TL;DR: Most physicalist theories of consciousness explain how brains process information, but they struggle to explain why any of that processing should feel like anything from the inside. I argue that, for a brain with our evolutionary history to avoid a crippling motivation problem, evolution had to develop conscious minds.

Argument Summary

Not a substitute for reading the whole thing (unfortunately)!

1. The Planner - Humans have an unusually powerful, flexible reasoning engine (the “Planner”). Natural selection has favoured ever more sophisticated reasoning in humans.
2. Regress - Unless there is an intrinsic notion of value anywhere, this Planner will not be able to make decisions and end up stalling. Neither a bare bodily stimulus, nor an overarching objective function will stop this motivational regress.

3. Valence - To end the regress and motivate action effectively, value data should have several properties. They should be: -

   1. Non-rationally compelling - these are bedrock impulses, unable to be ignored or reasoned away.
   2. Decision-relevant - either over-rideable by reason in the present, or relevant for future learning.
   3. Intelligible - its “message” about what’s better and worse must be clear and usable.
   4. Interest-aligning - motivate future planning towards what evolution “wants” (survival, reproduction, etc.).

   Wiring these signals in leads to the emergence of a self-model as a centralised locus of control and value.

4. Model of the World - Perceptual and memory systems must provide the Planner with a structured picture of the world and self, so that value signals can guide choices over time.
5. Cashing In the Chips - Putting all this motivational architecture together leads to something that looks exactly like phenomenal consciousness. Conscious experience isn’t an optional add-on, it’s what we need to solve the motivation problem.
6. Conclusion - Implications are there but let’s red-team the argument first. Comment and objections invited.

Introduction

“It is widely agreed that experience arises from a physical basis, but we have no good explanation of why and how it so arises.” Chalmers 1995

This quote marked the beginning of a new chapter in the philosophy of mind. In framing the “hard problem of consciousness”, David Chalmers got people talking about consciousness again, reigniting a debate that stretches back to Descartes.

The “why and how” dual nature of the hard problem makes it more than a little slippery. Neuroscientists making progress on how the brain processes information can always be hit with the objection: “yes but why? Why does it feel like anything?”

Attempts to answer this Why Question also usually hit a wall. Their story tends to follow a pattern - beings need to get on in the world, so they need to predict the future, so they need to ingest and process a lot of information, and that information processing is what makes us conscious. The wall tends to appear when anyone mentions the word “qualia” - the redness of red, the agony of pain or the taste of an apple.

Some physicalists will say: “Look, the processing of all this information JUST IS consciousness. Whatever you’re calling ‘qualia’ included. Be happy and don’t probe.” Justifiably, purists aren’t satisfied. Fundamentally, the question remains unanswered. Whatever the function of this information processing, why should that functional signature feel like anything from the inside?

Other physicalist attempts, most famously that of Daniel Dennett, argue that qualia are an illusion - a fiction the brain tells itself. Chalmers himself calls this “utterly unsatisfying”. Our intuitions just don’t support the idea that “redness” is anything other than very real, and ignoring those intuitions entirely seems obtuse.

For us physicalists (or “Camp 1” in Rafael’s LessWrong-definitive framing), there is a temptation to ignore the question and move on. It’s inconvenient, bordering on unfair. Surely neuroscience will figure out the details at some point??

And yet, if we take a good look in the mirror, those in the non-physicalist “Camp 2” do have a point. We haven’t answered the question well enough.

So here’s an attempt at a better answer.

The basic structure will be to explore whether we can safely say that, given what we are as humans, evolution had to make us conscious - not just its functional signature, but subjective consciousness itself. (My focus is on humans for now as we’re the easiest case; the argument will I’m sure apply to many animals, but that’s one for later.)

I appreciate this is well-worn and emotionally-charged territory, so I’m going to tread the ground as carefully as possible. I’ll define the key words up front in terms that I think most of us can agree on. Then I’ll go through the argument, building up as much philosophical currency as I can, and I’ll leave it to the LW community to decide whether or not the argument earns its supper.

First, some simple definitions, derived from Nagel / Chalmers.

A subjective state is one which possesses a phenomenal character - meaning there is something it is like to be in that state. It exists inherently from a first-person point of view. Unlike objective facts (which can be fully described from the outside), subjective states are dependent on an experiencer; they are not just "occurring," they are felt.

