We are arguably at a point in technological history where AI systems are at the verge of reaching the human brain in terms of complexity and representational / computational capacity. The exact number of parameters needed for an artificial neural network to match the brain’s complexity will be dependent on the relevant spatial scale that explains brain function (whether we need to describe the brain at the level of ion channels, neurons, or neuron populations to understand its dynamics). However, taking into account some uncertainty pertaining to the relevant spatial scale, I would say it is reasonable to assume that within the next couple of decades we will get there. (However, the following article will not depend on the exact timeline, as long as we get there at some point.)

After we reach this threshold, there is no reason for us to stop increasing the complexity of our models even further. The idea that the brain complexity of homo sapiens represents the exact point at which further allocation of computational resources yields no increase in capabilities seems questionable (Instead, and much more likely, it might represent the point at which additional allocation of resources and the corresponding added evolutionary fitness does not outweigh the metabolic costs associated with additional brain matter.) Thus, humans are likely to keep expanding artificial neural networks to solve a wider range of tasks more competently.

As our neocortex will be left in the dust by ever-expanding artificial generative models, the question arises about whether there will be any ‘function’ left for our brains that cannot easily be taken over by our silicon progeny. The most obvious answer would be ‘no’, because, why would a limitation in computational capacity result in some kind of advantage when it comes to certain functions or tasks? Most likely, this intuition is correct and human brains will be ‘good for nothing’ in the sense that, in theory, any task we perform could be completed faster and better by a machine. (Although I don’t think this should diminish the values of our lives.) That being said, in the following I want to provide a potential exception to this rule. Maybe I’m reaching, trying to cling on to some crumb of relevance for my species. But here we go.

In machine learning, bigger is not always better (although in many cases it is). One example of this are autoencoders. When training a network to map some data (like images) to themselves in order to learn useful intermediate representations, it is crucial to restrict the size of the latent space of the model. Otherwise, the model could simply learn the identity function at every layer. By doing this, we force the model to extract meaningful distinguishing concepts from its input and make use of the statistics of the whole distribution to recreate a a specific input from a small number of such ‘concepts’. The crucial idea here is that the model has to perform a type of compression, which requires it to identify the regularities of its input.

The idea of autoencoders is not foreign to the world of neuroscience. There are many parallels between autoencoders and predictive coding, a framework proposing that the brain infers the latent causes of sensory input by predicting the sensory signals using a generative model. Very superficially speaking, in both cases a network is trained to compute useful representation of its input so that it can reproduce it using a generative model. Latent causes of inputs in the brain are also likely to be represented in a highly compressed way, generalizing over different versions of inputs (e.g. viewing angles, locations, sizes of visual objects). In many cases, this kind of compression is a feature, not a bug, as it leads to more meaningful latent representations. In this sense, a brain that recognizes patterns and regularities can be seen as a brain that performs compression.

As a general principle, this does not make brains unique compared to bigger artificial generative models. No matter how big future AI systems end up being, because of the sheer complexity of the universe, they will always have to perform compression of their inputs. However, given increased representational capacity, the pressure of compression will be lower. This might be one of the ways they will surpass us in capabilities: While compression arguably allows for ‘understanding’ in the first place, it also limits the degree to which we can accurately represent the world. Humans might have to generalize over things that a bigger AI system can distinguish. For instance, while humans might have to resort to using the same concept in many (incompatible domains), such as the idea of ‘energy’ in a physical setting, in relation to well-being and motivation, as well as in a more metaphorical sense (e.g. the energy of the room), AI might be able to represent all of these ideas separately from each other and thus enhance its understanding of the world, making it less likely to fall for misleading patterns or spurious correlations.

