip onto the intact arms.
  • Birds

    • Similar brain and thalamus to mammals.
    • Parrots can count, cockatoos cna dance, srub jays can stash food for future needs.
  • Insects

    • Recent study found that the fruit fly Drosophila melanogaster displays a post-injury hypersensitivity to previously non-painful stimuli in a way that resembles ‘chronic pain’ in humans. Also anaesthetic drugs seem to work over all kinds of brains equally.

Artificial Intelligence

  • Rarely the creations don’t turn on their creators.
  • Author too believes that consciousness will emerge naturally once machines pass some intelligence threshold.
    • Necessary condition being functionalism. Information processing by itself is not sufficient for consciousness.
    • Second assumption being information processing that is sufficient for consciousness is also sufficient for consciousness that underpins intelligence.
  • If we tie up consciousness and intelligence too much, we might ignore all non-intelligent conscious lives.
  • There cann be intelligence that is simply augmenting humans without being conscious as well.
  • According to IIT as well, if machine generates integrated information will have some degree of consciousness, regardless of what its made up of.
  • The imperative for self-maintenance and regulation in living systems isn’t restricted to just one level. Goes all the way down to the level of individual cells.
  • Hiroshi Ishiguro spent decades building robots that are similar to human beings which he calls “Geminoids”. Once used this to deliver a 45 min lecture to 150 undergrad students.
  • It goes to the idea of “Uncanny valley”, a concept introduced by , Mashiro Mori in 1970. Mori proposed that as a robot begins to look humanlike, it will elicit increasingly positive and empathetic reactions from people. But once it passes a certain point where it looks more like human but falls short in few aspects, it will rapidly turn into revusion and fear - the uncanny valley - only recovering as resemblance becomes even more closer (almost indistinguishable).
  • We already are surrounded with faces generated by GANs which never existed.
  • We even have the capability to develop fully synthetic organisms built from the genes up which was done in 2019 in Cambridge. It created a variant of E.Coli with a fully synthetic genome.
  • So, biotechnology might be one that brings us closest to synthetic consciousness.
  • Human minds even adapt to removing the hemispheres of brain which makes the study even more diverse and difficult. Now, all the regions and knowledge of it becomes useless in face of that situation.
  • Studying this conscious level comparing it to states like comatose using integrated information, casual density etc. are not random and started teaching us simple things.

Examples from movies

  • Rabbi Judah’s idea of Golem from clay from banks of Vltava river.
  • Mary Shelley idea of Frankenstein
  • Alex Garland idea of Ex Machina
  • James Cameron’s Terminator
  • Ridley Scott’s replicants in Blade Runner
  • Stanley Kubrick’s HAL

Beast Machine Theory Conclusion

  • Everything in conscious experience is a perception of sorts, and every perception is a kind of controlled – or controlling – hallucination. Experiences of free will is a perception. Experience of time is a perception. Perhaps even the 3d world and sense that contents of perceptual experience as objectively real could be our perceptions too.

Reorganization and summarization

Terminologies

  • Qualia

Different states of body and consciousness

Locked In Syndrome
Consciousness is fully present despite the total Body Paralysis. Mostly caused by damage to brainstem.

Philosophies

Proposed Theories and Philosophies

Measuring consciousness

Questions on consciousness

  • What does anaesthesia do to your consciousness? Where does the state of being temporarily get abolished to?
  • Where does subjective universe come into play?

Highlights

General anaesthesia doesn’t just work on your brain, or on your mind. It works on your consciousness. By altering the delicate electrochemical balance within the neural circuitry inside your head, the basic ground state of what it is to ‘be’ is – temporarily – abolished. In this process lies one of the greatest remaining mysteries in science, and in philosophy too.

The novelist Julian Barnes, in his meditation on mortality, puts it perfectly. When the end of consciousness comes there is nothing – really nothing – to be frightened of.

For a conscious creature, there is something that it is like to be that creature.

But there is almost certainly nothing it is like to be a bacterium, a blade of grass, or a toy robot. For these things, there is (presumably) never any subjective experience going on: no inner universe, no awareness, no consciousness.

This way of putting things is most closely associated with the philosopher Thomas Nagel, who in 1974 published a now legendary article called ‘What is it like to be a bat?’ in which he argued that while we humans could never experience the experiences of a bat, there nonetheless would be something it is like for the bat, to be a bat.

These properties are sometimes also called qualia in philosophy: the redness of red, the pang of jealousy, the sharp pain or dull throb of a toothache.

Consciousness is first and foremost about subjective experience – it is about phenomenology. This may seem obvious, but it wasn’t always so. At various times in the past, being conscious has been confused with having language, being intelligent, or exhibiting behaviour of a particular kind.

But consciousness does not depend on outward behaviour, as is clear during dreaming and for people suffering states of total bodily paralysis

Some prominent theories in the science of consciousness continue to emphasise function and behaviour over phenomenology. Foremost among these is the ‘global workspace’ theory, which has been developed over many years by the psychologist Bernard Baars and the neuroscientist Stanislas Dehaene, among others.

INTERESTING THEORY

According to this theory, mental content (perceptions, thoughts, emotions, and so on) becomes conscious when it gains access to a ‘workspace’, which – anatomically speaking – is distributed across frontal and parietal regions of the cortex. (The cerebral cortex is the massively folded outer surface of the brain, made up of tightly packed neurons.†) When mental content is broadcast within this cortical workspace, we are conscious of it, and it can be used to guide behaviour in much more flexible ways than is the case for unconscious perception.

Another prominent theory, called ‘higher-order thought’ theory, proposes that mental content becomes conscious when there is a ‘higher-level’ cognitive process that is somehow oriented towards it, rendering it conscious.

Like global workspace theory, higher-order thought theories also emphasise frontal brain regions as key for consciousness

When a complex phenomenon is incompletely understood, prematurely precise definitions can be constraining and even misleading.

How does consciousness happen? How do conscious experiences relate to the biophysical machinery inside our brains and our bodies? How indeed do they relate to the swirl of atoms or quarks or superstrings, or to whatever it is that the entirety of our universe ultimately consists in? The classic formulation of this question is known as the ‘hard problem’ of consciousness. This expression was coined by the Australian philosopher David Chalmers in the early 1990s, and it has set the agenda for much of consciousness science ever since.

Here is how he describes

It is widely agreed that experience arises from a physical basis, but we have no good explanation of why and how it so arises. Why should physical processing give rise to a rich inner life at all? It seems objectively unreasonable that it should, and yet it does.

The easy problems cover all the things that beings like us can do and that can be specified in terms of a function – how an input is transformed into an output – or in terms of a behaviour.

More precisely, for Chalmers there is no conceptual obstacle to easy problems eventually yielding to explanations in terms of physical mechanisms.

‘[E]ven when we have explained the performance of all the functions in the vicinity of experience – perceptual discrimination, categorization, internal access, verbal report – there may still remain a further unanswered question: Why is the performance of these functions accompanied by experience?’

René Descartes’ seventeenth-century sundering of the universe into mind stuff, res cogitans, and matter stuff, res extensa. This distinction inaugurated the philosophy of dualism, and has made all discussions of consciousness complicated and confusing ever since

My preferred philosophical position, and the default assumption of many neuroscientists, is physicalism. This is the idea that the universe is made of physical stuff, and that conscious states are either identical to, or somehow emerge from, particular arrangements of this physical stuff. Some philosophers use the term materialism instead of physicalism, but for our purposes they can be treated synonymously.

At the other extreme to physicalism is idealism. This is the idea – often associated with the eighteenth-century Bishop George Berkeley – that consciousness or mind is the ultimate source of reality, not physical stuff or matter. The problem isn’t how mind emerges from matter, but how matter emerges from mind. Sitting awkwardly in the middle, dualists like Descartes believe that consciousness (mind) and physical matter are separate substances or modes of existence, raising the tricky problem of how they ever interact.

There are two more ‘isms’, then we’re done

The first is panpsychism.

There are more sophisticated versions of the idea, some of which we will meet in later chapters, but the main problems with panpsychism don’t lie with its apparent craziness – after all, some crazy ideas turn out to be true, or at least useful. The main problems are that it doesn’t really explain anything and that it doesn’t lead to testable hypotheses. It’s an easy get-out to the apparent mystery posed by the hard problem, and taking it on ushers the science of consciousness down an empirical dead end.

Finally, there’s mysterianism, which is associated with the philosopher Colin McGinn.

physical understanding of consciousness exists, but it lies as far beyond us as an understanding of cryptocurrency lies beyond frogs. It is cognitively closed to us by our species-specific mental limitations

One particularly influential flavour of physicalism is functionalism.

Functionalism is the idea that consciousness does not depend on what a system is made of (its physical constitution), but only on what the system does, on the functions it performs, on how it transforms inputs into outputs. The intuition driving functionalism is that mind and consciousness are forms of information processing which can be implemented by brains, but for which biological brains are not strictly necessary.

This means that if you get the functional relations right – if you ensure that a system has the right kind of ‘input–output mappings’ – then this will be enough to give rise to consciousness. In other words, for functionalists, simulation means instantiation – it means coming into being, in reality.

Is consciousness more like Go or more like the weather? Don’t expect an answer – there isn’t one, at least not yet. It’s enough to appreciate that there’s a valid question here. This is why I’m agnostic about functionalism.

A philosophical zombie is a creature that is indistinguishable from a conscious creature, but which lacks consciousness

Here’s why the zombie idea is supposed to provide an argument against physicalist explanations of consciousness. If you can imagine a zombie, this means you can conceive of a world that is indistinguishable from our world, but in which no consciousness is happening. And if you can conceive of such a world, then consciousness cannot be a physical phenomenon. And here’s why it doesn’t work. The zombie argument, like many thought experiments that take aim at physicalism, is a conceivability argument, and conceivability arguments are intrinsically weak. Like many such arguments, it has a plausibility that is inversely related to the amount of knowledge one has

Whether something is conceivable or not is often a psychological observation about the person doing the conceiving, not an insight into the nature of reality. This is the weakness of zombies. We are asked to imagine the unimaginable, and through this act of illusory comprehension conclusions are drawn about the limits of physicalist explanation.

According to the real problem, the primary goals of consciousness science are to explain, predict, and control the phenomenological properties of conscious experience.

The real problem is distinct from the hard problem, because it is not – at least not in the first instance – about explaining why and how consciousness is part of the universe in the first place. It does not hunt for a special sauce that can magic consciousness from mere mechanism (or the other way around).

Francis Crick (the co-discoverer, with Rosalind Franklin and James Watson, of the molecular structure of DNA) and his colleague Christof Koch, who were both based in San Diego in California, were setting out what would become the dominant method in the rise of consciousness science – the search for the neural correlates of consciousness.

gold-standard definition of a neural correlate of consciousness, or NCC, is ‘the minimal neuronal mechanisms jointly sufficient for any one specific conscious percept’. The NCC approach proposes that there is some specific pattern of neural activity that is responsible for any and every experience, such as the experience of ‘seeing red’. Whenever this activity is present, an experience of redness will happen, and whenever it isn’t, it won’t

‘binocular rivalry’

In binocular rivalry, a different image is shown to each eye – perhaps a picture of a face to the left eye and a picture of a house to the right eye. In this situation, conscious perception doesn’t settle on a weird face-house chimera. It flips back and forth between the face and the house, dwelling for a few seconds on each. First you see a house, then a face, then a house again … and so on. What’s important here is that conscious perception changes even though the sensory input remains constant

The NCC strategy has been impressively productive over many years, delivering reams of fascinating findings, but its limitations are becoming apparent. One problem is that it is difficult, and perhaps in the end impossible, to disentangle a ‘true’ NCC from a range of potentially confounding factors, the most important of which are those neural happenings that are either prerequisites for, or consequences of, an NCC itself.

Not so long ago, life seemed as mysterious as consciousness does today

This philosophy of vitalism reached a peak in the nineteenth century. It was supported by leading biologists like Johannes Müller and Louis Pasteur, and it persisted well into the twentieth century. Vitalists thought that the property of being alive could only be explained by appealing to some special sauce: a spark of life, an elán vital.

The fatal flaw of vitalism was to interpret a failure of imagination as an insight into necessity. This is the same flaw that lies at the heart of the zombie argument.

course, the parallel between life and consciousness is not perfect. Most conspicuously, the properties of life are objectively describable, whereas the explanatory targets of consciousness science are subjective – they exist only in the first person. However, this is not an insurmountable barrier; it mostly means the relevant data, because they are subjective, are harder to collect

Conscious level concerns ‘how conscious we are’ – on a scale from complete absence of any conscious experience at all, as in coma or brain death, all the way to vivid states of awareness that accompany normal waking life.

Conscious content is about what we are conscious of – the sights, sounds, smells, emotions, moods, thoughts, and beliefs that make up our inner universe. Conscious contents are all varieties of perception – brain-based interpretations of sensory signals that collectively make up our conscious experiences. (Perception, as we will see, can be both conscious and unconscious.)

there’s conscious self – the specific experience of being you, and the guiding theme of this book

The experience of ‘being a self’ is a subset of conscious contents, encompassing experiences of having a particular body, a first-person perspective, a set of unique memories, as well as experiences of moods, emotions, and ‘free will’.

Measuring Consciousness

The history of thermometry, and its impact on our understanding of heat, offers a vivid example of how the ability to make detailed quantitative measurements, on a scale defined by fixed points, has the power to transform something mysterious into something comprehensible. Could the same approach work for consciousness?

In fact, brain-based consciousness monitors have been deployed in operating theatres for many years already. The most common is the ‘bispectral index’ monitor

deeper problem, when it comes to consciousness science, is that the bispectral index is not based on any principled theory

Conscious level is not the same thing as physiological arousal. While the two are often highly correlated, consciousness (awareness) and wakefulness (arousal) can come apart in various ways, which is enough to show that they cannot depend on the same underlying biology.

vegetative state (also now known as ‘unresponsive wakefulness syndrome’), in which a person still cycles through sleep and wakefulness, but shows no behavioural signs of conscious awareness: the lights are occasionally on but there’s nobody home

Could it simply be the number of neurons involved? It doesn’t seem so. The cerebellum (the ‘little brain’ hanging off the back of your cortex) has about four times as many neurons as the rest of the brain put together, but seems barely involved in consciousness.

What about the overall degree of neuronal activity? Is the brain generally more active in conscious states than in unconscious states? Well, maybe, to some extent – but not by much. Although there are differences in the brain’s energy consumption across conscious levels, these differences are rather small and certainly there is no sense in which the brain ‘shuts down’ when consciousness fades.

Consciousness instead seems to depend on how different parts of the brain speak to each other.

The latest and most exciting approaches to measuring conscious level – and distinguishing it from wakefulness – are based on tracking and quantifying these interactions.

To test how different parts of the cortex were talking to each other, they stimulated activity in one location and recorded how this pulse of activity spread to other cortical regions over space and time

They did this by combining two techniques: EEG and transcranial magnetic stimulation (TMS). A TMS rig is a precisely controlled electromagnet which allows a researcher to inject a short and sharp pulse of energy directly into the brain through the skull, while EEG in this case is used to record the brain’s response to this zapping. It’s like banging on the brain with an electrical hammer and listening to the echo.

But for the most part, the huge disturbance in brain activity caused by TMS does not generate any alteration in conscious experience at all. Maybe this is not so surprising. It just shows we are not aware of what our neurons are doing – and why should we be

In unconscious states, like dreamless sleep and general anaesthesia, these echoes are very simple. There is a strong initial response in the part of the brain that was zapped, but this response dies away quickly, like the ripples caused by throwing a stone into still water. But during conscious states, the response is very different: a typical echo ranges widely over the cortical surface, disappearing and reappearing in complex patterns.

The complexity of these patterns, across space and time, implies that different parts of the brain – in particular the thalamocortical system – are communicating with each other in much more sophisticated ways during conscious states than during unconscious states.

It is possible to put a number to it to specify the magnitude of complexity. The approach is called ‘zap and zip’: use TMS to zap the cortex, and use a computer algorithm to ‘zip’ the response, the electrical echo, into a single number

The ‘zip’ algorithm – which calculates what’s called ‘Lempel-Ziv-Welch complexity’, or ‘LZW complexity’ for short – is a popular way of estimating the algorithmic complexity for any given sequence.