Consciousness itself is the ability to be in subjective states. A system is conscious if there is anything it is like to be that system.

1 - The Planner

From the very beginning, evolution has demanded motivation. Without the attraction of some chemicals to other chemicals, life could not have started.

Simple organisms developed processes like phototaxis (movement towards or away from light) to generate a fixed response to a stimulus. This approach was more than sufficient for much of our evolutionary history, but as environments became more complex, some flexibility became adaptive. “Go towards food source” may be a helpful heuristic, but only until one of your predators is lurking nearby.

In humans, this flexible architecture has reached an advanced level of reasoning. We can:

• Imagine detailed future end-states (“angry alpha male placated”)
• Cycle between counterfactuals (“what if I had…”)
• Encode and manipulate symbols (“This log should burn well”)
• Reason over time (“I may have more help with this tomorrow”)
• Plan multi-step sequences (“Sharpening this flint would enable me to…)
• Assess success chances (“better not risk it”)

Impressive as these capabilities are, we don’t need them for everything. In cases where complex reasoning would be slow and unnecessary, heuristics like “remove hand from hot stove” kick in automatically. What is needed, though, is the ability to recruit these abilities quickly when they’re needed.

Our reasoning circuitry has had various labels over the years. It roughly aligns with Dual Process Theory’s “System 2” and with what computational neuroscientists call “model-based planning” in reinforcement learning. I’ll use the term “the Planner”, which will hopefully help reinforce intuitions around the role these abilities play in our brains. To avoid confusion though - I’m not suggesting that they occupy one location in the brain, and I’m certainly not positing a little-person-in-the-head (“homunculus”). Although the Planner does seem to be rooted largely in the brain’s fronto-parietal networks, it is not a single physical module but a distributed set of capabilities. It’s a slow, computationally limited, metabolically expensive, far-from-perfect reasoning system, but one that has enabled us to create hand-axes, cathedrals and space stations.

In human beings at least, natural selection has favoured an ever more sophisticated Planner, involved in ever more decisions. The basic reason for this is simple: the world has structure. Physics is lawful, predators and prey have habits, other humans have patterns of behaviour. A creature that can pause, simulate a few futures, and then act accordingly will tend to outperform one that generates a fixed response. If we rewire from a simple stimulus-response loop to a stimulus-Planner-response loop, we often improve our odds.

2 - Regress

The complexity arises when we consider how this rewiring process could work in practice. What exact signal should the Planner receive? The flows above make it seem like we can just reroute the original stimulus to the Planner. But the original stimulus simply triggered an action. If that’s good enough - as it is in many cases - there’s no reason to recruit the Planner’s services. But if we want a stimulus to be overridable in favour of something else, we’re going to need something richer than simply “do this”.

Selection pressure favours decisions being taken according to their expected value. Actions should be chosen based on predicted outcomes weighted by value and the chances of success. For that, the Planner needs underlying data about value, i.e. what actually counts as good or bad.

An outcome can be “good” in one of two ways. Either it appears to the Planner as intrinsically good (a terminal goal), or its value is derived from some more fundamental goal that itself appears as intrinsically good.

If there is nothing intrinsic anywhere in this chain - if every “why care?” answer just points to another goal - we end up in a regress.

Why should the Planner hunt? Because organism needs food.
Why should the Planner care if organism needs food? Because organism might die.
Why should it care if organism dies? Because the Planner itself might cease to exist?

Even here, the regress doesn’t stop. The Planner is a reasoning engine: a system for modelling scenarios and comparing them. By default, nothing is valuable, not even its own continued operation.

Evolution has to halt the regress somehow, or the Planner stalls.

In practice, there seems to be only one plausible option for this, and that is to create signals that the Planner values intrinsically. This intuition is important, so I’ll cycle through a few possible alternatives.