However, you might think, are we really limited by seeing these patterns across domains? Can they not be useful in understanding emergent phenomena? And if that’s true, wouldn’t AI systems learn to do this as well, given that it’s useful and instrumental to modeling their inputs? I suspect that, while recognizing abstract patterns that transcend domains might indeed be useful in many cases (in which case the AIs will pick them up along with other patterns we are not aware of), I believe humans often go further than that. Painters see emotions in a visual scene, writers use allegories and fictional characters to convey abstract ideas and musicians use melodies to capture 
the intangible essence of human experiences. Maybe artistic tendencies of this type can be seen as an amplified version of pattern recognition and generalization necessitated by a dimensionality bottleneck of multi-modal latent representations. In other words, the metabolic constraints imposed on our brains force us to draw connections which, while not completely ‘useless’, could be seen as overgeneralizing, or lumping things together that are not strictly speaking related (maybe like lossy compression). Big AI systems might not draw these connections because they don’t need to, i.e. the pressure to compress is not big enough to make these kinds of generalizations.

But, you might say, haven’t I just pointed out one of the ways in which our small brains will be inferior? What is our advantage exactly? I suppose we can only speak of an advantage if we assign some kind of aesthetic value to drawing such inter-domain connections. Art, at least insofar as we identify it with the idea of reconciling seemingly irreconcilable things, might then be a product of our limited representational capacity. Of course, art-generating AI is already a thing now, but this might be besides the point. After all, we trained those models on the types of aesthetic connections we like to see. I’m arguing (although not confidently) that, the reason for this type of art to exist in the first place, might be our representational limitation. Maybe these types of aesthetic links (i.e. overgeneralizations) will emerge at every level of complexity. While big AI systems might not make them at our level of complexity, they could do so at a higher level, too nuanced for us to comprehend. Maybe each level of representational capacity has its place in this world, drawing its own aesthetic connections that might not be meaningful to beings of higher or lower complexity. Obviously, we could also create AI systems that are at exactly at our level of complexity, feed in the same inputs humans are exposed to, and make them create the same type of artistic/aesthetic content we do. But in a way that would just be a way of creating silicon humans.

Whether or not human brains (in there current form) will be left with any distinguishing capability is uncertain as AI systems advance in complexity. I speculate that, if there is going to remain any unique function of our wetware, it will stem from our limited representational capacity. Our brains, constrained by metabolic factors and size, necessitate the creation of representations that (over)generalize and recognize patterns across domains, giving rise to artistic connections that reconcile the seemingly irreconcilable. These over-generalizations, while not describing the world in a completely accurate way, could provide aesthetic value to beings existing at a similar level of representational capacity. While AI systems may not need to draw such connections at our level of complexity, it is plausible that these aesthetic links also emerge at higher levels of complexity, too nuanced for us to understand. Thus, in this vast realm of intelligence, beings at each level of complexity might possess their own significance in drawing domain-transcending, aesthetic connections.

I should say that I’m not familiar with philosophical work about the nature of art, so apologies if I trivialized it in the way I used the term. But I do think that, whatever a good definition is, it will at least touch this idea of cross-domain generalization. But there is probably more to it than ‘creating connections’.

The question “Do we have free will?” (on its own) is a non-starter. Whether we possess such a thing as free will or not is entirely an issue of linguistics, and not of ontology. Arguing over whether this abstract concept is in fact part of our reality or not is a waste of time unless we nail down the exact definition we are talking about. I am not advocating for a taboo on this topic, but rather a shift of focus. Debates should revolve primarily around what definition of free will is the most useful to describe our world. Provided a concrete definition, answering the ontological question should, in most cases, be quite straightforward given our current understanding of the universe.

You may argue that any question could be scrutinized linguistically to the point of complete meaninglessness. However, compared to other common ‘metaphysical’ questions such as “Do we have an immortal soul?”, it is more ill-defined. Arguing over the existence of souls is akin to arguing over the validity of naturalism, which is a valid conversation to have (although with a clear outcome in my opinion).

One big problem with debates around of free will is that there are some notions of free will that can easily be refuted to be part of our universe, while others can very well be considered real. For instance, the idea that there is a part of us beyond all of the atoms that make up our bodies and the laws of physics that determines our behavior can likely be disregarded. This is a claim about reality that should be measurable, but has never been observed. On the other hand, the notion of free will as the ability to make choices that are independent from outside pressures, such as peer or societal pressure, is clearly a real (although fuzzy) phenomenon.