Massimini and his team called their measure of the echoes recorded in their experiments the perturbational complexity index, or PCI. It uses LZW complexity to provide a measure (index) of the algorithmic complexity of the brain’s response to a perturbation – the TMS pulse.

the PCI values of a large number of patients with brain injuries who had disorders of consciousness. They found that PCI magnitudes correlated extremely well with levels of impairment, as independently diagnosed by neurologists. For example, people in a vegetative state, where consciousness is assumed to be absent despite preserved wakefulness, had lower PCI scores than those in minimally conscious states, in which behavioural signs of consciousness come and go. They were even able to draw a dividing line between PCI values indicative of consciousness from those suggesting its absence

In my research group at the University of Sussex, we’ve been working on similar methods to assess conscious level. But instead of using TMS to inject pulses of energy into the cortex, we’ve been measuring the algorithmic complexity of ongoing, natural – what we call ‘spontaneous’ – brain activity

we discovered that the complexity of spontaneous cortical activity – as measured by EEG – reliably drops in both early sleep and anaesthesia. We also found that complexity during rapid eye movement (REM) sleep is much the same as during normal conscious wakefulness, which makes sense because REM sleep is when dreaming is most likely – and dreams are conscious. Massimini and his team had found the same pattern of results with their PCI measure, further supporting the claim that these measures are tracking conscious level rather than wakefulness.

An extreme example is ‘locked-in syndrome’, where consciousness is fully present despite total paralysis of the body. This rare affliction can follow damage to the brainstem, a region at the base of the brain (and at the top of the spinal cord) which, among other roles, mediates control of muscles in the body and in the face

A former editor of Elle magazine – Jean-Dominique Bauby, who became locked-in in 1995 following a brain haemorrhage – wrote an entire book this way, The Diving Bell and the Butterfly

In Massimini’s 2013 study, locked-in patients had PCI values indistinguishable from healthy age-matched controls – indicating fully intact consciousness

The more challenging cases arise in the grey zone between life and death, in conditions like the vegetative state and the minimally conscious state.

However, studies with healthy people have shown that the brain regions engaged by imagining fluent movements (like playing tennis) are highly distinct from those activated by imagining navigating through spaces.

Owen concluded that the behavioural diagnosis of vegetative state was wrong, and that the young woman was in fact conscious. In effect, Owen and his team had repurposed a brain scanner to allow his patient to interact with her environment using her brain rather than her body

To pass the Owen test, patients still have to understand language and engage in extended periods of mental imagery, which some may not be able to do, despite being conscious. It is here that new methods like PCI are especially significant, since they promise the ability to detect residual awareness without requiring the patient to do anything at all. Just as a true consciousness meter should.

Is a baby less conscious than an adult? Is a tortoise less conscious than either?

There is of course a danger in thinking along these lines. Such questions seduce us into presuming that any form of consciousness that diverges from the healthy adult human is somehow lesser, or lower. This way of thinking is symptomatic of the human exceptionalism that has repeatedly plagued biology, as it has darkened the history of human thought everywhere.

An important consequence of taking multidimensional levels of consciousness seriously is that sharp distinctions between conscious level and conscious content disappear. It becomes meaningless to completely separate how conscious you are from what you are conscious of. A ‘one size fits all’ measure of consciousness – of the sort we might expect if we take the temperature analogy too literally – may never be enough.

One example of how conscious level and conscious content interact comes from a study we carried out a few years ago on brain activity in the psychedelic state. Among their many uses, psychedelic drugs offer unique opportunities for consciousness science since they induce profound alterations in conscious contents, arising from a simple pharmacological intervention in the brain.

In the psychedelic state, vivid perceptual hallucinations are frequently accompanied by unusual experiences of selfhood often described as ‘ego dissolution’, where the boundaries between self and world, and other people, appear to shift or dissolve. These departures from ‘normal’ conscious experience are so pervasive that the psychedelic state might represent not only a change in conscious contents, but also a change in overall conscious level

At the level of neurochemistry, the classic psychedelics – LSD, psilocybin, mescaline, and dimethyltryptamine (DMT, the active ingredient in the South American hallucinogenic brew called ayahuasca) – work primarily by affecting the brain’s serotonin system

Psychedelic drugs influence the serotonin system by binding strongly to a specific serotonin receptor, the 5-HT2a receptor, which is found throughout much of the brain. One of the main challenges for psychedelic research is to understand how these low-level pharmacological interventions alter global patterns of brain activity, so as to deliver profound changes in conscious experience.

Robin’s team had previously discovered that the psychedelic state involves striking alterations in brain dynamics, when compared to a placebo control condition.

results were clear and surprising: psilocybin, LSD, and ketamine all led to increases when compared to a placebo control. This was the first time anyone had seen an increase in a measure of conscious level, relative to a baseline of waking rest

The results from our psychedelic analyses raised a disturbing prospect. Would maximally random brain activity, as measured by algorithmic complexity, lead to a maximally psychedelic experience? Or to a different ‘level’ of consciousness of some other kind? The extrapolation seems unlikely. A brain with all its neurons firing willy-nilly would seem more likely to give rise to no conscious experience at all, just as free-form jazz at some point stops being music.

The issue here is that algorithmic complexity is a poor approximation to what ‘being complex’ usually means. Intuitively, complexity is not the same as randomness. A more satisfying notion of complexity is as the middle ground between order and disorder – not the extreme point of disorder

They made a simple but profound observation: that conscious experiences – all conscious experiences – are both informative and integrated.

From this starting point, they made claims about the neural basis of every conscious experience, not just of specific experiences of seeing red, or feeling jealous, or suffering a toothache.

Every conscious experience therefore delivers a massive reduction of uncertainty, since this experience is being had, and not that experience, or that experience, and so on. And reduction of uncertainty is – mathematically – what is meant by information.

Scoring high on information is not by itself enough. Conscious experiences are not only highly informative, they are also integrated.

We do not experience colours separately from their shapes, nor objects independently of their background.

The key move made by Tononi and Edelman was to propose that if every conscious experience is both informative and unified at the level of phenomenology, then the neural mechanisms underlying conscious experiences should also exhibit both of these properties. That it is in virtue of expressing both of these properties that neural mechanisms do not merely correlate with, but actually account for, core phenomenological features of every conscious experience

At one extreme (on the left) all the elements behave randomly and independently, like molecules in a gas. This kind of system has maximum information – maximum randomness – but shows no integration at all, because every element is independent from every other.

At the other extreme (on the right), all the elements do exactly the same thing, so that the state of each element is completely determined by the state of the other elements in the system. No randomness at all. This would be like the arrangement of atoms in a crystal lattice, in which the position of any individual atom is fully determined by the lattice structure, which is defined by the positions of all the other atoms. This kind of arrangement has maximum integration, but almost no information, because there are very few possible states that the system can be in

In the middle lie systems in which individual elements may do different things, but in which there is some degree of co-ordination so that the system behaves to some extent ‘as a whole’. This is the domain where both integration and information are to be found. It is also the middle ground between order and disorder, and it is here that systems are typically called ‘complex’.

In a maximally information-rich brain, all neurons would behave independently, firing randomly as if they were completely disconnected. In such a brain, measures of algorithmic complexity, like LZW complexity, would score very high

But this brain – with lots of information but no integration – would not support any conscious states. At the other extreme, a maximally ordered brain would have all neurons doing exactly the same thing, perhaps firing in lockstep together, somewhat like what happens during global epileptic seizures. Algorithmic complexity here would be very low. This brain would also lack consciousness, but for a different reason: lots of integration, no information.

What we are looking for are measures that are directly sensitive to both integration and information, from the same data, in the same way, and at the same time.

There are several measures that do fulfil these criteria, at least in theory. Back in the 1990s, Tononi and Edelman, together with their colleague Olaf Sporns, came up with a measure they called ‘neural complexity’, and ten years later I derived my own measure, which I called ‘causal density’, using a different kind of mathematics.

All these measures attempt to quantify, in one way or another, the extent to which a system occupies the middle ground between order and disorder, where integrated information is to be found. The problem, however, is that none has yet worked particularly well when applied to actual brain imaging data.

IIT says that subjective experience is a property of patterns of cause and effect, that information is as real as mass or energy, and that even atoms may be a little bit conscious.

Back in 2006, IIT was not well known. Today, it’s one of the highest-profile but also most hotly debated theories in consciousness science. It’s been eulogised by some of the biggest names in the field, besides Tononi himself.

One reason for this pushback is that its approach is deeply mathematical and unapologetically complex.

For me, the bigger problem is that IIT’s extraordinary claims require extraordinary evidence, yet it is precisely the ambition of IIT – to solve the hard problem – that renders its most distinctive claims untestable in practice.

At the core of IIT is a single measure called ‘Φ’ (the Greek letter phi, pronounced fy). The easiest way to think about Φ is that it measures how much a system is ‘more than the sum’ of its parts, in terms of information.

How can a system be more than the sum of its parts? A flock of birds provides a loose analogy: the flock seems to be more than the sum of the birds that make it up – it seems to have a ‘life of its own’. IIT takes this idea and translates it into the domain of information.

This underpins the main claim of the theory, which is that a system is conscious to the extent that its whole generates more information than its parts.

According to IIT, the level of Φ is intrinsic to a system (meaning that it does not depend on an external observer), and it is identical to the amount of consciousness associated with that system. High Φ, lots of consciousness. Zero Φ, no consciousness.

To have high Φ, the network has to satisfy two main conditions. First, the global state of the network – the network ‘as a whole’ – has to rule out a large number of alternative possible global states. This is information, and it reflects the observation from phenomenology that every conscious experience rules out a great many alternative possible conscious experiences. Second, there must be more information when considering the system as a whole than when dividing it into its parts (its individual neurons, or groups of neurons) and considering all the parts separately. This is integration, and it reflects the observation that all conscious experiences are unified, that they are experienced ‘all of a piece’. Φ is a way of putting a number to a system that measures how high it scores on both these dimensions.

There are many ways a system can fail to have high Φ. One is to score low on information. A minimal example is a single photodiode – a simple light sensor which can be either ‘on’ or ‘off’.

Another way a system can have low Φ is to score low on integration. Imagine a large array of photodiodes, perhaps like the sensor in your phone’s camera. The global state of the system is the state of the entire array, and this can carry a great deal of information. A sufficiently large sensor array will enter a different global state for every different state of the world it encounters, which is why cameras are useful. But this global information does not matter to the sensor itself. The individual photodiodes in the sensor are all causally independent from one another – their state depends only on the level of light they each encounter. Chop the sensor into a bunch of (causally independent) photodiodes and it will work just as well. The information conveyed by the sensor array as a whole is not more than that conveyed by all the sensors, all the photodiodes, independently. This means that the information it generates is no more than the sum of its parts and so its Φ will also be zero.

This example emphasises how Φ depends on the optimal way in which a system can be ‘cut up’ to minimise the difference between what the whole does and what the parts do. This is one of the distinctive aspects of IIT which sets it apart from the measures of complexity described in the previous chapter.

This example also implies that a real split brain – following surgery to divide the cortical hemispheres, as happens in some cases of otherwise untreatable epilepsy – might harbour two independent ‘consciousnesses’, but that there will be no single conscious entity spanning both hemispheres.

Remember from the previous chapter that the cerebellum doesn’t seem much involved in consciousness, despite containing about three quarters of all neurons in the brain. This is explained by IIT because the anatomy of the cerebellum is comparable to the sensor array in your camera – a vast number of Φ-unfriendly semi-independent circuits. The cerebral cortex, by contrast, is packed with densely interconnected wiring that is likely to be associated with high Φ

So then why does consciousness fade during dreamless sleep, anaesthesia, and coma, given that this wiring doesn’t change? IIT says that, in these states, the ability of cortical neurons to interact with other is compromised in ways such that Φ vanishes.

On IIT, any mechanism that has these properties, whether a brain or not, whether biological or not, will have non-zero Φ, and will have consciousness.

The primary claim of the theory is that the level of consciousness for a system is given by its Φ. Testing this requires measuring Φ for real systems, and this is where the trouble starts. It turns out that measuring Φ is extremely challenging and in most cases nearly or actually impossible. The main reason for this is that IIT treats ‘information’ in an unusual way.

standard use of information in mathematics, developed by Claude Shannon in the 1950s, is observer-relative

Observer-relative (or extrinsic) information is the degree to which uncertainty is reduced, from the perspective of an observer, by observing a system in a particular state. For example, imagine rolling a single die many times. Each time, you observe one outcome out of six possibilities: each time, five alternatives are ruled out. This corresponds to a reduction in uncertainty of a particular amount (measured in bits), and is information that is ‘for’ the observer.

An external observer can record all the different states the neurons enter into, calculate the probabilities associated with each state, and then measure the reduction in uncertainty associated with the network being in any one of these states.

For IIT, however, information cannot be treated in this observer-relative way. This is because on IIT information – integrated information, Φ – actually is consciousness, and so if we treat information as observer-relative then this would mean that consciousness is also observer-relative. But it isn’t observer-relative. Whether I am conscious or not should not and does not depend on how you or anyone else measures my brain.

It must be defined in a way that does not depend on any external observer. It must be information ‘for’ the system itself – not for anyone or anything else. If not, the identity relation between Φ and consciousness at the core of IIT cannot hold.

You – as the scientist, the external observer – have to know all the different ways a system could behave, even if it never actually behaves in all these ways. The distinction is between knowing what a system actually does over time (which is easy, at least in principle, and is observer-relative) and knowing what a system could do even if it never does it (which is usually difficult, if not impossible, but is observer-independent).

Compared to measuring the empirical distribution of a system by just observing it over time, measuring the maximum entropy distribution is in general a very difficult thing to do.

There are two ways one might go about it. The first is to perturb the system in all possible ways and see what happens, just as a child might push all the buttons on a new toy to see all the things it can do. The second is to infer the maximum entropy distribution from an exhaustive, complete knowledge of the system’s physical mechanism – its ‘cause–effect structure’. If you know everything about a mechanism, it is sometimes possible to know all the things it could do, even if in practice it doesn’t do them. If I know that a die has six sides, I can figure out that two dice can generate all the numbers from two to twelve without having to make a single throw.

Unfortunately, often all we have access to are the dynamics of a system, to what a system does, rather than to what it could do. This is certainly true for brains.

but in general this is a very difficult problem, since the number of possible ways of dividing up a system grows exponentially with its size.

Then there are even more fundamental questions about what counts as a system in the first place. What is the right granularity of space and time over which to calculate Φ? Is it neurons and milliseconds, or atoms and femtoseconds?

Just because a theory is difficult to test doesn’t mean it’s wrong. It just means it’s difficult to test.

It turns out that following IIT all the way through leads to some very weird consequences. Imagine I open up your skull and glom onto your brain a fistful of new neurons, each of which wires up to your existing grey matter in some specific way. Imagine further that, as you go about your day, these new neurons never actually do anything at all. No matter what happens, no matter what you do or who you see, they never fire. Your newly augmented brain appears, to all intents and purposes, just the same as the old one. But, here’s the thing, your new neurons are organised in such a way that they could fire if only the rest of your brain encountered some particular state which it never in fact encounters.

For example, let’s say these new neurons would only ever fire if you were to eat a Densuke watermelon, a rare fruit found only on the Japanese island of Hokkaido. Assuming that you never actually eat a Densuke, so that these new neurons never fire, IIT nevertheless predicts that all your conscious experiences would change – albeit very subtly. This is because there are now more potential states your brain could be in – the new neurons could potentially fire – so Φ must change too.

Even though the overall activity of the brain has not changed at all, IIT would again predict a change in conscious experience, since there are now fewer potential states the brain could enter into.

Optogenetics uses genetic techniques to modify specific neurons so that they become sensitive to light at specific wavelengths. Then, by using lasers or LED arrays to shine light into the brains of genetically modified animals, these neurons can be switched on or switched off. In principle, optogenetics could be used to inactivate already inactive neurons, with the effects on conscious perception – if any – being assessed.

Zooming out, another weirdness of IIT is that by making the strong claim that Φ is consciousness, IIT also implies that information itself exists – that it has some definite ontological status in our universe – a status like mass/energy and electrical charge.

And this leads to a final challenging implication: panpsychism. So long as there is the right kind of mechanism, the right kind of cause–effect structure in a system, there will be non-zero Φ, and there will be consciousness. IIT’s panpsychism is a restrained panpsychism, not the sort in which consciousness is spread out through the entire universe like a thin layer of jam. Rather, consciousness is to be found wherever integrated information – Φ – is to be found. This could be here and there, but not everywhere.

My colleagues Adam Barrett and Pedro Mediano and I have been following this strategy for many years now. We’ve developed several versions of Φ that work with observer-relative information, rather than with intrinsic information.

This allows us to measure Φ based on the observable behaviour of a system over time, without worrying about what it could do but never does.

Whenever we are conscious, we are conscious of something, or of many things. These are the contents of consciousness. To understand how they come about, and what I mean by controlled hallucination, let’s change our perspective. Imagine, for a moment, that you are a brain.

When trying to form perceptions, all the brain has to go on is a constant barrage of electrical signals which are only indirectly related to things out there in the world, whatever they may be. These sensory inputs don’t come with labels attached (‘I’m from a cup of coffee’, ‘I’m from a tree’). They don’t even arrive with labels announcing their modality – whether they are visual, auditory, sensations of touch, or from less familiar modalities such as thermoception (sense of temperature) or proprioception (sense of body position).