1. Hold on. Surely the original stimulus could communicate value?
   1. The original stimulus might be able to communicate information, like “Water reserves = 3%”. But unless this information is intrinsically motivating, or there’s a higher objective to derive value from, it won’t lead to action. Imagine, for example, that anger was a literal “red mist” - no actual feeling, just a hue overlaid on the visual field to communicate the data. In itself, this doesn’t exert any pressure on our decisions. Intrinsic value must exist somewhere in the chain.
2. We could just hardwire in some value weights which describe the relative importance of food, drink, sleep etc.
   1. If we drive behaviour directly off value weights, we’ve bypassed the Planner and created an advanced stimulus-response process. Heart-rate, insulin release and white blood cell production are all still determined in this way. The Planner is introduced for a different class of problem - open-ended, world-facing decisions where there are genuine alternatives that require its skills to evaluate. Should the organism hunt now or later? Which route to take? Which ally to back?
   2. If you feed those weights into the Planner as inputs, only intrinsic value will give them force. Data on its own would still need a higher objective function to end the regress.
3. OK, perhaps we could do that - give the Planner an overarching objective function. That way it can understand what’s important, and let its powerful reasoning engine derive specific actions?
   1. That seems promising in theory - perhaps we could imagine an objective function like “stay alive” or “maximise inclusive fitness”. Thermostats, chess engines and reinforcement learning systems all employ runtime objective functions to control behaviour, so perhaps we could have?
   2. This view treats evolution like an engineer with a whiteboard, when in reality evolution is a tinkerer - it works with what it has. As we’ve seen, organisms have always acted using “bottom-up” stimulus-response rules: move towards food sources, withdraw from damage, cling to caregivers. When evolution starts to add a Planner, every generation still has to act using whatever drives it already has, or it dies.
   3. Our brains encode local objectives, like “if glucose low, increase hunger” or “if tissue damaged, increase pain”. What evolution never gets is a clean, moment-by-moment read-out of “overall inclusive fitness: 73%”. It’s more like a CEO who only ever hears how loudly each department head is yelling. Hunger, thirst, pain, fear, curiosity, loneliness, sexual desire - these are the department heads. Evolution can tweak which departments exist, what counts as a crisis for each, and how loudly they shout when things go wrong. It cannot install a single, transparent “fitness” dial in the brain and let the Planner read it off.
   4. Even if it were possible, switching to a “top-down” objective function would not be adaptive unless the Planner is already highly sophisticated - somewhere close to perfectly rational - enabling them to derive each action rationally from a “procreate genes” prompt. By that time, the bottom-up architecture would be highly effective, largely mirroring the effects of any top-down objective function and rendering the switch redundant.
4. If not a single objective function, what about multiple objective functions? Isn’t that roughly what the brain looks like anyway with the dopaminergic systems etc?
   1. Adding more objectives only pushes the problem back a step. As soon as any two of these “local utilities” conflict, the system needs some way to decide which one wins. That decision procedure could be:
      1. itself specified by another, higher-level objective function (back in the regress).
      2. grounded in something of intrinsic value.
   2. Either way, multiple objective functions don’t remove the need for a special kind of signal that can actually end the “why care?” regress.

3 - Valence

With no single, top-down objective function available, the only realistic option is to work with the body’s existing signals. We’ve established that a bare stimulus is not enough; an extra element is thus needed to break the “why care?” motivational regress.

This “extra element” is more of a design challenge than it might initially appear. A few pushes when the organism needs food or drink aren’t going to be sufficient. Evolution is going to need to create an entire motivational architecture, the net effect of which should be to approximate what practical reality has denied it: an objective function serving its interests.

Given this role, we can start to identify some core properties that these signals will need.

Property 1: Signals to the Planner should be “Non-rationally Compelling”. They must be made to act as an irreducible tug on the Planner’s motivation, rather than showing up as more data to be analysed. This means appearing as inherently positive or inherently negative to the Planner itself*,* defining what it treats as better or worse. In other words, these signals should be valenced.

To emphasise, this is not how a thermostat works, or a chess engine, or even AlphaZero. In those architectures, all impetus to action at runtime is derived from a pre-defined “top-down” objective (maintain setpoint; maximise win-rate). Internal signals are treated as data about how well that external objective is being met. By contrast, valenced signals are themselves the source of “better or worse for this system”. They are wired to intrude into the Planner’s scope of attention and reshape its priorities. The compulsion is graded: the more extreme the underlying bodily condition, the more forcefully the signal should present itself.

To be clear, I’m not claiming to have shown that consciousness itself is necessary (yet). This is purely a functional story so far. Although these valenced signals will eventually become the subjective feelings of pain, hunger, pleasure etc., all we have shown thus far is that a series of circuits in the brain must be pushed and pulled in different directions in response to certain stimuli.