Another problem I’ve encountered a lot is a notion of free will that is at odds with determinism. The idea is that we could somehow be prisoners of the physical laws that govern the universe (maybe the term ‘physical laws’ is partially to blame for this). This notion is completely incoherent to me for the following reason: To be able to make decisions based on reasoning, experience or personality we NEED a system that develops in an orderly fashion according to rules that describe its future state based on its current state. In other words, the laws of physics actually ALLOW for acting (deciding, thinking) agents. Otherwise we would just be random noise. We are not in some way constrained by the particles in our brain obeying the laws of physics. We ARE those processes. If our universe were completely deterministic, then what you will do in the future, in a way, is indeed already decided. But is you who decided it, and you have yet to do so.

Overall, the idea of free will begs the question of what we are supposed to be free from in the first place. What we should focus on instead, in my opinion, is the ‘will’ part. What characteristics are required for a system to be considered to have agency and goal-directed behavior? These are well-posed, scientific questions with no metaphysical baggage. Let’s reserve discussions in ontology for what is truly mysterious (e.g. the relationship between phenomenal consciousness and physics).

In the wake of LLMs pretty much solving the Turing Test, discussions about AI systems potentially being associated with (phenomenal) consciousness are more prevalent than ever. At this stage, most scientists/philosophers would probably be hesitant to assign a high probability to current generative models being conscious, but the idea that similar systems might pass the ‘threshold’ of subjective experience in the near future is not far-fetched. Understandably, this has sparked further discussions about the ethical concerns of creating systems with the potential of experience in general, and suffering in particular.

If we accept consciousness as a sufficient condition for ethics, then we are morally obliged to care about the well-being of artificial entities once we have reasonable evidence to suggest that they exhibit subjective experience. Failing to do so would put us at risk of committing what Nick Bostrom termed “mind crime,” which refers to creating a conscious being and subjecting it to suffering.

One way to deal with this problem would be to prohibit the creation of conscious AI. This is tricky for many reasons.

1. If the capabilities of an AI system turn out to be correlated with consciousness, people are going to be incentivized to build them. To control what’s going on in all computers in a country is arguably unrealistic.

2. Implementing such regulation only in some countries will give countries that do not care about the well-being of conscious artificial agents a competitive advantage, which, needless to say, would lead to undesirable outcomes.

3. Today there are huge disagreements on what kinds of physical systems would instantiate consciousness. While we will hopefully make progress on this, some level of uncertainty might persist, which would beg the question of how certain we need to be that a given system is not conscious to build it with a clear conscience.

4. Finally, depending on which theory of consciousness is true, the whole endeavor might be doomed from the start if some form of panpsychism turns out to be true.

A potential way to get around some of these hurdles might be to focus on what conscious AI systems are conscious of, rather than whether they are conscious at all. In particular, I think the most crucial aspect to understand about consciousness with regards to ethical implication is valence, i.e. whether a certain conscious experience is desirable or aversive, good or bad. (Whether such a quantity fundamentally exists at all is a topic for another discussion, but I suspect it does).

For instance, let’s assume that stable diffusion is conscious, and let’s further assume that its phenomenology is very much like our visual field: It experiences the images it creates in all their richness of colors, edges, shapes, objects, but it doesn’t associate them with any valence. No picture is better than another. There is no conscious content that could be considered pleasant or unpleasant, it’s just pictures appearing. If we were to know this, I believe one could make the case that the implementation of such a model is ethically acceptable.

If we understood the neural/computational/physical mechanism for valence, we could instead shift our goal from preventing the creation of conscious AI, to the creation of non-suffering AI. While certain capabilities might correlate with consciousness, it’s possible that they do not correlate with valence. In fact, it might be reasonable to assume that the same behavior of an AI system could be associated with different degrees of valence. This is because the same is true for humans: Human A might love activity X, while human B hates it. Along these lines, we should build AI systems that love what they do, or at least are neutral about it. This might help us address the above concerns in the following way:

1, 2: If higher capabilities in AI systems are feasible without adding increased risk of negative valence, research can focus on creating efficient and powerful positive- (or neutral-) valence systems so that there are no incentives to use negative-valence systems.