How does the brain transform these inherently ambiguous sensory signals into a coherent perceptual world full of objects, people, and places? In Part Two of this book, we explore the idea that the brain is a ‘prediction machine’, and that what we see, hear, and feel is nothing more than the brain’s ‘best guess’ of the causes of its sensory inputs. Following this idea all the way through, we will see that the contents of consciousness are a kind of waking dream – a controlled hallucination – that is both more than and less than whatever the real world really is.

Signals flowing in the opposite direction – from the ‘top down’ or the ‘inside out’, the small grey arrows – serve only to refine or otherwise constrain the all-important bottom-up flow of sensory information.

This bottom-up view of perception fits well with what we know about the anatomy of the brain, at least at first glance. In the cortex, different perceptual modalities are associated with specific regions: visual cortex, auditory cortex, and so on. Within each region, perceptual processing is organised hierarchically. In the visual system, lower levels such as the primary visual cortex are close to sensory inputs, while higher levels, such as the inferotemporal cortex, are several stages of processing further away. In terms of connectivity, signals from each level are pooled together in the level above, so that neurons in higher levels can respond to features that may be spread out over space or time – just as one would expect.

Although it seems as though the sun goes around the Earth, it is of course the Earth rotating around its own axis that gives us night and day, and it is the sun, not the Earth, that sits at the centre of the solar system.

Although it may seem as though my senses provide transparent windows onto a mind-independent reality, and that perception is a process of ‘reading out’ sensory data, what’s really going on is – I believe – quite different.

The allegory is that our own conscious perceptions are just like these shadows, indirect reflections of hidden causes that we can never directly encounter.

Arab scholar Ibn al Haytham wrote that perception, in the here and now, depends on processes of ‘judgement and inference’ rather than providing direct access to an objective reality. Hundreds of years later again, Immanuel Kant realised that the chaos of unrestricted sensory data would always remain meaningless without being given structure by pre-existing conceptions, which for him included a priori frameworks like space and time. Kant’s term noumenon refers to ‘things in themselves’ – Ding an sich – a mind-independent reality that will always be inaccessible to human perception, hidden behind a sensory veil

Helmholtz proposed the idea of perception as a process of ‘unconscious inference’.

Helmholtz’s central idea of ‘perception as inference’ has been remarkably influential, taking on many different forms throughout the twentieth century. In the 1950s, the ‘new look’ movement in psychology emphasised how social and cultural factors could influence perception. For example, one widely circulated study found that children from poor families overestimated the size of coins, while those from well-to-do families didn’t. Unfortunately, many experiments of this kind – while fascinating – were poorly done by today’s methodological standards, so the results can’t always be trusted.

Perceptual content, for Gregory, is determined by the brain’s best-supported hypotheses.

A variety of new theories have blossomed under the general headings of ‘predictive coding’ and ‘predictive processing’. Although these theories differ in their details, they share the common proposal that perception depends on brain-based inference of some kind.

The essential ingredients of the controlled hallucination view, as I think of it, are as follows.

First, the brain is constantly making predictions about the causes of its sensory signals, predictions which cascade in a top-down direction through the brain’s perceptual hierarchies (the grey arrows in the image opposite). If you happen to be looking at a coffee cup, your visual cortex will be formulating predictions about the causes of the sensory signals that originate from this coffee cup. Second, sensory signals – which stream into the brain from the bottom up, or outside in – keep these perceptual predictions tied in useful ways to their causes: in this case, a coffee cup. These signals serve as prediction errors registering the difference between what the brain expects and what it gets at every level of processing. By adjusting top-down predictions so as to suppress bottom-up prediction errors, the brain’s perceptual best guesses maintain their grip on their causes in the world. In this view, perception happens through a continual process of prediction error minimisation.

The third and most important ingredient in the controlled hallucination view is the claim that perceptual experience – in this case the subjective experience of ‘seeing a coffee cup’ – is determined by the content of the (top-down) predictions, and not by the (bottom-up) sensory signals. We never experience sensory signals themselves, we only ever experience interpretations of them. Mix these ingredients together and we’ve cooked up a Copernican inversion for how to think about perception. It seems as though the world is revealed directly to our conscious minds through our sensory organs. With this mindset, it is natural to think of perception as a process of bottom-up feature detection – a ‘reading’ of the world around us. But what we actually perceive is a top-down, inside-out neuronal fantasy that is reined in by reality, not a transparent window onto whatever that reality may be.

The difference is that in ‘normal’ perception, what we perceive is tied to – controlled by – causes in the world, whereas in the case of hallucination our perceptions have, to some extent, lost their grip on these causes.

If perception is controlled hallucination, then – equally – hallucination can be thought of as uncontrolled perception. They are different, but to ask where to draw the line is like asking where the boundary is between day and night.

Ask a neurophysiologist and she may say that you perceive a particular colour when the colour-sensitive cone cells in your retina are activated in a certain proportion. This is not wrong, but it’s far from the whole story. There is no one-to-one mapping between activities of colour-sensitive cells and colour experience. The colour you experience depends on a complex interplay between the light reflected from a surface and the general illumination within the environment in which you happen to be. More precisely, it depends on how your brain makes inferences – best guesses –about how this interaction plays out.

When I look at a red chair, the redness I experience depends both on properties of the chair and on properties of my brain. It corresponds to the content of a set of perceptual predictions about the ways in which a specific kind of surface reflects light. There is no redness-as-such in the world or in the brain. As Paul Cézanne said, ‘colour is the place where our brain and the universe meet.’

People quickly pointed out that as a photo The Dress is overexposed and lacking in context – the dress itself fills most of the image – which might play tricks with how the brain generates colour from context. If, for some reason, your visual system is accustomed to yellowish ambient light – perhaps you spend too much time indoors – then your visual system might be more likely to infer blue–black on the assumption of a yellowish illuminant. If, on the other hand, you are a happy healthy outdoor person, with a visual cortex frequently bathed in blueish sunlight, then maybe you’d see white–gold.

Collectively, these examples – while admittedly and deliberately simple – reveal perception to be a generative, creative act; a proactive, context-laden interpretation of, and engagement with, sensory signals. And as I mentioned earlier, the principle that perceptual experience is built from brain-based predictions applies across the board – not only to vision and hearing, but to all of our perceptions, all of the time.

Locke proposed that the primary qualities of an object are those that exist independently of an observer, such as occupying space, having solidity, and moving. An oncoming train has these primary qualities in abundance, which is why jumping in front of one is a bad idea, whether or not you are observing it, and whatever beliefs you might hold about the nature of perception. Secondary qualities are those whose existence does depend on an observer. These are properties of objects that produce sensations – or ‘ideas’ – in the mind, and cannot be said to independently exist in the object. Colour is a good example of a secondary quality, since the experience of colour depends on the interaction of a particular kind of perceptual apparatus with an object.

The second objection has to do with our ability to perceive new things. One might worry that we’d need a pre-formed best guess for anything that we might ever perceive, so that we are forever trapped in a perceptual world of the already expected. Imagine that you’ve never seen a gorilla, not in real life, on TV, or in a film – nor even in a book – and then you unexpectedly encounter one ambling along the street. I guarantee that you will now see a gorilla, a new and probably rather scary perceptual experience. In a world of the already expected, how can this happen?

More generally, gorillas are objects that have defined (though furry) edges, that move in reasonably predictable ways, and that reflect light in the same way that other objects of similar size, colour, and texture do. The novel experience of ‘seeing a gorilla’ is built up from perceptual predictions operating over many different levels of granularity and acquired over many different timescales – from predictions about luminance and edges to predictions about faces and posture – that together sculpt a new overall perceptual best guess, so that you see a gorilla for the first time. The longer answer involves learning more about how the brain performs the wickedly complex neural gymnastics involved in perceptual inference – and this is precisely where we are headed in the next chapter.

When knowledge fails to affect perception, we call that perception ‘cognitively impenetrable’.

Like inductive reasoning, abductive reasoning can also get things wrong. In seeking the ‘best explanation’, abductive reasoning can be thought of as reasoning backward, from observed effects to their most likely causes, rather than forward, from causes to their effects – as is the case for deduction and induction.

Bayes’ rule is a mathematical recipe for going from what we already know (the prior) to what we should believe next (the posterior), based on what we are learning now (the likelihood

Perceptual priors can be encoded at many levels of abstraction and flexibility. These range from very general and relatively fixed priors such as ‘light comes from above’ to situation-specific priors like ‘the approaching furry object is a gorilla’. Likelihoods in the brain encode mappings from potential causes to sensory signals.

In other words, predictive processing is a theory about how brains work, whereas the controlled hallucination view takes this theory and develops it to account for the nature of conscious experiences. Importantly, both rest on the bedrock process of prediction error minimisation.

And it is prediction error minimisation that provides the connection between controlled hallucinations and Bayesian inference.

It’s worth clarifying here that ‘prediction’ in prediction error minimisation is not necessarily about the future. It simply means going beyond the data by using a model.

In statistics, the essence of prediction is in catering for the absence of sufficient data. Whether this is because predictions are about the future – one can think of the future as ‘insufficient data’ – or about some current but incompletely unknown state of affairs, doesn’t matter.

I say ‘estimated precision’ rather than simply ‘precision’ because the precision of sensory signals is not something that is directly given to the perceiving brain. It also has to be inferred. The brain is faced not only with the challenge of figuring out the most likely causes of its sensory inputs, but also with figuring out how reliable the relevant sensory inputs are

Remarkably, in some situations unattended sensory data may have no influence at all. In 1999, the psychologist Daniel Simons developed a well-known video demonstration of this phenomenon, which he calls ‘inattentional blindness’. If you haven’t seen it, I suggest taking a look before you read on.†

What’s happening is that focusing attention on the players in white means that the sensory signals from the players in black – and the gorilla – are afforded low estimated precision, and so have little or no influence on updating perceptual best guesses.

’s tempting to think of our interaction with the world in the following way. First, we perceive the world as it is. Then we decide what to do. Then we do it. Sense, think, act. This may be how things seem, but once again how things seem is a poor guide to how they actually are. It’s time to bring action into the picture. Action is inseparable from perception. Perception and action are so tightly coupled that they determine and define each other. Every action alters perception by changing the incoming sensory data, and every perception is the way it is in order to help guide action. There is simply no point to perception in the absence of action

It may even be that action comes first. Instead of picturing the brain as reaching perceptual best guesses in order to then guide behaviour, we can think of brains as fundamentally in the business of generating actions, and continually calibrating these actions using sensory signals, so as to best achieve the organism’s goals

Prediction errors can also be quenched by performing actions in order to change the sensory data, so that the new sensory data matches an existing prediction. Minimising prediction error through action is called active inference – a term coined by the British neuroscientist Karl Friston.

Like all aspects of predictive processing, active inference depends on generative modelling. More specifically, active inference relies on the ability of generative models to predict the sensory consequences of actions. These are predictions of the form ‘if I look over there, what sensory data am I likely to encounter?’ Such predictions are called conditional predictions – predictions about what would happen were something to be the case

Another example would be tidying away all the mugs on my desk in order to help find my car keys. In each case, choosing the relevant action relies on having a generative model able to predict how sensory data would change as a result of that action.

The most counterintuitive aspect of active inference is that action itself can be thought of as a form of self-fulfilling perceptual prediction.

In active inference, actions are self-fulfilling proprioceptive predictions. Proprioception is a form of perception which keeps track of where the body is and how it is moving, by registering sensory signals that flow from receptors situated all over the skeleton and musculature. We probably don’t think much about proprioception because, in some sense, it’s always there, but the simple fact that you can touch your nose with your eyes shut – try it! – demonstrates the essential role it plays in all our actions

This can be thought of as the opposite of paying attention – a kind of ‘disattention’ to the body, which allows it to move.

Thinking about action in this way underlines how action and perception are two sides of the same coin. Rather than perception being the input and action being the output with respect to some central ‘mind’, action and perception are both forms of brain-based prediction.

We now know that this brain is far from isolated. It swims in a torrent of sensory signals from the world and the body, continually directing actions – self-fulfilling proprioceptive predictions – which proactively sculpt this sensory flow.

The incoming sensory barrage is met by a cascade of top-down predictions, with prediction error signals streaming upward to stimulate ever better predictions and elicit new actions. This rolling process gives rise to an approximation to Bayesian inference, a GoodEnough Bayesianism in which the brain settles and resettles on its evolving best guess about the causes of its sensory environment, and a vivid perceptual world – a controlled hallucination – is brought into being.

One of the most influential ideas emerging from the age of insight is the ‘beholder’s share’, first introduced by Riegl and later popularised by one of the major figures in twentieth-century art history, Ernst Gombrich – himself born in Vienna in 1909. Their idea highlighted the role played by the observer – the beholder – in imaginatively ‘completing’ a work of art. The beholder’s share is that part of perceptual experience that is contributed by the perceiver and which is not to be found in the artwork – or the world – itself.

that is, all visual perception is based on classifying concepts and interpreting visual information. One cannot perceive that which one cannot classify.’

Yair used a set-up called ‘continuous flash suppression’ in which different images are presented to the left and right eyes. One eye is presented with a picture – in this case either a house or a face – while the other eye is presented with a rapidly changing pattern of overlapping oblongs. When the brain tries to fuse the two images into a single scene, it fails, and the changing oblongs tend to dominate, so that’s what the person consciously sees.

Conscious perception of the picture is suppressed by the ‘continuous flashing’ shapes.

To discover how perceptual expectations affected conscious perception, we cued participants with either the word ‘house’ or the word ‘face’ before each experimental trial. Importantly, these expectations were only partially valid. When participants were cued with the word ‘face’, a face would be present on 70 per cent of trials, but on the other 30 per cent a house would be present. The reverse was true when we led them to expect a house. By measuring how long it took for each image to emerge from flash suppression, we could determine how quickly people consciously saw a particular image – a house or a face – when it was expected, compared to when it was unexpected.

As we had predicted, people were faster and more accurate at seeing houses when houses were what they were expecting – and the same for faces. The difference in speed was small – about one tenth of a second – but it was reliable. In our experiment, valid perceptual expectations do indeed lead to more rapid and more accurate conscious perceptions

Single letters – like ‘U’ – are easier to identify when they form part of a word, such as ‘HOUSE’, than when they are part of a nonword letter string, such as ‘AEUVR’. De Lange’s team showed volunteers many examples of words and nonwords, always presented against visually noisy backgrounds. Confirming the word superiority effect, they found that individual letters were easier to read in the word condition than in the nonword condition.

The twist in this study came from a clever way of analysing the volunteers’ brain activity, which they had recorded using fMRI. By analysing the data using a powerful technique called ‘brain reading’, they found that the neural signature of a letter in the visual cortex was ‘sharper’ – by which I mean more distinguishable from neural representations of other letters – when the letter formed part of a word than when it was part of a nonword. This means that perceptual expectations provided by the word context are able to alter activity at early stages of visual processing in a way that enhances perception, just as the controlled hallucination view suggests should happen.

For anyone who doubts that the brain is the organ of experience, the hallucinations produced through LSD deliver a powerful corrective.

For humans – and some other animals – the significance of faces means that our brains come preloaded with strong face-related prior expectations.

What we call ‘hallucination’ is what happens when perceptual priors are unusually strong, overwhelming the sensory data so that the brain’s grip on their causes in the world starts to slide.

By programming the hallucination machine in slightly different ways, we can generate different kinds of simulated hallucinatory experience.

The hallucination machine is an exercise in what we might call ‘computational phenomenology’: the use of computational models to build explanatory bridges from mechanisms to properties of perceptual experience.

Objecthood is a property of how visual conscious contents generally appear, rather than being a property of any single conscious experience.

A large part of Pablo Picasso’s Cubist portfolio investigates how our perception of objecthood depends on our first-person perspective

Picasso’s work in particular draws the observer into imaginatively creating perceptual objects out of a jumble of possibilities. As the philosopher Maurice Merleau-Ponty put it, the painter investigates through painting the means by which an object makes itself visible to our eyes.

In cognitive science, the phenomenology of objecthood has been most thoroughly explored by ‘sensorimotor contingency theory’. According to this theory, what we experience depends on a ‘practical mastery’ of how actions change sensory inputs. When we perceive something, the content of what we perceive is not carried by the sensory signals; instead it emerges from the brain’s implicit knowledge about how actions and sensations are coupled. On this view, vision – and all our perceptual modalities – are things an organism does, not passive information feeds for a centralised ‘mind’.

According to sensorimotor contingency theory, I become perceptually aware of the back of the tomato, even though I cannot directly see it, because of implicit knowledge, wired into my brain, about how rotating a tomato will change incoming sensory signals.