That being said, it’s worthwhile considering what we do observe to illustrate the functional role that these signals play.

We observe:

• Feelings do force themselves upon us. We don’t decide whether severe pain matters - the system is built so that it “just does”. We cannot entirely shift our attention elsewhere and remove it from the Planner’s view.
• Signals do indeed motivate action in a graded way. Acute hunger, thirst and breathlessness all essentially force remedial action above all other considerations.

• Clinically, if valence breaks, motivation disappears. The best-known example of this is pain asymbolia, a condition where patients can sense pain, but its unpleasantness is absent or massively reduced. They’re typically aware that the sensations should be unpleasant, but they report the sensation as neutral or even detached.

  “The patient still feels the nociceptive aspect of pain but is not distressed by it and consequently not motivated to avoid it” Gerrans 2024

  “The pain is experienced as happening to my body but is not experienced as mine.” Gerrans 2020

  Sufferers can self-harm out of pure curiosity, and typically require close supervision due to risk of injury. Pain asymbolia is associated with lesions in areas of the brain involved in motivation and valence, particularly the anterior cingulate cortex and the insula.

Property 2: they should be Relevant for Planner decision-making**.** The regress logic shows that valence serves a purpose - it enables the Planner to make decisions. Only data that affects a current or future decision is worth escalating.

As noted, there are plenty of decisions that don’t require the Planner’s metabolically-expensive reasoning capabilities. The Planner is like a highly-salaried CEO - there’s no point in “bothering the boss” with irrelevant information.

Valenced signals could be relevant in two ways.

1. Current control: Overrideability
   The signals could affect some immediate decision where, at the Planner’s level of abstraction, there are genuine alternatives. In other words, the Planner could, at least sometimes, override the signal’s suggestion in favour of other goals.
2. Future control: Lesson learnt
   Alternatively, the signal can matter as a “lesson learnt”: a way of updating future behaviour, policies, or expectations (“don’t fall off the bike again”). Agony after the fact seems to live here - an extreme initial effect may not be practically overrideable, but the Planner can still learn from what happened.

What we observe:

• The vast majority of the brain’s regulatory activities operate without conscious knowledge. The brain regulates heart rate, white blood cell production, and liver enzyme levels without generating any specific valence. Just as a CEO doesn’t need to manage stationary purchases, the Planner does not need to manage insulin release.
• Planners do detect hunger signals because obtaining food requires complex, novel navigation of the environment - a task only Planners can execute. We feel breathlessness because, unlike heart rate, breathing is a semi-voluntary system where the Planner can intervene (holding one’s breath) and must be effectively forced to stop doing so when oxygen levels drop.
• Straightforward real-world decisions are taken before the Planner is engaged. Studies show that we remove our hand from a hot stove before the conscious mind registers the heat; the fight-or-flight mode is engaged before fear kicks in.

Property 3: Valenced signals should be Intelligible.

To be useful, these impulses can’t just be random jolts. They must be categorised and context-sensitive enough so that the Planner can treat them as intelligible data about the organism’s state: “this kind of bad usually means damage”, “this kind of bad means social risk”, “this pleasant anticipation means a likely future status reward”. Without this, they’d be noise rather than guidance.

Property 4: They should align interests with evolution.

As we’ve seen, evolution cannot govern our behaviour by encoding a “top-down” objective function in our brains. What it can do, though, is approximate an objective function, assembling it from the bottom up, using valence as its raw material.

It may be helpful to consider the Planner’s role. It has, in a sense, been “hired” by natural selection to manage an asset (the organism), to accomplish a goal (gene promulgation). This relationship becomes a version of the “principal-agent” problem. If an agent is going to do its job well, its interests should be aligned as closely as possible with the principal’s.

To do this, evolution uses another of the brain’s specialist functions - building internal representations. We use these every day to recognise recurring patterns like faces, objects and voices, which the Planner can use in reasoning and prediction.

Body Model

Since the Planner has some executive control over the organism, it is adaptive for it to reason in a body-integrated way: to co-ordinate legs, arms, eyes in response to threats and opportunities without internal conflict. It needs to bend its legs to duck from a missile that would hit its head, with no part of the system treating that as “someone else’s problem”. This requires a holistic, real-time map of the physical asset - a body model - which the Planner can incorporate into its reasoning.