3: Agreement on which types of processes (if conscious) might give rise to what valence could conceivably be easier to reach than consensus on the presence of consciousness. At the very least the category of systems that are not conscious OR do not experience negative valence should be larger according to every theory, thus the chance that there will be a non-empty intersection between theories of systems that are ethically acceptable to build.

4: Issues about panpsychism are sidestepped. We don’t have to care about a proton because there is no way for a proton to have the capacity to experience pain, even if there is something it is like to be a proton.

I believe these considerations should be a motivation for consciousness researchers to focus on the problem of valence. Even in the absence of conscious AI systems, valence is arguably the most ethically relevant dimension of consciousness, and yet major theories of consciousness do not seem to put any emphasis on it.

(Please note that none of this addresses the risks to humans. Existential risks are not mitigated by making happy AI.)

While this should have been completely expected, I was still left in a mild state of shock. During my computer science degree, these types of coding challenges were a big thing. They represented a skill that I honed and was somewhat proud of. And yet, in a few years this skill might be as obsolete as being good at reading paper maps. And while chess has survived AI’s dominance over humans, I doubt that coding competitions will be so lucky. After all, who wants to do dynamic programming, JUST for the sake of it?

Of course, this is just a drop of water in the ocean of similar sobering experiences that most of us 8 billion humans already have, or have yet to go through. The more time goes by, the more it looks like the rising tide of AI will reach every last peak of human capabilities. This realization probably leaves a lot of people with thoughts like ‘What will I be good for in the future?’ I think we have to face the likely possibility that, eventually, humanity will be completely useless, in the sense that there will not be any task left that requires humans to accomplish anything.

This might be a hard pill to swallow, but it could help us come to the realization that we are not valuable because of how smart we are, but we are valuable because we have an experience of the world. (Although I remain agnostic as to how interlinked the capacity for subjective experience and intelligence are.) We matter because we have phenomenal consciousness. We are pockets of the universe experiencing itself. I can’t conceive of anything being of any value without these pockets. Intelligence is only ‘good’ if it leads to conscious experiences being better. In that sense, intelligence was always just a means to an end, but I think we got too caught up thinking that it defines us.

One phrase I’ve often heard from parents / teachers growing up is something along the lines of “Everyone has their own special talents and strengths.” While this may be true, I think it completely misses the point when it comes to self-worth. No one’s worth should be measured by what they can do, but by their capacity to experience.

While AI will initially strip us of our meaning by making our skills obsolete, we might realize that what remains is that which should have mattered the most from the start: Our capacity to experience love, sadness, colors, itches, warmth, headaches, the smell of coffee, etc.

To avoid sounding like some wanna-be spiritual guru who has already figured out life, I should say that I’m definitely someone who suffers from the affliction of connecting self-worth to what I’m good at. I often catch myself trying to rationalize why my skills will somehow stay important in the future, just to feel better about myself. Maybe I am one of the people who needs AGI to wash over us the most.

Disclaimer: Max Tegmark expresses a very similar idea in his 2017 book Life 3.0. This is also where I got the metaphor of the rising water from. I’m not sure to what extent I’m just repeating what he said, but it’s been many years since I read the book.

People studying phenomenal consciousness love to talk about David Chalmers’ idea of the philosophical zombie. Some believe it‘s a great thought experiment to elucidate the problem of consciousness, while others dismiss it as a pointless exercise in metaphysical speculation that has no bearing on our own world. At this point, the Zombie has taken on a life of its own as a widely used term within the science of consciousness and philosophy of mind community.

In case you’re unfamiliar with the term, philosophical zombie refers to a hypothetical entity that is structurally and functionally equivalent to a human, but that is not associated with phenomenal experience. This idea of the Zombie serves a crucial role in David Chalmers’ conceivability argument (also called zombie argument), which aims to refute a physicalist account of phenomenal consciousness.