Another situation where objecthood is typically absent is in ‘grapheme-colour synaesthesia’. The term synaesthesia refers to a kind of ‘mixing of the senses’.

People with the grapheme-colour variety have experiences of colour when seeing letters: for example, the letter A may elicit a luminous redness, regardless of its actual colour on the page. Although these colour experiences happen consistently and automatically – meaning that the same colour is experienced whenever a particular letter is encountered, and that no conscious effort is needed for this to happen – synaesthetes do not normally confuse their synaesthetic colours with real ‘out there in the world’ colours.

I think this is because synaesthetic colours, when compared to ‘real’ colours, do not support a rich repertoire of sensorimotor predictions. A synaesthetic ‘red’ does not vary much as you move around, or as the ambient lighting changes, and so there won’t be any phenomenology of objecthood

We predicted that normally behaving virtual objects would break through into consciousness sooner than those that violated sensorimotor predictions, and this is precisely what we found.

Many experiments have shown that physical change – change in the world – is neither necessary nor sufficient for the perception of change. The snake-filled image below provides a striking example in which nothing is moving, yet there can be a perceptual impression of movement, especially if you let your eyes rove around the image. What’s happening is that the fine details of the picture, when seen in the periphery of your vision – out of the corner of your eye – convince your visual cortex to infer motion even though no motion is present.

The opposite situation, physical change without perceptual change, happens in ‘change blindness’. This can occur when some aspects of an environment change very slowly, or when everything is changing at once with only some features being relevant. In one powerful video example of this phenomenon,§ the entire lower half of an image can change colour – from red to purple – but because it happens slowly, over about forty seconds, most people don’t notice the change at all, even if they are looking directly at the changing part of the image.

The difference here is that what people are failing to see is ‘change’ itself.

Some people think that change blindness exposes a philosophical dilemma: after the image has changed colour, are you still experiencing red (even though it’s now purple), or are you now experiencing purple, in which case what were you experiencing before, given that you didn’t experience any change? The resolution is to deny the premise of the question and to recognise that perception of change is not the same as change of perception. The experience of change is another perceptual inference, another variety of controlled hallucination. And if experiences of change are perceptual inferences, so too are experiences of time.

flows for us too, though sometimes it crawls along while at other times it streaks

We experience seconds, hours, months, and years, but we have no ‘time sensors’ inside our brains.

What’s more, setting aside the circadian rhythm, which provides us with jet lag among other things, there is no evidence for any ‘neuronal clock’ inside the head which measures out our experiences in time – and which would in any case be a prime example of what Daniel Dennett called a double transduction, in which a property of the world is re-instantiated in the brain for the benefit of an assumed internal observer. Instead, like change, like all our perceptions, experiences of time are controlled hallucinations too.

His idea is that we infer time based not on the ticking of an internal clock, but on the rate of change of perceptual contents in other modalities – and he devised a clever way to test it.

Led by Warrick, our team recorded a library of videos of different durations and in different contexts – a crowded city street, an empty office, a few cows grazing in a field near the university. We then asked volunteers to watch these videos and judge how long each lasted. When they did so, they all showed characteristic biases: underestimating the duration of long videos, and overestimating short videos. They also showed biases according to the context of each scene, rating busy scenes as lasting longer than quiet scenes, even for videos that were objectively of the same duration.

Warrick then showed the same videos to an artificial neural network that mimicked the operation of the human visual system. This network was in fact the same one we’d used in our hallucination machine. For each video, an estimate of duration was computed, based – roughly speaking – on the accumulated rate of change of activity within the network. These estimates did not involve any ‘inner clock’ whatsoever. Remarkably, the neural network estimates and the human estimates were virtually identical, showing the same biases by duration and by context.

This shows that time perception can emerge, at least in principle, from a ‘best guess’ about the rate of change of sensory signals, without any need for an internal pacemaker.

In a study led by postdoctoral researcher Maxine Sherman, we used fMRI to record people’s brain activity while they watched the same set of videos and estimated their durations. We wanted to know whether we could use activity in the visual cortex to predict how long each video seemed, as we’d been able to do in Warrick’s previous study using a computational model of vision. Maxine found that we could. Brain activity in the visual system, but not in other brain regions, neatly predicted subjective duration. This is strong evidence that experiences of duration indeed emerge from perceptual best guessing, and not from the ticking of any neuronal clock.

Other experiments that might have revealed an ‘inner clock’ have failed to do so. My favourite, by a distance, involved jumping off cranes. The neuroscientist David Eagleman set out to test the common intuition that subjective time slows down in moments of high drama, such as in the moments before a car crash. He reasoned that this subjective slowing down might be due to an internal clock running faster – more ticks of the clock in a given time period, longer perceived duration. This in turn should lead to a ‘speeding up’ of the rate of perception since a faster clock should mean an improved ability to perceive short durations. To test this idea, Eagleman and his team designed a special digital watch which displayed a series of numbers that flickered so quickly that they were impossible to read in normal conditions. Then he persuaded some brave volunteers to repeatedly perform scary adrenaline-loaded leaps into the void while staring at their flickering watches. If an internal clock was indeed speeding up, then – his reasoning went – the volunteers should see the blur resolve into readable numbers while in freefall. They couldn’t, so his study provided no evidence for an internal clock. Of course, absence of evidence is not evidence of absence, but still … what an experiment!

The idea is simple. As with the hallucination machine we pre-record some panoramic video footage, but this time the footage is of the interior of the exact same VR/AR laboratory in which we conduct our experiments. When volunteers arrive in the lab, they sit on a stool in the middle of the room and put on a head-mounted display, which has a camera attached to the front. They are invited to look around the room through the camera mounted on their headset. At a certain point, we switch the feed so that the camera displays not the live real-world scene but the pre-recorded video. Remarkably, most people in this situation continue to experience what they are seeing as ‘real’, even though it no longer is.

Long ago, back in the eighteenth century, David Hume made a similar observation about causality – another pervasive feature of the way we experience the world. Rather than physical causality being an objective property of the world, ready to be detected by our senses, Hume argued that we ‘project’ causality out into the world on the basis of repeated perception of things happening in close temporal succession.

To put it another way, even though perceptual properties depend on top-down generative models, we do not experience the models as models. Rather, we perceive with and through our generative models, and in doing so out of mere mechanism a structured world is brought forth.

Starting from the principle that the brain must infer the hidden causes of its sensory inputs, we’ve reached a new understanding of why and how our inner universe is populated with everything from coffee cups to colours to causality-ness – things that seem to be properties of an external objective reality, where this seeming-to-be is itself a property of perceptual inference.

A chair has a mind-independent existence; chairness does not.

Delirium – a word surfacing from the depths of the sixteenth century – comes from the Latin delirare, to deviate, to be deranged.

Unlike dementia, which is a chronic and degenerative condition, delirium is usually temporary.

Up to a third of elderly patients entering acute care develop hospital-induced delirium, and the proportion is even higher for those having surgery. Even though it usually recedes given time, there can be severe long-term consequences including reduced cognitive capacity, an increased chance of dying in the subsequent months, and heightened risk of subsequent episodes of delirium and dementia.

Delusions are rarely random. The specifics of my mother’s particular delirium have a twisted logic. She believes – she knows – that she is the victim of a devious experiment that everyone is ‘in on’ and in which I am complicit. Well, I do do experiments on people, and when I’m in the hospital I often take on a strangely fluid identity, one part son and another part doctor, trying to comfort her while at the same time going through medical notes and murmuring to the consultants and junior doctors about paraneoplastic encephalopathies and other horrid things. The brain is always trying to find closure, to make its best guess.

But what is a ‘self’? Is it the sort of thing that can be departed from and returned to? The self, too, is not what it might seem to be.

It may seem as though the self – your self – is the ‘thing’ that does the perceiving. But this is not how things are. The self is another perception, another controlled hallucination, though of a very special kind. From the sense of personal identity – like being a scientist, or a son – to experiences of having a body, and of simply ‘being’ a body, the many and varied elements of selfhood are Bayesian best guesses, designed by evolution to keep you alive

There are two philosophical problems raised by the teletransportation paradox. The consciousness-in-general problem is whether we can be sure that the replica will have conscious experiences, or whether it will be a perfectly functioning equivalent but without any inner universe. I don’t find this problem very interesting. If the replica is created in sufficient detail – every molecule identical! – then there’s no reason to doubt it would be conscious, and conscious in exactly the same way as the original. If the replica is not completely identical, then we’re back to arguments about different kinds of philosophical zombie – and there’s no need to go over all that again. The more interesting problem is that of personal identity. Is the Eva on Mars (let’s call her Eva2) the same person as Eva1 (the Eva still in London)? It’s tempting to say, yes, she is: Eva2 would feel in every way as Eva1 would have felt had she actually been transported instantaneously from London to Mars. What seems to matter for this kind of personal identity is psychological continuity, not physical continuity.* But then if Eva1 has not been vaporised, which is the real Eva?

I think the correct – but admittedly strange – answer is that both are the real Eva.

We intuitively treat experiences of self differently from experiences of the world. When it comes to the experience of being you it seems harder to resist the intuition that it reveals a genuine property of the way things are – in this case an actual self – rather than a collection of perceptions. One intuitive consequence of assuming the existence of an actual self is that there can be only one such self, not two, or two thirds, or many.

Kant, in his Critique of Pure Reason, argued that the concept of the self as a ‘simple substance’ is wrong, and Hume talked about the self as a ‘bundle’ of perceptions. Much more recently, the German philosopher Thomas Metzinger wrote a very brilliant book called Being No One – a powerful deconstruction of the singular self. Buddhists have long argued that there is no such thing as a permanent self and through meditation have attempted to reach entirely selfless states of consciousness. Ayahuasca ceremonies in South America, and increasingly elsewhere, strip away people’s sense of self with heady mixtures of ritual and dimethyltryptamine.

Most curious of all are craniopagus twins, who are not only physically conjoined but also share some of their brain structures. What could it mean to be an individual self, when it turns out that one craniopagus twin can feel the other drinking orange juice? Being you is not as simple as it sounds.

Even if Eva1 had been standing right next to Eva2 there would have been small differences in sensory inputs leading to subtle differences in behaviour and before you know it, Eva1 and Eva2 are experiencing different things, laying down different memories, becoming different people.

The relationship between Eva1 and Eva2 might even be a little bit like the relationship that holds between you now and you ten years ago – or ten years in the future.

There are experiences of embodied selfhood that relate directly to the body. These include feelings of identification with the particular object that happens to be your body – we feel a certain sense of ownership over our body that doesn’t apply to other objects in the world

Emotions and moods are also aspects of embodied selfhood, as are states of arousal and alertness. And running below these experiences we can find deeper, formless feelings of simply being an embodied living organism – of being a body – without any clearly definable spatial extent or specific content. We’ll come back to this bedrock layer of selfhood later. For now it’s enough to think of it as the ‘feeling of being alive’.

Moving on from the body, there’s the experience of perceiving the world from a particular point of view, of having a first-person perspective – a subjective point of origin for perceptual experience which usually appears to reside inside the head, located somewhere between the eyes and slightly behind the forehead. This perspectival self is nowhere better illustrated than in the Austrian physicist Ernst Mach’s self-portrait, also known as ‘View from the Left Eye’.

Experiences of volition, of intending to do things – intention – and of being the cause of things that happen – agency – are also central to selfhood. This is the volitional self. When people talk about ‘free will’ it is these aspects of selfhood that they are usually talking about.

All these ways of being-a-self can be in place prior to any concept of personal identity – the identity that can be associated with a name, a history, and a future. As we saw with the teletransportation paradox, for personal identity to exist, there has to be a personalised prior history, a thread of autobiographical memories, a remembered past and a projected future.

This sense of personal identity, when it emerges, can be called the narrative self. With its appearance comes the ability to experience sophisticated emotions like regret, as opposed to mere disappointment. (We humans can also suffer ‘anticipatory regret’ – the feeling of certainty that what I’m about to do will turn out badly, that despite knowing this I will do it anyway, and that I and others will suffer as a consequence.) Here we can see how different levels of selfhood ramify and interact – the emergence of personal identity both changes and is partly defined by the increased range of emotional states on offer. The social self is all about how I perceive others perceiving me. It is the part of me that arises from my being embedded in a social network.

For each of us – in normal circumstances – these diverse elements of selfhood are bound together, all of a piece, all subsumed within an overarching unified experience – the experience of being you.

Just as experiences of redness are not indications of an externally existing ‘red’, experiences of unified selfhood do not signify the existence of an ‘actual self’. Indeed, the experience of being a unified self can come undone all too easily.

The sense of personal identity, built on the narrative self, can erode or disappear entirely in dementia and in severe cases of amnesia, and it can be warped and distorted in cases of delirium, whether hospital-induced or not. The volitional self can go awry in conditions like schizophrenia and alien hand syndrome, when people experience a reduced sense of connection with their own actions, or in akinetic mutism, a disorder in which people stop interacting with their surroundings altogether. Out-of-body experiences and other dissociative disorders affect the perspectival self, while disorders of body ownership range from phantom limb syndrome – the experience of persistent, often painful sensations located in a limb that is no longer there – to somatoparaphrenia – the experience that one of your limbs belongs to someone else.

The self is not an immutable entity that lurks behind the windows of the eyes, looking out into the world and controlling the body as a pilot controls a plane. The experience of being me, or of being you, is a perception itself – or better, a collection of perceptions – a tightly woven bundle of neurally encoded predictions geared towards keeping your body alive.

The best-known example is the ‘rubber hand illusion’, first described more than twenty years ago, and now a cornerstone of research on embodiment.

Is nirvana being rubber hand syndrome to the entire world? If yes, how does it happen with meditation?

The idea is that when the hands are stroked synchronously, she will develop the uncanny feeling that the rubber hand is somehow actually part of her body, even though she knows that it isn’t. But when the hands are stroked asynchronously, out of time, the illusion should not develop and she will not assimilate the rubber hand into her experience of what is her body.

The idea is that in the synchronous-stroking condition, the combination of seeing the rubber hand be touched and simultaneously feeling (but not seeing) the touch on the real hand provides enough sensory evidence for the brain to reach a perceptual best guess that the rubber hand is somehow part of the body. This happens in the synchronous – but not the asynchronous – condition because of a prior expectation that simultaneously arriving sensory signals are likely to have a common source – the rubber hand.

experiments were based on the rubber hand illusion, but now extended to the entire body. In one of the studies, conducted by a group in Lausanne led by Olaf Blanke, volunteers wore a head-mounted display through which they saw a virtual reality representation of the back of their own body from a distance of about two metres (see opposite). From this perspective, they then saw the virtual body being stroked with a paintbrush, again either synchronously or asynchronously with strokes applied to their own (real) body. When the stroking was synchronous, most participants reported that they felt the virtual body was, to some extent, their ‘own’ body, and when asked to walk to where they felt their body was, they showed a drift in space towards the location of the virtual body

But there’s no need to reach for such dualistic skyhooks if you take the first-person perspective to be another species of perceptual inference. This view is supported not only by the virtual reality experiments of Olaf Blanke and others, but also by brain stimulation studies going all the way back to a series of seminal experiments conducted in the 1940s by the Canadian neurologist Wilder Penfield.

Among Penfield’s patients was a woman known as G.A. who, when electrically stimulated in her right superior temporal gyrus – part of the brain’s temporal lobe – spontaneously exclaimed, ‘I have a queer sensation that I am not here … As though I were half here and half not here.’ Blanke himself first became fascinated by OBEs when a patient of his, stimulated in a similar part of the brain – the angular gyrus, at the junction of the temporal and parietal lobes – reported a similar experience: ‘I see myself lying in bed, from above, but I only see my legs.’

seems that when normal activity in these systems becomes disrupted, the brain can reach an unusual ‘best guess’ about the location of its first-person perspective, even while other aspects of selfhood are left unaltered

In the virtual world, the ability to alter first-person perspectives is generating some fascinating applications, many of them driven by the intriguingly named ‘body swap’ illusion, which was described in a 2008 study led by the Swedish researcher Henrik Ehrsson. In the body swap set-up, two people wear head-mounted displays, each with a camera attached. By swapping the camera feeds between the headsets, each person can see themselves from the other’s point of view. The effect kicks in properly only when they shake hands. The idea is that seeing and simultaneously feeling the handshake provides the multisensory stimulation so that, when combined with top-down expectations, each person feels they are now somehow located in the other person’s body, shaking hands with themselves. This experience puts you, albeit virtually, in the shoes of another.

Supporting this hypothesis, we found that individual differences in the strength of the illusion correlated with how suggestible a person is, when measured on a standard scale of hypnotisability. People who were highly hypnotisable reported strong feelings of ownership (for synchronous stroking), while those who scored low on the scale were hardly susceptible at all.