It is well-known in cognitive neuroscience that the brain uses sensory inputs to construct a stable, bounded body symbol. We have circuitry specifically designed to integrate proprioception (where my limbs are), vestibular data (which way is up), and vision into a unified map.

A body model is necessary but not sufficient to solve the principal-agent problem. To close the loop properly, we need a self model.

Self model

Aligning interests is not only a present-tense project. If the Planner is going to discharge its duties well, it should be able to formulate long term plans. Thus valence, as the ultimate source of motivation, should encourage the Planner to build shelters, collect firewood for the winter, and sharpen flint tools for later use.

Future planning requires:

1. A persistent locus of control (“The organism will respond to instruction tomorrow as well”)
2. A persistent terminus of valence (“The organism getting damaged tomorrow will lead to badness”)
3. Memories of past valenced experiences (“Sabre-toothed tigers are to be avoided”)

In a “persistent terminus of valence”, we have a requirement for a stable model, available to the Planner’s circuitry, which matters over time - i.e. a basic self-model.

The theory of the self is well covered by the likes of Antonio Damasio, Thomas Metzinger and Anil Seth. Damasio describes the transition from the “proto-self” (essentially the body model) to a “core self” (what I’m calling the self-model) as the moment when “somatic markers” (valenced signals) are bound to that body model.

The self-model is created like a portrait being painted. Initially brushstrokes (valence tags) stick to the canvas (body). As more and more are added, an image (self) emerges; further brushstrokes affect both the image and the canvas.

The self-model must be useful for the Planner for natural selection to create it. Its “job” in this case is to solve the principal-agent problem. The most effective way of doing that is for the Planner to develop a direct identification with the self-model. The self, as a construct of valence, is evolution’s “person-on-the-inside”, sculpted by natural selection to represent its interests. If the Planner can be convinced that it is the self, then it will plan accordingly. Even the supposed interests of the organism, where they diverge from that of the self, will likely be subordinated. Good mating opportunities, for example, may often be worth the likely drain on the organism’s resources.

I’ve only sketched a fairly minimal self-model so far, but richer versions can be built on this foundation. Anil Seth identifies five “selves”, including a “narrative self” - the story that we tell ourselves built on our memories - and a “social self”, how we perceive others perceiving us. In reality of course, there are no bright lines separating any of these - they are useful lenses on what is in practice a messy, fluid set of processes. For our purposes, the key distinction is motivating valence - hence the non-valenced “body-model” and the valenced “self-model”.

It may help to revisit pain asymbolia to illustrate what is going on. Pain asymbolia isn’t the absence of the sensory mechanism for detecting bodily damage, it’s the absence (or massive reduction) of the affective quality of the pain, i.e. its valence.

Here’s the quote from earlier plus another from Philip Gerrans (Gerrans 2020)

“The pain is experienced as happening to my body but is not experienced as mine.”

“Pain asymbolics no longer assign or feel emotional significance in response to bodily damage... Consequently, they report that the experience is painful but that it does not matter and feels as if it is not happening to them. What this shows is that ‘mineness’ can be lost locally....”

Without valence, the self-model seems to disconnect. This is to be expected if it is, at base, constituted by valence.

To reiterate then, the self is evolution’s way of approximating an objective function. Its role is to complete the union of the Planner with its organism and impose evolution’s agenda. It enables evolution to say convincingly: “You and this body and this self are one thing. It matters therefore you matter. Its wants are your wants. You must act in its interests.”

4 - A Model of the World

So far I’ve focused on valence for the role it plays in resolving the motivational regress. But the Planner also needs data about the physical world to steer the organism effectively. This is where non-valenced sensory data enters the picture - the redness of an apple, the sound of a twig snapping, the smell of a possible food source.

Sensory data helps the Planner to do its job. Valence tells the system what’s better or worse; perception, along with memory, provides the raw material for its reasoning. A compulsion to eat is useless without visual information about where the food is, auditory cues about approaching predators, or memories of where berries grow. What the Planner needs is a world-model: a representation of its environment that it can use to simulate outcomes. The role of sensory input and memory is to construct that model.