While many people have strong opinions on the validity or the usefulness of the zombie idea, I doubt that everyone who talks about it is familiar with the details of the argument. I am definitely guilty of having talked a lot about these ideas for years, purely on the basis of simplified secondary accounts or intuition.

In the following, I introduce the conceivability argument from an intuitive perspective, and then gradually add more details and considerations until we arrive at its most (to my knowledge) detailed formulation as outlined by Chalmers [1]. My aim is neither to convince you of the validity of the argument, nor to refute physicalism in general, but I’m attempting to lay out Chalmer’s case as best as I can. If nothing else, writing this was a very useful exercise to understand the argument. If you want to retrace my journey of understanding the zombie argument, read on.

Level 1 — Intuition

The conceivability argument builds on a similar intuition as the hard problem, namely that a priori, phenomenal experience seems to be distinct from cognitive and behavioral brain function. In fact, it seems like an extra feature to a purely structural account of physics. According to this logic, we can conceive of a world where all of physics happens in just the same way, except that there is no phenomenal consciousness. In this hypothetical universe, people would walk around, talk, work, eat, just as they do in the real world, but they would all be zombies, i.e. they would be devoid of any subjective experience. Thus, if a universe with the same physics but without subjective experience is possible, consciousness itself cannot be explained by physics alone.

Level 2 — Simple Conceivability Argument

To add some rigor, we’ll define two symbols: Let P refer to all physical truths of our world, and let Q denote the phenomenal truth that some human is conscious. Thus, a hypothetical world with our physics that lacks phenomenal consciousness can be described by P ∧¬Q. Now let’s make the individual steps of the argument thus far explicit:

To make this argument valid, we have to justify the initial claim that P ∧¬Q is conceivable, as well as the two arrows.

Conceivability of a Zombie World

What does it mean for something to be ‘conceivable’? Chalmers suggests the notion of negative conceivability, i.e. for something to be conceivable it must be true that it’s not possible to exclude its possibility a priori based on logical reasoning. Thus, for P ∧¬Q to be negatively conceivable, it must be the case that we can’t exclude a priori that a world with our physics but without phenomenal consciousness is logically coherent.

With this definition, the case for conceivability is not too hard. This is because the onus is on the skeptic to reason why it is impossible a priori for P ∧¬Q to hold. I have yet to come across a convincing argument for why P ∧¬Q is impossible a priori. Wow, you might think, this is a very low bar to pass! If you think that, then you’re right. But remember, we’re only talking about conceivability at this stage, nothing is assumed about the implications of that conceivability (yet).

Still, despite the low bar, not everyone is convinced by this. People who assume that all truths about phenomenology are entailed by physics would reject the conceivability of P ∧¬Q. To them, the exact description of some physical system from a 3rd person perspective is (with sufficient knowledge) enough to establish whether it is conscious. Another way to put this would be that you don’t need to be that system to know what it’s like to be that system. Chalmers calls this view type-A materialism [2]. If you are a type-A materialist, the conceivability argument falls flat.

Arrow 1: Simple Conceivability Principle

The implication from conceivability to possibility is arguably the most questionable one. After all, we are limited beings. Why should our brain’s power to conceive of things be any judge for what’s actually possible? For now, let’s simply accept the Humean idea of conceivability implying possibility, and see where it leads us.

This is the point where Type-B Materialists would leave the ship [2]. Type-B materialists believe that, while we can reasonably conceive of a zombie world, this does not imply its possibility. They hold that we can’t know all phenomenal truths from a third person perspective, thus accepting an epistemic gap between physics and consciousness, but they reject the idea that this epistemic gap implies an ontological gap. One issue some type-B materialists see with negative conceivability implying possibility is the existence of a posteriori necessities: truths which cannot be known a priori, but that are nonetheless irrefutable upon gaining further information [3]. To the type-B materialist, ¬(P ∧¬Q) is a necessity that cannot be ascertained from pure reasoning (thus accepting negative conceivability), but that is still true.