These higher reaches of selfhood are fully dissociable from the embodied self. Many non-human animals, as well as human infants, may experience embodied selfhood without having – or missing – any accompanying sense of personal identity. And while adult humans normally experience all these forms of selfhood in an integrated and unified way, when the narrative and social aspects of self are diminished or destroyed, the impact can be devastating.

suffered a devastating brain infection, a herpes encephalitis that wrought massive damage to his hippocampus in both cerebral hemispheres, producing in him one of the most profound amnesias ever documented.‡ Wearing has immense problems recalling old memories (retrograde amnesia) and, especially, laying down new memories (anterograde amnesia). Remarkably, he seems to exist in a permanent present of between seven and thirty seconds. He’s now in his eighties, and it’s likely that he still experiences his life as a continuous series of mini-awakenings, as if – every twenty seconds or so – he has just emerged from a coma, or from anaesthesia. His narrative self has been annihilated

His inability to string together a self-narrative over time means that what-it-is-to-be-him has been, for more than thirty years, a continual starting from scratch, a fleeting presence with no stable ‘I’ around which to organise the flow of perceptions of world and self

His love for his wife remains undiminished even though sometimes he cannot remember meeting her – they married just a year before his illness. And when Clive plays the piano, or sings, or conducts, the music falls freely from him with a fluency that makes him seem whole again.

Humans, like many other animals, are social creatures. Perceiving the state of mind of another is a crucial ability for social creatures in all sorts of contexts and in all manner of societies. This ability – sometimes called ‘theory of mind’ – is often thought to develop rather slowly in humans, but it comes to play a key role for almost all of us throughout our lives.

Much of our social perception is automatic and direct. We form perceptions of others’ beliefs, emotions, and intentions as naturally and effortlessly as we form perceptions of cats and coffee cups and chairs and even of our own bodies.

One reason for the inherent ambiguity of social perception is that the relevant causes are more deeply hidden. The light waves that elicit perception of a wine glass originate more or less directly from the glass itself, but the sensory signals relevant to others’ mental states must pass through a number of intermediate stages – through facial expressions, gestures, and speech acts – with each stage creating a new opportunity for an off-target inference.

For example, we smile not only to express our own pleasure but also to change the way our companion is feeling, and when we speak we are trying to insert thoughts into another’s mind.

My best model of your mental states will include a model of how you model my mental states.

striking but often overlooked aspect of conscious selfhood is that we generally experience ourselves as being continuous and unified across time. We can call this the subjective stability of the self. It holds not only in terms of a continuity of autobiographical memory, but in a deeper sense of experiencing oneself as persisting from moment to moment, whether at the level of the biological body or at the level of personal identity

Self-related experiences seem to change much less. Even though we know we change over time – most of us have more than enough photographic evidence for that – it still seems to us that we don’t change all that much.

This distinction is exemplified by the phenomenon of ‘change blindness’, in which slowly changing things (in the world) do not evoke any corresponding experience of change. The same principle will apply to self-perceptions too.

This subjective blindness to the changing self has consequences. For one thing, it fosters the false intuition that the self is an immutable entity, rather than a bundle of perceptions.

We do not perceive ourselves in order to know ourselves, we perceive ourselves in order to control ourselves.

Self-perception is not about discovering what’s out there in the world, or in here, in the body. It’s about physiological control and regulation – it’s about staying alive.

For centuries, especially in Europe, the Great Chain of Being – or Scala Naturae – provided a stable template by which humans could understand their place in nature, as well as their value compared to other humans – kings being higher up the Chain than peasants, for instance. Then, in the seventeenth century, René Descartes did away with the many gradations of the Scala by cleaving the universe into just two modes of existence: res cogitans (mind stuff) and res extensa (matter stuff).

If animals had elements of res cogitans, any signs of minds, what was to prevent them from aspiring to the spiritual realm, as humans could? And any attempt to intellectually investigate the soul – as Descartes seemingly advocated – was sure to annoy the powerful Catholic Church.

When it came to non-human animals, it is often claimed that he thought they lacked consciousness entirely. Although it is hard to be sure, this was probably not his view.* Descartes’ primary claim about non-human animals was that they lacked souls and all the rational, spiritual, and conscious attributes that came along with having a soul.

By reinforcing the specialness of humans Descartes was able to placate many of his would-be persecutors. But a dangerous door was now ajar. If non-human animals are beast machines, and if humans are a kind of animal too – humans, after all, certainly seemed to be made of the same sort of flesh, blood, cartilage, and bone – then surely the faculties of mind and reason should also be explicable in mechanistic, physiological terms?

Julien Offray de La Mettrie, writing in the middle of the eighteenth century, certainly saw things this way

He extended Descartes’ beast machine argument to human beings, arguing that humans were machines too – l’homme machine (man machine) – and in doing so denying any special immaterial status for the soul while also questioning the existence of God.

While in the Cartesian view mind and life are independent, for La Mettrie they were deeply connected in the sense that mind could be viewed as a property of life

My sympathies in this debate lie with La Mettrie, but rather than speaking in general terms about ‘mind’, my focus is on consciousness. This brings us to the heart of my ‘beast machine’ theory of consciousness and self. Our conscious experiences of the world around us, and of ourselves within it, happen with, through, and because of our living bodies. Our animal constitution is not merely compatible with our conscious perceptions of self and world. My proposal is that we cannot understand the nature and origin of these conscious experiences, except in light of our nature as living creatures.

Underneath the layered expressions of selfhood involving memories of the past and plans for the future, before the explicit sense of personal identity, beneath the ‘I’ and even prior to the emergence of a first-person perspective and experiences of body ownership, there are deeper layers of selfhood still to be found.

These bedrock layers are intimately tied to the interior of the body, rather than to the body as an object in the world, and they range from emotions and moods – what psychologists call ‘affective’ experiences – to a basal, formless, and ever-present sense of simply ‘being’ an embodied, living organism.

These world-oriented varieties of sensation and perception are collectively called exteroception.

Perception of the body from within is known as interoception –

Interoceptive sensory signals are typically transmitted from the body’s internal organs – the viscera – to the central nervous system, conveying information about the state of those organs, as well as about the functioning of the body as a whole.

These signals travel through a complex network of nerves and deep-lying brain regions in the brainstem and thalamus before arriving at parts of the cortex specialised for interoceptive processing – in particular the insular cortex.‡

In other words, how good a job the brain is doing of keeping its body alive.

Interoceptive signalling has long been connected with emotion and mood. Back in 1884, William James and – independently – Carl Lange argued that emotions were not the ‘eternal and sacred psychic entities’ of the ancient philosophers, nor were they hard-wired into brain circuits by evolution, as Darwin had proposed not long before.

Instead, they argued that emotions are perceptions of changes in bodily state. We don’t cry because we are sad, we are sad because we perceive our bodily state in the condition of crying. The emotion of fear, in this view, is constituted by (interoceptive) perception of a whole gamut of bodily responses set off by the organism recognising danger in its environment. For James, the perception of bodily changes as they occur is the emotion: ‘We feel sorry because we cry, angry because we strike, afraid because we tremble, and not that we cry, strike, or tremble, because we are sorry, angry, or fearful.’

Appraisal theories solve the problem of emotional range because each specific emotion now no longer needs a dedicated bodily state.

Two closely related emotions – for example, listlessness and ennui – might be based on the same bodily state.

My favourite experimental investigation of appraisal theory comes from an inventive study by Donald Dutton and Arthur Aron in 1974, in which a female interviewer approached male passers-by while they were crossing one of two bridges across the Capilano River in North Vancouver.

The researchers wondered whether the men on the rickety bridge would misinterpret the physiological arousal caused by their precarious condition as sexual attraction, rather than as fear or anxiety.

The female interviewer received more calls from men who had been crossing the rickety bridge than from those who had been crossing the sturdy bridge.

Dutton and Aron called this a ‘misattribution of arousal’: increased physiological arousal, induced by the rickety bridge, had been misinterpreted by higher-level cognitive systems as sexual chemistry.

Supporting this appraisal theory interpretation (and assuming heterosexuality), when the questionnaire-bearer was a man, there was no effect of type of bridge on the number of follow-up calls made.§

But it still vividly illustrates the view that emotional experiences depend on how physiological changes are evaluated by higher-level cognitive processes.

I called the idea interoceptive inference

In the same way that ‘redness’ is the subjective aspect of brain-based predictions about how some surfaces reflect light, emotions and moods are the subjective aspects of predictions about the causes of interoceptive signals.

However, unfortunately for appraisal theories, the brain does not separate neatly into ‘cognitive’ and ‘non-cognitive’ domains.

In this way, interoceptive inference solves the problem of emotional range without needing any bright-line distinction between the non-cognitive and the cognitive.

Interoceptive inference is therefore more parsimonious than appraisal theory, because it involves just one process (Bayesian best guessing) rather than two (non-cognitive perception and cognitive evaluation), and because of this it also maps more comfortably onto the underlying brain anatomy.

Measuring interoceptive inference

One promising approach explores the possibility that brain responses to heartbeats might be signatures of interoceptive prediction errors. The German neuroscientist Frederike Petzschner has recently shown that such responses, called ‘heartbeat evoked potentials’, are modulated by paying attention, as predicted by interoceptive inference.

Keisuke Suzuki we found that people experienced greater ownership over a virtual reality ‘rubber hand’ when it flashed in time with their heartbeat than when it flashed out of time – suggesting that body ownership depends on the integration of both exteroceptive and interoceptive signals

more research is needed here too, in part because these experiments didn’t take account of individual differences in hypnotic suggestibility –

also depend on how aware a person is of their own heartbeat, a trait which has proven frustratingly difficult to measure

From the perspective of the beast machine theory, the most important implication of interoceptive inference is that affective experiences are not merely shaped by interoceptive predictions but constituted by them. Emotions and moods, like all perceptions, come from the inside out, not the outside in. Whether it’s fear, anxiety, joy, or regret – every emotional experience is rooted in top-down perceptual best guessing about the state of the body (and about the causes of this state).

To take the next step, we need to ask what these perceptions of the body ‘from within’ are for. Perception of the outside world is obviously useful for guiding action, but why should our internal physiological condition be built into our conscious lives from the ground up?

This way of thinking about machines – as potentially having ‘purposes’ – provided a new bridge from the non-living to the living.

cybernetics

One of these insights comes from a 1970 paper by William Ross Ashby and Roger Conant, which describes their so-called ‘Good Regulator Theorem’. The concept is nicely encapsulated by the title of their paper: ‘Every good regulator of a system must be a model of that system’.

System B is better because it has a model of the house, which allows it to predict how the temperature inside the house will respond to the actions it can take. A top-end System B might even be able to anticipate upcoming temperature-related challenges – perhaps a cold day on the way – and alter the boiler output in advance, so as to guard against even a temporary drop in warmth. As Conant and Ashby said, every good regulator of a system must be a model of that system.

Let’s take this example a step further. Imagine that System B has been fitted with imperfect ‘noisy’ temperature sensors that only indirectly reflect the ambient temperature in the house.

System B works by inferring the ambient temperature from sensor readings, just as our brain makes best guesses about the causes of its sensory signals in order to infer states of the world (and body) and how they change over time.

But the goal for System B is not to figure out ‘what’s there’ – in this case the ambient temperature. The goal is to regulate this inferred hidden cause, to take action so as to keep the temperature within a comfortable range, and ideally at a single fixed value. Perception, in this context, is not for figuring out what’s there, it’s for control and regulation.

Control-oriented perception – the sort of thing implemented by System B – is therefore a form of active inference,

What cybernetics brings to the table is the idea that, for some systems, control comes first. From the perspective of the Good Regulator Theorem, the entire apparatus of predictive perception and active inference emerges from a fundamental requirement about what it takes to adequately regulate a system.

To answer the question of what perceptions of emotion and mood are for, we need one more concept from cybernetics – that of an essential variable.

essential variables are physiological quantities, such as body temperature, sugar levels, oxygen levels and the like, that must be kept within certain rather strict limits in order for an organism to remain alive. By analogy, a desired room temperature would be the ‘essential variable’ for a central heating system.

Putting these pieces together, emotions and moods can now be understood as control-oriented perceptions which regulate the body’s essential variables.

This perspective on emotions and moods ties them even more closely to our flesh-and-blood nature.

Experiences of fear, jealousy, joy, and pride are very different, but they are more similar to each other than any one of them is to a visual experience, or to an auditory experience. Why is this

Predictions geared towards finding out things will have a very different phenomenology from those geared towards controlling things.

Now consider a more active example: catching a cricket ball. You might think the best way to do this would be to figure out where the ball is going to land and run there as fast as possible. But ‘figuring out what’s there’ is not, in fact, a good strategy, and is not what expert fielders do. Instead, you should keep moving so that the ball always ‘looks the same’ in a particular way – specifically, so that the angle of elevation of your gaze to the ball increases, but at a steadily decreasing rate. It turns out that if you follow this strategy – psychologists call it ‘optic acceleration cancellation’ – you are guaranteed to intercept the ball.††

Your actions – and your brain’s predictions about their sensory consequences – are not about finding out where the ball is. They are geared towards controlling how the ball perceptually appears.

Perception in this situation is a controlling hallucination just as much as it is a controlled hallucination.

This idea has substantial historical pedigree. In the 1970s, the psychologist James Gibson argued that we often perceive the world in terms of what he called ‘affordances’.

An affordance, for Gibson, is an opportunity for action – a door for opening, a ball for catching – rather than an action-independent representation of the ‘way things are’.

According to William Powers’ ‘perceptual control theory’, we don’t perceive things in order to then behave in a particular way. Instead, as in the example of catching a cricket ball, we behave so that we end up perceiving things in a particular way.

Anxiety doesn’t have a back, sadness doesn’t have sides, and happiness is not rectangular. The perceptions of the body ‘from within’ on which affective experiences are built do not deliver experiences of the shape and location of my various internal organs – my spleen here, my kidneys over there. There is no phenomenology of objecthood, as when looking at a coffee cup on the table, nor is there anything like movement in a spatial frame, as when catching a cricket ball.

The control-oriented perceptions that underpin emotions and moods are all about predicting the consequences of actions for keeping the body’s essential variables where they belong. This is why, instead of experiencing emotions as objects, we experience how well or badly our overall situation is going, and is likely to go.

At the very deepest layers of the self, beneath even emotions and moods, there lies a cognitively subterranean, inchoate, difficult-to-describe experience of simply being a living organism

Here, experiences of selfhood emerge in the unstructured feeling of just ‘being’. This is where we reach the core of the beast machine theory: the proposal that conscious experiences of the world around us, and of ourselves within it, happen with, through, and because of our living bodies. It is at this point that all of the ideas I’ve been putting forward about perception and self fall into place. So let’s take things step by step, from the beginning.

The primary goal for any organism is to continue staying alive. This is true almost by definition – an imperative endowed by evolution.

All living organisms strive to maintain their physiological integrity in the face of danger and opportunity. This is why brains exist.

Evolution’s reason for providing organisms with brains is not so they can write poetry, do crossword puzzles, or pursue neuroscience. Evolutionarily speaking, brains are not ‘for’ rational thinking, linguistic communication, or even for perceiving the world. The most fundamental reason any organism has a brain – or any kind of nervous system – is to help it stay alive, through making sure that its physiological essential variables remain within the tight ranges compatible with its continued survival.

These essential variables, whose effective regulation determines the life-status and future prospects of an organism, are the causes of interoceptive signals. Like all physical properties, these causes remain hidden behind a sensory veil. Just as with the outside world, the brain has no direct access to physiological states of the body, and so these states have to be inferred through Bayesian best guessing.

As with all predictive perception, this best guessing is achieved through a brain-based process of prediction error minimisation. In the context of interoception, this is called interoceptive inference.

Just as with vision and with hearing – just as with all perceptual modalities – interoceptive perception is a kind of controlled hallucination.

Whereas perceptual inference about the world is often geared towards finding out things, interoceptive inference is primarily about controlling things – it is about physiological regulation.

Interoceptive inference exemplifies active inference, in that prediction errors are minimised by acting to fulfil top-down predictions, rather than by updating the predictions themselves (though this happens too).

This kind of predictive control can support anticipatory responses, through predictions about future bodily states and their dependence on this-or-that action.

Running away from a bear before being eaten is an example of external anticipatory regulation. The transient increase in blood pressure needed to run away effectively, or even to stand up from your desk after you’ve been working for a while, is an internal anticipatory response.

There is a useful term in physiology to describe this process: allostasis. Allostasis means the process of achieving stability through change, as compared to the more familiar term homeostasis, which simply means a tendency towards a state of equilibrium.

We can think of interoceptive inference as being about the allostatic regulation of the physiological condition of the body.

These affective experiences have their characteristic phenomenology because of the control-oriented and body-related nature of the perceptual predictions that they depend on. They are controlling hallucinations just as much as they are controlled hallucinations.