Building the World-Model

Exactly why something like “redness” appears as what you or I know as redness is beyond my scope. What matters is that colour is a useful format for the Planner to differentiate wavelengths of light. Evolution could have played a different musical note in our heads when we detected these wavelengths, but that would be much less useful. Every shift in gaze would trigger a jumble of overlapping chords, the sounds blurring together rather than cleanly marking the surface boundaries. Colour, on the other hand, “sticks” to the surface of objects, like a paint that allows the Planner to track an object through its world-model.

Audio, visual and other sensory cues help it answer relevant questions quickly and easily - “where is that?”, “how is it moving?”, “can I reach it?” or “what happens if I push it?”.

Helping build Body and Self

Sensory content serves another purpose. As the default “camera” (vision) and “microphone” (hearing) are physically located in the organism’s head, they help to reinforce the unity of the body and self models. Their position confirms one geometric centre of perception - somewhere “behind the eyes”. The Planner is “looking out” from this point.

Further, if you tap the back of your hand with your finger, you’ll notice at least three senses which reinforce this coherent story of “one body, one self” - touch, sight and audio cues all help locate the action within the body. Valence can add motivation and meaning to this picture. If you pinch yourself hard instead, you’ll get evolution’s message: the body matters, it’s part of your “self”, look after it.

The Planner, then, can model a rich, narrative self, instantiated in a body, situated in a world. It can examine its “self” as it can examine any internal representation - it can categorise its properties, assess what it’s capable of or imagine it in detailed future end-states - it can “self-reflect”.

In this context, there’s selection pressure to develop a legitimate rationale for the original “non-rationally compelling” valence tags. A negative-valenced impulse in the stomach needs a story that is consistent with the more advanced incarnations of the self, as does the sensation that develops if the local water source dries up, or if the organism’s sexual rival appears.

5 - Cashing in the chips

I may have strayed a bit in the text, but we can still explain all the implications of the regress in purely functional language.

We could say: “The reasoning circuitry in our brain requires signals that are intelligible, non-rationally compelling, relevant for decision-making and bound to a self-model, within the context of its world-model.”

This is already fundamentally different from any artificial system we’ve yet created. From thermostats to self-driving cars, all still have “top-down” objective functions at runtime, negating the need for any valence or self-model.

Still, we can say all this and be firmly on the functional side of the philosophical Rubicon. On the other side lies the language of subjectivity and consciousness. Many in the physicalist Camp 1 will have happily swum across well before reading this article. Many in Camp 2 will deny that the river can ever be crossed. Here’s my attempted bridge.

Firstly, as argued in part 3, the Planner must identify with the self-model in order to solve the principal-agent problem. Together, they form a composite “Planner-Self” - a reasoning actor in the world-model with a unified locus of sensory and valenced content.

Next, the original regress showed that if nothing intrinsically “mattered”, the Planner would stall, as no overarching objective function is possible. For the Planner-Self, each “non-rationally compelling” valence tag would appear in the world model as a bedrock concern with a legitimate rationale. That is, it should intrinsically matter to it. Words like pain, love, and joy are the labels the Planner-Self uses to backfill a rationale for these mattering appearances.

Pain is to be avoided because it will affect it. Love and joy are to be sought because they will appear as good to it. It can imagine future events happening to it and classify them as worth seeking or worth avoiding. It lives in a richly-structured inner landscape of possible situations that could affect it. In this world, the role the Planner-Self plays is “me”. “Me” is the entity to which things matter.

Low glucose signals present as “my hunger*”, damage to the organism presents as “my pain”, social exclusion presents as “my loneliness”*. The Planner-Self will rehearse conversations, anticipate “pain”, crave “relief”, even reflect on the strangeness of its own condition. From its perspective, there is something-that-it-is-like to be in these states. It can feel them. If it couldn’t, they wouldn’t be performing their essential function of motivating the whole system into action. They must be “subjective”.

Each one of these subjective states has a different job to do. Evolution has many priorities for the organism, from eating, drinking and sleeping, to navigating the social world of friendships, romantic attachments and rivalries. We would expect our internal world to be a constant flux of feelings and imaginings relating to these goals, each playing its own role in keeping us broadly on evolution’s script. We would expect a rich narrative and social self, formed as we try to feel out where we fit in, what strategies are open to us, and who to befriend.