Despite being firmly rooted in physiological regulation, emotions and moods are still mostly experienced at least in part as relating to things and situations beyond the self, outside the body. When I feel fear, I am usually afraid of some thing. But the very deepest levels of experienced selfhood – the inchoate feeling of ‘just being’ – seem to lack these external referents altogether. This, for me, is the true ground-state of conscious selfhood: a formless, shapeless, control-oriented perceptual prediction about the present and future physiological condition of the body itself. This is where being you begins, and it is here that we find the most profound connections between life and mind, between our beast machine nature and our conscious self.

We are not the beast machines of Descartes, for whom life was irrelevant to mind. It is exactly the opposite. All of our perceptions and experiences, whether of the self or of the world, all are inside-out controlled and controlling hallucinations that are rooted in the flesh-and-blood predictive machinery that evolved, develops, and operates from moment to moment always in light of a fundamental biological drive to stay alive.

At the end of the previous chapter I noted that while perceptions of the world come and go, experiences of selfhood seem to be stable and continuous over many different timescales. We can now see that this subjective stability emerges naturally from the beast machine theory

To effectively regulate the body’s physiological condition, priors on interoceptive signals need to have high precision, so that they tend to become self-fulfilling

For example, my body temperature is predicted to be rather constant over time, which – according to active inference – is why it actually turns out this way. The experience of the bodily self as being relatively unchanging therefore stems directly from the need to have precise priors – strong predictions – about stable bodily states, for the purposes of physiological regulation.

To put it another way: for as long as we live, the brain will never update its prior belief of expecting to be alive.‡‡

What’s more, given that ‘change’ is itself an aspect of perceptual inference, the brain may attenuate prior expectations related to perceiving changes in the condition of the body, in order to further ensure that physiological essential variables stay where they ought to be.

This implies a form of ‘self-change-blindness’ – a notion also introduced in the previous chapter.

On this view, we might not perceive our physiological condition as changing even when it does, in fact, change.

Awesome how he is connectig points

Putting these ideas together, we perceive ourselves as stable over time in part because of a self-fulfilling prior expectation that our physiological condition is restricted to a particular range, and in part because of a self-fulfilling prior expectation that this condition does not change.

In other words, effective physiological regulation may depend on systematically misperceiving the body’s internal state as being more stable than it really is, and as changing less than it really does.

We will be better able to maintain our physiological and psychological identity, at every level of selfhood, if we do not (expect to) perceive ourselves as continually changing.

This may seem obvious, but remember from chapter 6 that the experience of things in the world as ‘really existing’ is not evidence of direct perceptual access to an objective reality, but a phenomenological property that needs to be explained.

There, I proposed that to be useful for the perceiving organism, our perceptual best guesses need to be experienced as really existing out there in the world, rather than as the brain-based constructions that in truth they are. The same reasoning holds for the self too. Just as it seems as though the chair in the corner really is red, and that a minute really has passed since I started writing this sentence, the predictive machinery of perception when directed inwardly makes it seem as though there really is a stable essence of ‘me’ at the centre of everything.

And in the same way that our perceptions of the world can sometimes lack the phenomenology of being real, the self too can lose its reality. The experienced reality (and subjective stability) of the self may wax and wane during illness, and it can be severely attenuated or even abolished in the psychiatric condition of depersonalisation.

The embodied self is so far gone in the Cotard delusion that sufferers believe they do not exist, or that they are already dead. Of course, the experience that the self is unreal does not mean that any essence-of-self has upped and left. It just means that the control-oriented perceptions associated with the deepest layers of bodily regulation have gone significantly awry.

What I am claiming is that in order to understand why our conscious experiences are the way they are, what experiences of self are like, and how they relate to experiences of the world, we will do well to appreciate the deep roots of all perception in the physiology of the living.

By extending the controlled hallucination view to the very deepest layers of selfhood, by revealing the experience of the-self-as-really-existing as one more aspect of perceptual inference, the intuitions on which the hard problem implicitly rest are eroded even further.

What we might call the ‘soul’ in this view is the perceptual expression of a deep continuity between mind and life.

It seems right to call this an echo of the soul because it revives even more ancient conceptions of this eternal notion, conceptions – such as the Ātman in Hinduism – which contemplated our innermost essence more as breath than as thought. We are not cognitive computers, we are feeling machines.

One might wonder whether there is a difference between ‘being’ a model and ‘having’ a model. I think that systems that possess explicit generative models capable of producing conditional or counterfactual predictions, like System B, can be said to ‘have’ models. Regulators that are relatively fixed and inflexible, like a simple feedback thermostat such as System A, may merely ‘be’ a model.

What it means for something to exist is that there must be a difference – a boundary – between that thing and everything else. If there were no boundaries there would be no things – there would be nothing.

This boundary must also persist over time, because things that exist maintain their identity over time. If you add a drop of ink to a glass of water it will rapidly disperse, colouring the water and losing its identity. If instead you add a drop of oil, although the oil will spread out over the surface, it will remain recognisably separate from the water. The oil drop continues to exist because it has not dispersed itself evenly throughout the water.

But neither oil drops nor rocks actively maintain their boundaries, they just get dispersed slowly enough for us to notice them as existing while this happens.

Living systems are different. Unlike the examples above, living systems actively maintain their boundaries over time – through moving, or sometimes even just through growing.

The starting point for the FEP is that living systems, simply by virtue of existing, must actively resist the dispersion of their internal states.

For you, or me, or even a bacterium, our internal states are less disordered when we are alive than when we decompose into mush. Being alive means being in a condition of low entropy.

The second law tells us that instances of organised matter, like living systems, are intrinsically improbable and unstable, and that – in the long run – we’re all doomed. But somehow, unlike rocks or ink drops, living systems temporarily fend off the second law, persisting in a precarious condition of improbability.

According to the FEP, for a living system to resist the pull of the second law it must occupy states which it expects to be in.

It is by taking advantage of this openness that living systems are able to engage in the energy-thirsty activity of seeking out statistically expected states, minimising entropy, and warding off the second law.

From the perspective of an organism, the entropy that matters is the entropy of its sensory states – those states that bring it into contact with its environment. Imagine a very simple living system, like a single bacterium. This bacterium requires a particular nutrient to survive, and it can sense the concentration of this nutrient in its immediate environment. By expecting to sense high nutrient concentrations, and by actively seeking out such expected sensory signals through its movements, this simple organism will maintain itself in the set of states that define it as being alive.

In other words, sensing high nutrient concentrations is a statistically expected state for the bacterium, which it proactively seeks to keep visiting.

Normally, to minimise a quantity, a system has to be able to measure it. The problem here is that sensory entropy cannot be directly detected or measured.

sensory entropy is very different from things like levels of light, or concentrations of nearby nutrients, which can be directly detected by an organism through its senses, and used to guide behaviour

free energy as a quantity which approximates sensory entropy. Crucially, it is also a quantity that can be measured by an organism – and therefore it is something that the organism can minimise

Following the FEP, we can now say that organisms maintain themselves in the low-entropy states that ensure their continued existence by actively minimising this measurable quantity called free energy. But what is free energy from the perspective of the organism?

It turns out, after some mathematical juggling, that free energy is basically the same thing as sensory prediction error

One implication of this connection is that the FEP licenses the idea from the previous chapter that living systems have – or are – models of their environment.

(More specifically, models of the causes of their sensory signals.)

This is because in predictive processing, as we saw in chapter 5, models are needed to supply the predictions that in turn define prediction errors.

According to the FEP, it is in virtue of having or being a model that a system can judge whether its sensations are (statistically) surprising. (If you believe that the number 6 you see comes from the roll of a die, you can judge exactly how surprising it is.)

Seen this way, the ideas of predictive perception and controlled (or controlling) hallucination follow seamlessly from Friston’s ambitious attempt to explain the whole of biology.

Paraphrasing Friston, the view from the FEP is of organisms gathering and modelling sensory information so as to maximise the sensory evidence for their own existence. Or, as I like to say, ‘I predict myself therefore I am.’

s worth noting that minimising free energy – sensory prediction error – does not mean that a living system can get away with retreating into a dark and silent room and staying there, staring at the wall. You might think this would be an ideal strategy, since sensory inputs from the external environment will become highly predictable. But it is far from ideal. Over time, sensory inputs signalling other things, like levels of blood sugar and so on, will start to deviate from their expected values: you’re going to get hungry if you stay in the dark room too long.

Sensory entropy will start to grow, and non-existence will loom.

Complex systems like living organisms need to allow some things to change in order for other things to stay the same.

Another common worry about the FEP is that it is not falsifiable, in the sense that it cannot be proven wrong by experimental data.

The theory of predictive processing, for example, can be falsified if it turns out that the brain doesn’t use sensory prediction errors in the process of perceiving.

In the end, the FEP will be judged on how useful it is, not on whether it is empirically true or false itself.

Let’s summarise the main steps of the FEP.

In order for organisms to stay alive they need to behave so as to maintain themselves in the (low entropy) states they ‘expect’ to be in. A fish swimming above a coral reef, searching for food, is proactively seeking expected sensory states compatible with its continued survival. In general, living systems do this by minimising a measurable approximation to the entropy of these states, which is free energy. Minimising free energy requires the organism to have, or to be, a model of its environment (which includes the body). Free-energy-minimising organisms then use these models to reduce the difference between predicted and actual sensory signals, by updating predictions and by performing actions. Indeed, given plausible mathematical assumptions, free energy turns out to be exactly the same thing as prediction error. Altogether, this means that the entirety of predictive processing and controlled hallucinations, of active inference and control-oriented perception – and of the beast machine theory too – can be understood through the lens of the FEP as flowing from a fundamental constraint on what it means to be alive, on what it means to exist.

the FEP is worth the journey because it enhances the beast machine theory in at least three significant ways.

First, the FEP grounds the beast machine theory in the bedrock of physics, and in particular within a physics relevant to what it means to be alive. The beast machine’s ‘drive to stay alive’ resurfaces in the FEP as an even more fundamental imperative to remain in statistically expected states, to withstand the insistent pull of the second law of thermodynamics

retelling it in reverse. Over the previous chapters, we began with the challenge of inferring what the outside world is like from within the vault of a bony skull, and then following the thread of ideas inwards to the body – dealing first with experiences of selfhood as perceptual best guesses, and finally identifying the most deep-seated of these experiences with control-oriented perception of the body itself. With the FEP, it’s the other way around. We start with the simple statement that ‘things exist’ and proceed outwards from there to the body and to the world

The third benefit of the FEP lies in the rich mathematical toolbox that it brings to the table. This toolbox offers many new opportunities to further develop the ideas I’ve presented in previous chapters. Here’s one example. When we unpack the mathematics of the FEP in more detail, we discover that what I really need to do, in order to stay alive, is to minimise free energy in the future – not just in the here and now.

it turns out that minimising this long-term prediction error means I need to seek out new sensations now that reduce my uncertainty about what would happen next, if I did such-and-such. I become a curious, sensation-seeking agent – not someone content to self-isolate in a dark room

The FEP and IIT are both grand theories, but they are grand in different ways. The FEP starts from the simple statement that ‘things exist’ and derives from this the whole of neuroscience and biology, but not consciousness. IIT starts from the simple statement ‘consciousness exists’ and from there launches a direct assault on the hard problem. It’s not surprising that they often talk past each other.

My own ideas about controlled hallucinations and beast machines chart a middle course. They share with the FEP a deep theoretical grounding in the nature of the self, and they leverage the powerful mathematical and conceptual machinery of the predictive brain. They share with IIT a clear focus on the subjective, phenomenological properties of consciousness – though with the real problem, not the hard problem, in the cross hairs. Rather than pitting the FEP against IIT, my hope is that the beast machine theory of consciousness and self provides a way to bring them together, weaving insights from both into a satisfying picture of why we are what we are

How can statistically expected states also be improbable? This is possible when a system inhabits only a restricted repertoire or subset of states – a so-called ‘attracting set’ – out of a large number of possible states. The attracting set is statistically expected because that is where the system is usually found, but it is also improbable because there are many more states outside the set than inside it. There are many more ways of being mush than there are of being alive.

What is the aspect of being you that you cling to most tightly? For many, it’s the feeling of being in control of your actions, of being the author of your thoughts. It’s the compelling but complex notion that we act according to our own free will.

But amid all this confusion there is one stable intuition. When we exercise free will there is – in the words of the philosopher Galen Strawson – a feeling of ‘radical, absolute, buck-stopping up-to-me-ness in choice and action’.

Perhaps more than any other kind of experience, experiences of volition make us feel that there is an immaterial conscious ‘self’ pulling strings in the material world. This is how things seem.

But experiences of volition do not reveal the existence of an immaterial self with causal power over physical events. Instead, I believe that they are distinctive forms of self-related perception. More precisely, that they are self-related perceptions associated with voluntary actions. Like all perceptions – whether self-related or world-related – experiences of volition are constructed according to the principles of Bayesian best guessing, and they play important – likely essential – roles in guiding what we do.

Let’s first be clear about what free will is not. Free will is not an intervention in the flow of physical events in the universe, more specifically in the brain, making things happen that wouldn’t otherwise happen. This ‘spooky’ free will invokes Cartesian dualism, demands freedom from the laws of cause and effect, and offers nothing of explanatory value in return.

Whether determinism matters for free will has been the topic of endless debate.

My former boss Gerald Edelman summed it up well with a provocative one-liner: Free will – whatever you think about it, we’re determined to have it.

In the early 1980s, at the University of California in San Francisco, the neuroscientist Benjamin Libet performed a series of experiments on the brain basis of voluntary action which have remained controversial ever since. Libet took advantage of a well-known phenomenon called the ‘readiness potential’ – a small slope-like EEG signal, originating from somewhere over the motor cortex, that reliably precedes voluntary actions. Libet wanted to know whether this brain signal could be identified not only prior to a voluntary action, but before the person was even aware of the intention to make the action.

His experimental set-up, shown in the image opposite, was straightforward. Libet asked his participants to flex their dominant wrist at a time of their own choosing – to make a spontaneous voluntary action, just as Briony does in McEwan’s novel. Each time they did this, he measured the precise time of the movement, while using EEG to record brain activity both before and after the onset of the movement. Crucially, he also asked his volunteers to estimate when they experienced the ‘urge’ to make each movement: the precise moment of conscious intention, the crest of the breaking wave. They did this by noting the angular position of a rotating dot on an oscilloscope screen at the time they experienced the intention to move, and then reporting this position later on.

The data were clear. After averaging across many trials, the readiness potential was identifiable hundreds of milliseconds before the conscious intention to move. In other words, by the time a person is aware of their intention, the readiness potential has already started ramping up.

A common interpretation of Libet’s experiment is that it ‘disproves free will’. Indeed, it is clearly bad news for spooky free will (not that more bad news is needed) because it seems to exclude the possibility that the experience of volition caused the voluntary action

If there isn’t any genuine (i.e. spooky) free will, Libet thought, maybe there’s still ‘free won’t’. This is a cute trick, but of course it doesn’t work. Conscious inhibition is no more a little miracle than the original conscious intention.

It wasn’t until 2012 that a new idea and a clever experiment properly shook things up, when the neuroscientist Aaron Schurger realised that readiness potentials might not be signatures of the brain initiating an action, but might instead be artefacts of the way they are measured.

Readiness potentials are typically measured by looking backward in time, at the EEG, starting from all those moments at which a voluntary action actually occurred. What Schurger realised is that, by doing this, researchers systematically ignore all the other times when voluntary actions don’t happen. What would the EEG look like at these other times? Perhaps there is activity similar to readiness potentials going on all the time, but we don’t see it, because we aren’t looking for it?

Schurger attacked this problem through a clever modification of the Libet design in which people continued to make spontaneous voluntary actions, but were also occasionally prompted, by a loud beep, to make the same action in a non-voluntary, stimulus-driven way. His key finding was that when his volunteers were quick to respond to the beep, their EEG showed what looked like a readiness potential, extending back long before the beep, even though they hadn’t been preparing any voluntary action at these times. By contrast, when looking at the EEG preceding slow responses to the beep, there was little sign of anything resembling a readiness potential.

Schurger interpreted his data by proposing that the readiness potential is not a signature of the brain initiating a voluntary action, but a fluctuating pattern of brain activity that occasionally passes a threshold, triggering a voluntary action when it does so.

And this is why, when an action is triggered by a beep, the behavioural response will be quicker if this fluctuating activity happens to be close to the threshold and slower if it happens to be far away

My preferred interpretation returns to the idea I started with: that experiences of volition are forms of self-related perception. Through the lens of Schurger’s experiment, readiness potentials look a lot like the activity associated with the brain accumulating sensory data in order to make a Bayesian best guess. In other words, they are the neural fingerprints of a special kind of controlled hallucination.

I just made a cup of tea

Let’s use this example to develop the view of experiences of volition – and voluntary actions too – as self-related perceptions. There are three defining features that characterise most, if not all, experiences of volition.

The first defining feature is the feeling that I am doing what I want to do.

I was thirsty and tea was available, nobody was restraining me or force-feeding me hot chocolate, so I made some tea and drank it.

Although making tea was fully consistent with my beliefs, values, and desires, I did not choose to have these beliefs, values, and desires. I wanted a cup of tea, but I did not choose to want a cup of tea.

They are voluntary because they express what I, as a person, want to do, even though I cannot choose these wants. As nineteenth-century philosopher Arthur Schopenhauer put it, ‘Man can do what he wills, but he cannot will what he wills.’

The second defining feature is the feeling that I could have done otherwise. When I experience an action as voluntary, the character of the experience is not only that I did X, but that I did X and not Y, even though I could have done Y. I made tea. Could I have done otherwise? In one sense, yes. There’s coffee in the kitchen too, so I could have made coffee.

The third defining feature is that voluntary actions seem to come from within rather than being imposed from somewhere else. This is the difference between the experience of a reflex action, like the rapid withdrawal of my foot when I accidentally stub my toe, and its voluntary equivalent – such as when I deliberately swing my foot backwards as I prepare to kick a ball

Altogether, we perceive an action as being voluntary – as being ‘freely willed’ – when we infer that its causes come predominantly from within, in a way that is aligned with one’s beliefs and goals, detached from alternative potential causes in the body or in the world, and that suggests the possibility of having done otherwise. This is what experiences of volition feel like from the inside, and also what voluntary actions look like from the outside.†

The next step is to ask how the brain enables and implements such actions. Here’s where ‘degrees of freedom’ – the title of this chapter – enters the story. In engineering and mathematics, a system has degrees of freedom to the extent that it has multiple ways of responding to some state of affairs.

Voluntary behaviour depends on the competence to control all these degrees of freedom, in ways that are aligned with our beliefs, values, and goals, and that are adaptively detached from the immediate exigencies of the environment and body.

Following the neuroscientist Patrick Haggard, we can think of this network as implementing three processes: an early ‘what’ process specifying which action to make, a subsequent ‘when’ process determining the timing of the action, and a late-breaking ‘whether’ process, which allows for its last-minute cancellation or inhibition.

The ‘what’ component of volition integrates hierarchically organised sets of beliefs, goals, and values together with perceptions of the environment, in order to specify a single action out of many possibilities. I move my hand to the kettle because I am thirsty, I like tea, it is the right time of day, the kettle is within reach, no wine is available – and so on.

The ‘when’ component specifies the timing of a selected action, and is most closely associated with the subjective urge to move –

The brain basis of this process localises to the same regions that are associated with the readiness potential.

Indeed, gentle electrical brain stimulation of these regions – in particular the supplementary motor area – can generate a subjective urge to move, even in the absence of any movement.

And the final ‘whether’ component provides a last-minute check on whether the planned action should go ahead. When we call off an action at the very last moment – perhaps I’m out of milk – it’s this process of ‘intentional inhibition’ that kicks in. These inhibitory processes are also localisable to more frontal parts of the brain.

And it is the perception of the operation of this network, its looping through the body, out into the world, and back again, that underpins subjective experiences of volition.

The perceptual experience of volition is a self-fulfilling perceptual prediction, another distinctive kind of controlled – again perhaps a controlling – hallucination.

Experiences of volition are useful for guiding future behaviour, just as much as for guiding current behaviour.

Experiences of volition flag up instances of voluntary behaviour so that we can pay attention to their consequences, and adjust future behaviour so as to better achieve our goals.

I mentioned earlier that our sense of free will is very much about feeling we ‘could have done differently

The feeling that I could have done differently does not mean that I actually could have done differently. Rather, the phenomenology of alternative possibilities is useful because in a future similar, but not identical, situation I might indeed do differently.

The physical world does not duplicate itself from day to day, not even from millisecond to millisecond. At the very least, the circumstances of my brain will have changed, because I’ve had an experience of volition on Monday and paid attention to its consequences. This, by itself, is enough to affect how my brain can control my many degrees of freedom when setting out to work again on Tuesday.‡ The usefulness of feeling ‘I could have done otherwise’ is that, next time, you might.

Altogether, the ability to exercise and to experience ‘free will’ is the capacity to perform actions, to make choices – and to think thoughts – that are uniquely your own.

The renowned psychologist Daniel Wegner captured this spirit with his book The Illusion of Conscious Will, which has remained influential since its publication nearly twenty years ago

Spooky free will certainly isn’t real.

When I play the piano, or make a cup of tea, most of the time these voluntary actions unfold with an automaticity and fluency that undermines not only the intuition that I somehow cause the actions, but also the less frequently examined intuition that such actions even seem to be caused by anything.

When people talk about being ‘in the moment’ or in a ‘state of flow’ – when deeply immersed in an activity they have extensively practised – the phenomenology of volition may be entirely absent. Much of the time, our voluntary actions, and our thoughts – well, they ‘just happen’. When it comes to free will it’s not only that how things seem is not how they are. How things seem deserves closer examination too.

This kind of freedom is both a freedom from and a freedom to. It is a freedom from immediate causes in the world or in the body, and from coercion by authorities, hypnotists and mesmerists, or social-media pushers.

It is not, however, freedom from the laws of nature or from the causal fabric of the universe. It is a freedom to act according to our beliefs, values, and goals, to do as we wish to do, and to make choices according to who we are.

People with anarchic hand syndrome make voluntary actions which they do not experience as being theirs, while those with akinetic mutism are unable to make any voluntary actions at all. An awkwardly located brain tumour can transform an engineering student into a mass school shooter, as happened in the case of Charles Whitman, the ‘Texas Tower Sniper’, or engender in a previously blameless teacher a rampant paedophilia – a tendency which disappeared when the tumour was removed, and returned when it grew back.

Einstein stated in a 1929 interview that because he didn’t believe in free will he took credit for nothing.

It is also a mistake to call the experience of volition an illusion. These experiences are perceptual best guesses, as real as any other kind of conscious perception, whether of the world or of the self. A conscious intention is as real as a visual experience of colour. Neither corresponds directly to any definite property of the world – there is no ‘real red’ or ‘real blue’ out there, just as there is no spooky free will in here – but they both contribute in important ways to guiding our behaviour, and both are constrained by prior beliefs and sensory data.

However, the ability to exercise free will might come in degrees not only among us humans, but more widely among the animals we share our world with. And if the ability to exercise free will extends to other species, what can be said about the extent of consciousness itself? It’s time to look beyond the human.

Heraclitus: ‘No man ever steps in the same river twice, for it’s not the same river and he’s not the same man.’

From the early ninth century all the way up until the mid 1700s, it was not uncommon for European ecclesiastical courts to hold animals criminally responsible for their actions. Pigs were executed or burned alive, as were bulls, horses, eels, dogs and, on at least one occasion, dolphins.

In the almost two hundred cases documented in E. P. Evans’ 1906 history of animal criminal prosecution, pigs were the most common offenders, probably because they roamed rather freely in medieval villages.

Their crimes varied from eating children to consuming consecrated biscuits. Sometimes they were charged with abetting – through their grunting and snorting – the crimes of another; often they were hanged, and occasionally they were acquitted.

In one celebrated sixteenth-century case, the French lawyer Bartholomew Chassenée successfully exonerated some rats with the clever argument that they could not reasonably be expected to turn up to trial, given the dangers posed to them by the many cats lying in wait along the route.

In other cases, including various weevil infestations, the offending animals were issued with written orders to leave a property or a barley crop, often on a specific day and even by a specific hour

The idea that animals could comprehend, and reasonably submit to, the arcane procedures of ecclesiastical law was and is borderline insane. But along with this idea comes a recognition that animals might have conscious experiences, and might be equipped with minds able – in some sense – to make decisions.

The first thing to say is that we cannot judge whether an animal is conscious by its ability – or inability – to tell us that it is conscious. Absence of language is not evidence for absence of consciousness. Neither is absence of so-called ‘high-level’ cognitive abilities like metacognition – which is the ability, broadly speaking, to reflect on one’s thoughts and perceptions.

Although animal experiments can shed light on the mechanisms of consciousness in humans, it is unwise to infer the existence of animal consciousness solely on the basis of superficial similarity to Homo sapiens.

Doing so carries the twin risks of anthropomorphism – the attribution of humanlike qualities to the non-human – and anthropocentrism – the tendency to interpret the world in terms of human values and experiences.

Above all, we should be suspicious of associating consciousness too closely with intelligence. Consciousness and intelligence are not the same thing. Using the latter as a litmus test for the former commits a number of errors

And it encourages a methodological laziness, since it justifies accepting ‘intelligent’ capabilities like language and metacognition – which are easier to assess than consciousness itself – as sufficient for inferring consciousness.

Intelligence is not irrelevant to consciousness. Other things being equal, intelligence opens up new possibilities for conscious experience. You can be sad or disappointed without much cognitive competence at all, but to feel regret – or anticipatory regret – requires enough mental capability to consider alternative outcomes and courses of action. Even rats, one study suggests, might experience a rodent version of regret rather than just disappointment, when things don’t turn out as hoped for.‡

Inferences about non-human consciousness must tread a fine line. We need to be wary of imposing our anthropocentric view, but at the same time we have little option but to use humans as a known quantity; a firm foundation from which to reach outwards.

The beast machine theory developed in this book makes the case that consciousness is more closely connected with being alive than with being intelligent. Naturally, this applies as much to other animals as it does to us humans. On this view, consciousness may be more widespread than it would seem, were we to take intelligence as the primary criterion. But it does not mean that wherever there is life there is also consciousness.

Let’s start with mammals – a grouping which includes rats, bats, monkeys, manatees, lions, hippos, and of course humans too. I believe that all mammals are conscious. Of course, I don’t know this for sure, but I am pretty confident. This claim is not based on superficial similarity to humans, but on shared mechanisms. If you set aside raw brain size – which has more to do with body size than with anything else – mammalian brains are strikingly similar across species.

Back in 2005, the cognitive scientist Bernard Baars, myself, and David Edelman – the son of Gerald Edelman, and an expert on animal cognition – made a list of the properties of human consciousness which we thought could be readily tested for in other mammals. We came up with seventeen distinct properties. This was in some ways an arbitrary number, but it demonstrated the reasonableness of asking experimentally testable questions about animal consciousness.

The first properties we thought about had to do with anatomical features of the brain.

There are common features of brain activity too. Among the most striking are the changes in brain dynamics as animals fall asleep and wake up – the dynamics underlying conscious level.

In normal waking states, all mammals show irregular, low-amplitude, and fast electrical brain activity. And when sleep comes, all mammalian brains switch to more regular, large-amplitude brain dynamics.

These patterns and changes closely resemble what is seen in humans during waking and sleeping. General anaesthesia, too, has similar effects across mammalian species – a widespread breakdown of communication between brain regions that accompanies total behavioural unresponsiveness.

There are of course many differences as well, especially in patterns of sleep. Seals and dolphins sleep with half their brain at a time, koalas sleep for about twenty-two hours each day while giraffes get by on less than four, and newborn killer whales do not sleep at all in the first month of life. Nearly all mammals have periods of rapid eye movement (REM) sleep – though seals do so only while sleeping on land, and dolphins apparently not at all.

Besides conscious level, there will also be substantial differences in conscious contents across mammalian species.

Mice rely on their whiskers, bats on their echolocating sonar, and naked mole rats on their acute sense of smell – especially when meeting other naked mole rats. These differences in perceptual dominance will mean that each animal will inhabit a distinctive inner universe.

More intriguing still are differences that relate to experiences of selfhood. In humans, a notable marker of the development of high-level self-consciousness, of the sort related to personal identity, is the ability to recognise oneself in the mirror. In humans, this ‘mirror self-recognition’ ability tends to develop some time between eighteen and twenty-four months of age.

In animals, the capacity for self-recognition has been extensively investigated using a test developed by the psychologist Gordon Gallup Jr in the 1970s. In the classic version of his mirror self-recognition test, an animal is anaesthetised and then marked, usually with paint or a sticker, in a place on its body that it cannot normally see. When the anaesthesia wears off, the animal is allowed to interact with a mirror so that it can see the mark. If, having looked into the mirror, it spontaneously looks for the mark on its own body, instead of investigating the mirror image, it has passed the test.

This criterion is based on the reasoning that the creature has recognised that the mirror image depicts its own body, rather than the body of another animal.

Who passes the mirror test? Among mammals, some great apes, a few dolphins and killer whales, and a single Eurasian elephant. A parade of other mammalian creatures, including pandas, dogs, and various monkeys, have failed – at least so far.

Animals could fail the mirror test for many reasons besides lacking self-recognition abilities. These include not liking mirrors, not understanding how they work, or even just preferring to avoid eye contact. Recognising this, researchers are continually developing new versions of the test that are tuned ever more astutely to different interior universes

For example, dog self-recognition can now be tested with ‘olfactory mirrors’ – though they still don’t do very well (delightfully, cognition in dogs is known as ‘dognition

These differences strike me especially forcefully in the case of monkeys. Although chimpanzees and the great apes are our nearest evolutionary neighbours, monkeys are not too far away and they’ve long been used in neuroscientific experiments as ‘primate models’ of humans, especially when it comes to vision. In some studies, monkeys have even been trained to deliver ‘reports’ – for example, by pressing a lever – about whether they ‘saw’ something or not. These experiments can be directly compared with human studies in which people say what they experience, or don’t experience, providing a primate equivalent of a key method in consciousness research.

Given their many similarities to humans, for me there is no doubt that monkeys possess some kind of conscious selfhood. If you hang around with monkeys for any length of time, the impression of being among other conscious entities – other conscious selves – is completely convincing.

Cayo Santiago is also known as ‘monkey island’ because its only permanent residents are rhesus macaque monkeys – more than a thousand of them.

The population was transplanted there in 1938 from Kolkata by an eccentric American zoologist, Clarence Ray Carpenter, who had become tired of trekking all the way to India.

When two monkeys took turns to climb a tree and leap from its branches into a pond below, it seemed to be for no other reason than the sheer spontaneous pleasure of it. They were having fun.¶

Equally compelling are videos of capuchin monkeys reacting to deliberate unfairness. In one video, popularised by the primatologist Frans de Waal, two monkeys are housed in adjacent cages and are rewarded, one after the other, for passing a small stone to an experimenter. Monkey One passes the stone through the cage mesh and is rewarded with a small slice of cucumber, which she happily eats. Monkey Two does the same thing and is given, not cucumber, but a grape – a much tastier morsel. Monkey Two eats the grape while Monkey One looks on. When Monkey One repeats the task and is again given cucumber, she looks at it, throws it back at the experimenter, and rattles the cage in apparent indignation.

Having fun and throwing tantrums are powerful intuition pumps. These behaviours are so distinctive that it’s almost impossible to interpret them as anything other than the outward manifestation of apparently humanlike inner states.

Monkeys, as mentioned, have consistently failed the mirror test. While monkeys are undoubtedly conscious, and while I also believe they experience some kind of selfhood, they are not furry little people. The shaping of our intuitions by anthropomorphism and anthropocentrism becomes even more apparent when we look beyond mammals. Especially when we look as far as some of our most distant evolutionary relatives.

Fiorito’s lab was arranged with two rows of tanks lining a central walkway, one octopus in each tank. (Octopuses are generally not social creatures and can even be cannibalistic.) On this particular day, David had chosen a tank in the left-hand row, about halfway along. When I walked in to see what was going on, I was astonished to see all the octopuses on the other side of the walkway pressed up against the glass of their tanks, every one of them staring intently at David while he repeatedly lowered his objects into his chosen tank. The observing octopuses seemed to be trying to figure out what was going on for no other reason than the sheer interest of it.

The mind of an octopus is an independently created evolutionary experiment, as close to the mind of an alien as we are likely to encounter on this planet. As scuba-diving philosopher Peter Godfrey-Smith put it, ‘If we want to understand other minds, the minds of cephalopods are the most other of all.’

These extraordinary bodies are complemented by highly sophisticated nervous systems. Octopus vulgaris has about half a billion neurons, roughly six times more than a mouse. Unlike in mammals, most of these neurons – about three fifths – are in its arms rather than in its central brain, a brain which nonetheless boasts about forty anatomically distinct lobes.

Octopuses do things differently even at the level of genes. In most organisms, genetic information in DNA is transcribed directly into shorter sequences of RNA (ribonucleic acid) which are then used to make proteins – the molecular workhorses of life. This is a well-established, textbook-level principle of molecular biology. But in 2017 this principle was upended by the discovery that RNA sequences in octopuses – and in a few other cephalopods – can undergo significant editing before being translated into protein.

It’s as if the octopus is able to rewrite parts of its own genome on the fly.