We would also expect to experience sensory content as a subject. Just as a bump to the head is interpreted as “my pain”, the redness of an apple is how it looks “to me”. While the relationship here may technically be “access consciousness” rather than valence-laden phenomenal consciousness (i.e. there may be little or no “motivational feel” of the colour red on its own) the Planner-Self is always the “client”. It is the centralised point of view for which all these percepts have been formatted.

The recognition of ourselves as the subject in this picture may still be an intuitive leap away. We all have the impression that we’re somehow more than just the sum of our parts, we’re a bona fide being, with moral worth, perhaps even a soul. Personally, if I really try to introspect, although it’s not easy to see through what Anil Seth calls “the controlled hallucination”, I see nothing which can’t be traced back to the models that valence has built.

David Hume got there first. In 1739, he wrote:

“For my part, when I enter most intimately into what I call myself, I always stumble on some particular perception or other, of heat or cold, light or shade, love or hatred, pain or pleasure. I never can catch myself at any time without a perception, and never can observe any thing but the perception.”

While Hume might not have covered everything here, we do experience this flux of feelings, thoughts and emotions. We do imagine scenarios and create plans. We do care about our “selves” over time, more-or-less in alignment with evolution’s interests. This is the “something-that-it’s-like” to be human - our version of consciousness.

For the “why” part of the hard problem to be coherent, we have to posit “something that it’s like” over-and-above this - i.e. over-and-above what it’s actually like. If you think that this something does exist, let’s discuss what it is or what it does.

A Word on Zombies

The idea of a “Philosophical Zombie” - a human with all of its faculties intact, yet without any subjective experience - is often used to illustrate the hard problem. If “p-zombies” can exist without consciousness, then consciousness would be an epiphenomenon - an unnecessary, non-functional add-on.

I’ve argued that, far from being epiphenomenal, consciousness has an essential role to play in the functioning of the brain. Without what we know as subjective feelings, humans would have no motivational structure.

There is some clinical evidence pointing in this direction. In akinetic mutism, patients are awake, can track stimuli, and often retain language and basic comprehension, but they don’t tend to speak or move; they lie or sit motionless unless strongly prompted. When they do later describe their state (if recovery allows), they often report a kind of extreme apathy or indifference rather than paralysis or confusion. Neurologically, akinetic mutism is typically associated with damage to medial frontal and anterior cingulate regions - areas heavily involved in motivation and initiation. Notably, unconsciously motivated behaviour is not affected. Patients breathe, regulate their own temperature and heart rate. Their stimulus-response circuitry is working fine; it's their valence that is broken.

According to reports:

“Patients are in a wakeful state of profound apathy, seemingly indifferent to pain, thirst, or hunger.” (Arnts et al. 2020)

“She didn’t talk because she had ‘nothing to say’. Her mind was ‘empty’. Nothing ‘mattered’. She apparently was able to follow . . . conversations even during the early period of the illness, but felt no ‘will’ to reply.” (Holroyd & Yeung 2012)

Cases like this don’t prove the argument, but they are consistent with what we would expect if valence plays an essential role in motivation.

Still, if the argument does hold, zombies aren’t a well-defined biological or physical possibility for our world - in taking away subjectivity you’re removing the very thing that animates us.

The zombie argument often sets the bar at “conceivability”. As Chalmers explains: “at
minimum, we can say that the conceivability of S requires that the truth of S cannot be ruled out a priori.” I believe that, with enough philosophical window-dressing, the regress logic could take a meaningful swipe at “conceivable”, but any attempt to do so now would be a distraction until the argument has been through proper quality control.

P-zombies will of course remain prima facie “imaginable”, but this really is a low bar to clear. I can imagine a pig without any internal organs - this alone doesn’t suggest any metaphysics.

Where does this leave us?

The argument I’ve presented concerns the Why Question. There of course remains the How Question - how exactly is conscious experience generated by physical processes in the brain. The exact neurology is beyond the scope of this essay (and well beyond my capabilities), but there are some implications worth mentioning.

Although the argument is strictly neutral on whether consciousness has to be nothing but physical, it does leave less space for non-physicalists to occupy. The key cleavage here seems to be whether you concede that it’s possible to create consciousness physically. If you accept that it probably is, then the argument suggests that evolution has indeed found a physical solution to the problem of motivation. Non-physical solutions are not parsimonious - no evidence has yet been found of any other such solution in biology.