What’s more, for the octopus, much of this RNA editing seems to be related to the nervous system. Some researchers have suggested that this prolific genome rewriting ability may partly underlie the impressive cognitive abilities of octopuses.

And their cognitive abilities certainly are impressive. They can retrieve hidden objects – usually tasty crabs – from within nested Plexiglas cubes, find their way through complex mazes, try out a range of different actions to solve a particular problem, and – as Fiorito himself has shown at the Stazione – learn simply through observing other octopuses.

These feats of cephalopod intelligence are certainly compelling evidence of a mind at work. But what kind of mind? I’ve already said that we should not put too much weight on intelligence as a benchmark for consciousness. So what can be said about what it is like to be an octopus? To get a handle on this, we need to connect octopus behaviour to octopus perception.

Their camouflage system must therefore somehow encode knowledge about the visual abilities of these predators.

What’s even more surprising is that octopuses do all this despite being colour-blind. The light-sensitive cells in human eyes respond to three different wavelengths of light, creating from their mixtures a universe of colour. The cells in octopus eyes, however, contain only one photopigment.

The same colour blindness is true for the light-sensitive cells embedded throughout their skin: it turns out that octopuses can ‘see’ with their skin, as well as with their eyes. Added to this, octopus chromatophore control is thought to be ‘open-loop’, meaning that the neurons in the chromatophore lobe do not generate any obvious internal copy of the signals sent out to the chromatophores in the skin. The central brain might not even know what its skin is doing.

Along with vision, octopuses share some of the other classic sensory modalities with mammals and other vertebrates. They can taste, smell, and touch – and they can hear too, though not very well. There’s still bizarreness to contend with, because octopuses can taste with their suckers as well as with their central mouth parts. This again points to a remarkable decentralisation of mind in these creatures.

Octopus arms behave like semi-autonomous animals: a severed octopus arm can still execute complex action sequences, like grasping pieces of food, for some time after being separated from the body.

These degrees of freedom and decentralised control together pose an intimidating challenge for any central brain trying to maintain single, unified perception of what is, and what is not, part of its body. Which is why octopuses may not even bother. As odd as it sounds, what it is like to be an octopus may not include an experience of body ownership in anything like the sense in which it applies to humans and other mammals.

This doesn’t mean that octopuses do not distinguish ‘self’ from ‘other’. They clearly do – and they need to. For a start, they need to avoid getting tangled up with themselves.

It turns out that this ability depends on a simple but effective system of taste-based self-recognition. Octopuses secrete a distinctive chemical throughout their skin. This chemical serves as a signal that can be detected by the suckers, so that they do not reflexively attach.

The detached arms would readily grip onto the arms from which the skin had been removed, but would never grip onto the intact arms.

Octopuses push hard at our intuitions about how different animal consciousness might be from our own. But in leaping straight from monkeys to cephalopods we’ve skipped over an enormous menagerie.

Birds make a pretty strong case for sentience. Avian brains, while significantly different from mammalian brains, nevertheless have an organisation which can be mapped quite closely onto the mammalian cortex and thalamus.

Parrots can count, cockatoos can dance, and scrub jays can stash food according to their future needs. While these examples of smartness suggest that some birds may enjoy complicated states of consciousness, remember that intelligence is not a litmus test for awareness. Non-food-hiding, non-speaking, non-dancing birds likely have conscious experiences too.

Instead of basing these inferences on similarity to mammalian brains and behaviours, a better strategy might be to adopt the beast machine perspective – mine, not the Cartesian version – which traces the origin and function of conscious perception to physiological regulation, and to preservation of the integrity of the organism. This suggests that one place to look for evidence of awareness is in how animals respond to supposedly painful events.

And while there are infinitely many ways in which living creatures may suffer, the most widely shared likely involve basic challenges to their physiological integrity.

Most vertebrates (animals with backbones) will tend to an injured body part. Even the tiny zebrafish will pay a ‘cost’ to access pain relief upon injury, shifting from a natural environment to a barren, brightly lit tank when that tank is suffused with analgesia. Whether this implies that fish are conscious – and there are many types of fish – is unclear, but it is certainly suggestive.

What about insects? Ants do not limp when a leg is damaged. However, their hard-bodied exoskeletons might be less susceptible to pain, and insect brains do possess forms of the opiate neurotransmitter system that is commonly associated with pain relief in other animals.

recent study found that the fruit fly Drosophila melanogaster displays a post-injury hypersensitivity to previously non-painful stimuli in a way that resembles ‘chronic pain’ in humans

And, remarkably, anaesthetic drugs seem to be effective across all animals, from single-celled critters all the way to advanced primates.

Conscious experience is central to our lives, but this doesn’t mean its biological basis is straightforward.

Setting aside its inevitable uncertainties, the study of animal consciousness delivers two profound benefits. The first is a recognition that the way we humans experience the world and self is not the only way. We inhabit a tiny region in a vast space of possible conscious minds, and the scientific investigation of this space so far amounts to little more than casting a few flares out into the darkness. The second is a newfound humility.

Not only can consciousness exist without all that much intelligence – you don’t have to be smart to suffer – but intelligence can exist without consciousness too.

The possibility of being smart without suffering brings us to the final leg of our journey through consciousness science. It’s time to talk about artificial intelligence, and about whether there could ever be conscious machines.

Rabbi Loew’s golem reminds us of the hubris we invite when attempting to fashion intelligent, sentient creatures – creatures in the image of ourselves, or from the mind of God.

In Prague, in the late sixteenth century, Rabbi Judah Loew ben Bezalel took clay from the banks of the Vltava River and from this clay shaped a humanlike figure – a golem.

It rarely goes well. From the monster in Mary Shelley’s Frankenstein to Ava in Alex Garland’s Ex Machina, by way of Karel Čapek’s eponymous robots, James Cameron’s Terminator, Ridley Scott’s replicants in Blade Runner, and Stanley Kubrick’s HAL, these creations almost always turn on their creators, leaving in their wake trails of destruction, melancholy, and philosophical confusion.

Running beneath many of these worries, especially the more existential and apocalyptic, is the assumption that AI will – at some point in its accelerating development – become conscious. This is the myth of the golem made silicon.

Why might we even think that a machine – an artificial intelligence – could become aware? As I just mentioned, it is quite common – though by no means universal – to think that consciousness will emerge naturally once machines pass some as yet unknown threshold of intelligence. But what drives this intuition? I think that two key assumptions are responsible, and neither is justifiable. The first assumption is about the necessary conditions for anything to be conscious. The second is about what is sufficient for a specific thing to be conscious.

The first assumption – the necessary condition – is functionalism. Functionalism says that consciousness doesn’t depend on what a system is made out of, whether wetware or hardware, whether neurons or silicon logic gates – or clay from the Vltava River. Functionalism says that what matters for consciousness is what a system does.

My attitude towards functionalism is one of suspicious agnosticism.

For artificially intelligent computers to become conscious, functionalism would have to be true.

But functionalism being true is, by itself, not enough: information processing by itself is not sufficient for consciousness. The second assumption is that the kind of information processing that is sufficient for consciousness is also that which underpins intelligence. This is the assumption that consciousness and intelligence are intimately, even constitutively, linked: that consciousness will just come along for the ride.

We are conscious, we are intelligent, and we are so species-proud of our self-declared intelligence that we assume that intelligence is inextricably linked with our conscious status and vice versa.

If we persist in assuming that consciousness is intrinsically tied to intelligence, we may be too eager to attribute consciousness to artificial systems that appear to be intelligent, and too quick to deny it to other systems – such as other animals – that fail to match up to our questionable human standards of cognitive competence

The problem with exponential curves – as many of us learned during the recent coronavirus pandemic – is that wherever you stand on them, what’s ahead looks impossibly steep and what’s behind looks irrelevantly flat.

Then there are our Promethean fears that our creations will turn on us in some way or another – fears which have been recognised, repackaged, and sold back to us by any number of science fiction movies and books.

In this depiction, you’ll notice that current AI is located quite low on the intelligence scale. This is because it’s unclear whether current AI systems are intelligent in any meaningful sense. Much of today’s AI is best described as sophisticated machine-based pattern recognition, perhaps spiced up with a bit of planning. Whether intelligent or not, these systems do what they do without being conscious of anything.

What’s more, there may be many forms of intelligence that deviate from the humanlike, complementing rather than substituting or amplifying our species-specific cognitive toolkit – again without consciousness being involved.

Just making computers smarter is not going to make them sentient. But this does not mean that machine consciousness is impossible. What if we were to try to design in consciousness from the outset? If not intelligence, what would it take to build a conscious machine?

For example, according to a proposal in the journal Science in 2017, a machine could be said to be conscious if it processes information in ways that involve ‘global availability’ of the information, and that allow ‘self-monitoring’ of its performance.

According to IIT, any machine that generates integrated information, whatever it is made out of, and no matter what it might look like from the outside, will have some degree of consciousness. However, IIT also leaves open the possibility that machines might appear to an external observer to be conscious, or intelligent, or perhaps both, but have mechanisms that generate no integrated information at all, and therefore be lacking in consciousness altogether.

The beast machine theory grounds experiences of world and self in a biological drive towards physiological integrity – towards staying alive. What does this theory say about the possibility of conscious machines?

This robot behaves autonomously, doing the right thing at the right time to fulfil its goals. In doing so, it gives the outward impression of being an intelligent, sentient agent. Internally, its mechanisms map directly onto the predictive machinery which I’ve suggested underlies basic human experiences of embodiment and selfhood. It is a silicon beast machine. Would such a robot be conscious?

The unsatisfying but honest answer is that I don’t know for sure, but probably not. The beast machine theory proposes that consciousness in humans and other animals arose in evolution, emerges in each of us during development, and operates from moment to moment in ways intimately connected with our status as living systems

One reason for this is that the imperative for regulation and self-maintenance in living systems isn’t restricted to just one level, such as the integrity of the whole body. Self-maintenance for living systems goes all the way down, even down to the level of individual cells. Every cell in your body – in any body – is continually regenerating the conditions necessary for its own integrity over time.

This shouldn’t be taken to imply that individual cells are conscious, or that all living organisms are conscious. The point is that the processes of physiological regulation that underpin consciousness and selfhood in the beast machine theory are bootstrapped from fundamental life processes that apply ‘all the way down’. On this view, it is life, rather than information processing, that breathes the fire into the equations.

Garland’s dialogue so elegantly captures the challenge of ascribing consciousness to a machine that the term ‘the Garland test’ is now itself gaining traction – a rare example of science fiction feeding back into science.

In Alex Garland’s 2014 film Ex Machina, reclusive billionaire tech genius Nathan invites hotshot programmer Caleb to his remote hideout to meet Ava, the intelligent, inquisitive robot he has created. Caleb’s task is to figure out whether Ava is conscious, or whether she – it – is merely an intelligent robot, with no inner life at all.

The Japanese roboticist Hiroshi Ishiguro has spent decades building robots that are as similar as possible to human beings. He calls them ‘Geminoids’. Ishiguro has created Geminoid copies of himself (see opposite) and his daughter (then six years old), as well as a Geminoid Japanese-European TV anchorwoman based on a blend of about thirty different people. Each Geminoid is constructed from detailed 3D body scans and has pneumatic actuators able to generate a wide range of facial expressions and gestures. There is precious little AI as such in these devices – they are all about human mimicry, with possible applications in, among other things, remote presence or ‘telepresence’. Ishiguro once used his Geminoid to deliver a forty-five-minute remote lecture to 150 undergraduate students

This kind of reaction exemplifies the so-called ‘uncanny valley’, a concept originated by another Japanese researcher, Masahiro Mori, in 1970. Mori proposed that as a robot begins to look humanlike, it will elicit increasingly positive and empathetic reactions from people (think C-3PO in Star Wars

But once it passes a certain point, where it appears strikingly human in some ways but falls short in others, these reactions will turn rapidly to revulsion and fear – the uncanny valley – only recovering as the resemblance becomes even closer, to the point of indistinguishability. There are many theories about why the uncanny valley exists, but there is little doubt that it does. Although real-world robots find it hard to escape from the uncanny valley, developments in the virtual world are already clambering back up the slope and out the other side. Recent advances in machine learning using ‘generative adversarial neural networks’ – GANNs for short – can generate photorealistic faces of people who never actually existed (see opposite).* These images are created by cleverly mixing features from large databases of actual faces, employing techniques similar to those we used in our hallucination machine (described in chapter 6). When combined with ‘deepfake’ technologies, which can animate these faces to make them say anything, and when what they say is powered by increasingly sophisticated speech recognition and language production software, such as GPT-3, we are all of a sudden living in a world populated by virtual people who are effectively indistinguishable from virtual representations of real people. In this world, we will become accustomed to not being able to tell who is real and who is not.

Another possibility is that the circle of our moral concern will be distorted by our anthropocentric tendency to experience greater empathy for entities towards which we feel greater similarity. In this scenario we may care more about our next-generation Geminoid twins than we do about other humans, let alone about other animals.

Imagine a system for which the distinction between positive and negative feelings does not even apply, for which there is no corresponding phenomenological dimension. The ethical challenge here is that we would not even know what the relevant ethical issues were.

We should not blithely forge ahead attempting to create artificial consciousness simply because we think it’s interesting, useful, or cool. The best ethics is preventative ethics

We even have the capability to develop fully synthetic organisms built from the ‘genes up’: in 2019, researchers in Cambridge created a variant of Escherichia coli with a fully synthetic genome. The ethics of creating new forms of life is suddenly very relevant indeed.

And perhaps it will be biotechnology, rather than AI, that brings us closest to synthetic consciousness. Here, the advent of ‘cerebral organoids’ is of particular significance. These are tiny brain-like structures, made of real neurons, which are grown from human pluripotent stem cells (cells which can differentiate into many different forms).

Why is the prospect of machine consciousness so alluring? Why does it exert such a pull on our collective imagination? I’ve come to think that it has to do with a kind of techno-rapture, a deep-seated desire to transcend our circumscribed and messily material biological existence as the end times approach

A hemispherotomy involves complete neural disconnection of the brain’s dysfunctional right hemisphere. The surgeon enters the brain through the right-hand side, removes (‘resects’) the temporal lobe, and then cuts through all the bundles of connections – the white-matter tracts – that link the right hemisphere with the rest of the brain and body. The isolated hemisphere remains inside the skull, and is still connected to its blood supply. It is a living but isolated island of cortex

the operation is carried out early enough, the young brain is often sufficiently adaptable that the remaining hemisphere can pick up most or all of the slack. Despite the radical nature of this surgery, and although every case is different, outcomes are generally good

Peering into the brightly lit cavities of the child’s brain, I tried to register my abstract knowledge of different regions and pathways to the welter of tissue illuminated before me. It made little sense. The crisp cortical hierarchies and counterflowing weaves of bottom-up and top-down signalling that I knew from my research were nowhere to be seen. The brain was newly inscrutable, and I was left in awe both of the neurosurgeon’s skill and of the material reality of this most magical of objects. It felt almost transgressive. A curtain had been pulled back revealing something too intimate to be so openly on view. I was looking directly into the mechanics of a human self.

Faced with this mystery, philosophy has provided a range of options, from panpsychism (consciousness everywhere, more or less) to eliminative materialism (no consciousness, at least not how we think of it) and everything in between. But the science of consciousness isn’t about choosing from a set menu, however swanky the restaurant or skilled the chef.

This is the essence of the real problem approach to consciousness. Accept that consciousness exists, and then ask how the various phenomenological properties of consciousness – which is to say how conscious experiences are structured, what form they take – relate to properties of brains, brains that are embodied in bodies and embedded in worlds.

We began with conscious level – the difference between being in a coma and being wide awake and aware – where we focused on the importance of measurement. The key point here is that candidate measures, like causal density and integrated information, are not arbitrary. Rather, they capture highly conserved properties of all conscious experiences, namely that every conscious experience is simultaneously unified and distinct from all other conscious experiences.

The first challenge was to understand perception as an active, action-oriented construction, rather than as a passive registration of an objective external reality.

The second challenge turned this insight inwards, to the experience of being a self. We explored how the self is itself a perception, another variety of controlled hallucination.

The final challenge was to see that the predictive machinery of conscious perception has its origin and primary function not in representing the world or the body, but in the control and regulation of our physiological condition.

This is my theory of the beast machine, a twenty-first-century version – or inversion – of Julien Offray de La Mettrie’s ‘l’homme machine’. And it’s here that the deepest shifts in how to think about consciousness and selfhood take place.

Everything in conscious experience is a perception of sorts, and every perception is a kind of controlled – or controlling – hallucination. What excites me most about this way of thinking is how far it may take us. Experiences of free will are perceptions. The flow of time is a perception. Perhaps even the three-dimensional structure of our experienced world and the sense that the contents of perceptual experience are objectively real – these may be aspects of perception too.