Non-physicalists can still claim impossibility, and many do. This isn’t the place to tackle these claims head-on, but it now seems a long-odds bet.

1. Firstly, I recognise and share the intuition that there’s something “extra extra special” about consciousness. We all feel this, and there’s a reason for it: there is direct selection pressure on that intuition being strong. The body-model, the self-model and the world-model all conspire to reinforce it. Some people may have followed along happily up until this point, but still be tempted to say something like “OK, I get that a system like that would need to feel something from the inside. I understand why it might need to love, hope and see the colour orange. But that’s not me. I’m something different”. Again, I feel this too. It shows that we’ve all been effectively house-trained.
2. As many have pointed out, we would do well to remember how fallible our past intuitions have been. The history of claiming “something extra” is littered with now-defunct ideas. An élan vital for life, caloric or phlogiston for heat and combustion, a luminiferous ether for light - all have eventually been replaced by detailed physical accounts.
3. Many of those ideas were posited without a good reason to believe that an expert physical designer (evolution) had a structural requirement to devise a physical solution. The presence of a structural requirement, even in the absence of the detailed physical account, should give some pause for thought.
4. Growing up surrounded by life, it’s easy for us to take its complexity for granted. Perhaps it’s worth taking a moment to reflect on evolution’s design achievements. Could we, with all our technological hubris, build a mosquito? Could we get anywhere near creating a self-replicating cell from scratch? Let alone something as infinitely intricate as a squirrel? Now consider that evolution has done far more than this with no refined metals, no industrial heat, no downtime for maintenance, no clean slate to restart and no foresight to design holistically. Evolution is a staggeringly capable engineer. We bet against it at our own risk.

6 - Conclusion

Consciousness is, perhaps, the greatest trick that evolution ever pulled. Stripped of the ability to control our actions directly, it developed a series of pushes and pulls to keep us broadly on script. As I’ve argued, in practice these must cash out as “ours” - subjective experiences over which we feel ownership.

There’s a real point of contact here with Camp 2 hard problem absolutists. These subjective experiences are very real - I agree that there’s “something missing” from the textbook functionalist - certainly the illusionist - story. My point of difference is that I believe that this “something” also bears a functional load.

As with any physicalist theory, there’s the inevitable sense that some of life’s magic is lost with each reductionist step - like watching a fairy being dissected on a lab bench. I accept this. But for me at least, there’s a certain magic here too.

We tend to view our motivational urges as a nuisance, a primitive system trying to force our hand. There’s a sense in which the truth is precisely the opposite. These signals become conscious not to overrule our rational mind, but to recruit it. Emotions demand our attention only because our bodies face problems they cannot solve alone. Our feelings are not trying to constrain us - they’re crying out for help. If they never needed that help, “we” wouldn’t be here in the first place. We have the privilege of exploring and understanding the world only by virtue of our talents in helping navigate an organism through the evolutionary environment. What we do with that privilege is up to us.

There’s a lot more ground to explore if all this survives scrutiny. Even if the hard problem discussion doesn’t move an inch, the argument’s claim about the role of valence leads to predictions about how the mind is constructed, and there’s a raft of empirical evidence to sift through. I’ve only scratched the surface here, partly to avoid trying anyone’s patience, but largely to invite comment on the argument itself without getting too bogged down in clinical cases. There are also implications to explore, not least for AI, for how we think about animal consciousness and for how we manage our own minds.

Taxonomising Objections

It would be amazing to have some feedback. To make things easier, it may make sense to distinguish some subcamps. Here’s a list to get things started, but feel free to subdivide or invent new categories.

Physicalists
Camp 1A: You may have a few niggles, but you broadly accept the argument.

Camp 1B: You think I’m overreaching, or I’ve gone off the rails at some point, but you’re still a physicalist.

Non-physicalists
Camp 2A: You reject the evolutionary logic of the argument. You don't accept that, for human-like brains, natural selection would develop non-rationally compelling, intelligible, relevant impulses that are bound to a self-model with which the Planner identifies.

Camp 2B: You accept the evolutionary logic but deny that subjectivity or consciousness follows. You think that a system could have all these functional properties without subjective experience.

Camp 2C: You accept that the argument does imply some form of subjectivity for these beings, but maintain there's “something extra” about our own consciousness that remains unexplained.

Over to you, and thank you for reading!