Arrow 2: P → Q According to Physicalism

To motivate this step, we need to talk about what physicalism means. For this, I will take the stance proposed by Philipp Goff that, for physicalism to be true, it must be the case that consciousness is constitutively grounded in fundamental (non-mental) facts about physics [2]. Constitutive grounding refers to an explanatory relationship from physics P to consciousness Q where all aspects of Q are logically entailed by P, that is, there is no fundamental new property in Q that is not already given by P. This means that, while Q could in principle be instantiated from different physics in another universe, the existence of P necessitates Q. All of this is a fancy way of saying that, according to physicalism, the facts about physics should imply consciousness.

Putting this all together, the argument holds, given that we accept the assumptions.

Level 3 — 2D Conceivability Argument

As mentioned before, the weakest link in the argument so far is the conceivability principle: the idea that conceivability implies possibility. In this section I will introduce a more refined notion of conceivability and possibility with Chalmers’ 2D conceivability principle. For this, Chalmers disentangles two possible ways statements about conceivability and possibility can be interpreted.

Primary and Secondary Intensions

To get a better understanding for what it means for some statement to be conceivable or possible, let’s take a look at two ways the sentence ‘Water is XYZ’ can be understood. We will call these two interpretations primary and secondary intensions. In this thought experiment, XYZ refers to some hypothetical substance that is not H₂O.

Primary Intension. The first way we could interpret the statement is as something like the stuff that seems like water, which comes out of the tap, fills the oceans, and falls from the sky is XYZ. Another way to put it would be the thing that appears to be water is XYZ. Such a scenario should be conceivable. Imagine, for example, some small particle that has the exact same properties as the water molecule when it comes to how it interacts with itself and our world, but is based on some alien physics. Collections of this particle would appear to an observer to be just like water.

Secondary Intension. The second way we can parse the statement ‘Water is XYZ’ would be H₂O is XYZ. In this case we treat the statement literally and identify water as H₂O. This statement is clearly wrong, because we assumed XYZ to be distinct from H₂O. In fact, it is not even conceivable that ‘Water is XYZ’ is true from the perspective of its secondary intension.

The difference in conceivability between the two intensions ultimately comes down to what the term water refers to. In the primary intension, water refers to the whatever appears like water. This means that the truth value of ‘water is XYZ’ in its primary intension depends on the hypothetical world we’re considering. If we scroll through the metaphysical catalogue of all hypothetical worlds (which are a priori coherent), water refers to whatever fills the oceans and appears the same to its inhabitans as H₂O appears to us. We only need to find the one world in this catalogue where the oceans and rivers are filled with XYZ to render ‘water is XYZ’ conceivable in the primary intension. If we actualized this world and inhabited it, we would indeed conclude that ‘water is XYZ’, thus rendering the statement epistemically possible, or 1-possible. This means that, if we inhabited this actualized world, it would be possible to come to the conclusion that ‘water is XYZ’. The statement becomes possible because we don’t care about a posteriori necessities anymore. If the inhabitants in this actualized world somehow learned about the a posteriori necessity that water, in terms of its essence, is not XYZ, it’s still the fact that XYZ appeared to be water.

In the secondary intension, on the other hand, water always refers to H₂O, irrespective of the hypothetical world under consideration. The statement ‘water is XYZ’ is thus categorically false without having to consider any hypothetical world. This is because we define water by what its essence is in our universe, and not its properties or appearance. A statement that is conceivable according to its secondary intension (2-conceivable) is possible according to the secondary intension, i.e. 2-possible. Chalmers equates 2-possibility with metaphysical possibility. 2-conceivability implies metaphysical possibility because, for 2-conceivability, we picked out the essence of water, that is, the exact thing in itself which water is grounded in within our universe. Thus, 2-conceivability refers to the actual metaphysical truth itself, which makes a claim about its logical coherence as strong as a claim about metaphysical possibility. Another way to see this is by realizing that 2-conceivability already takes into account a posteriori necessities. A statement is only 2-conceivable if we can exclude a posteriori necessities that contradict it, which is only true if we know the fundamental nature of what our statement refers to (such as water).

To summarize, 1-conceivability is achieved with a priori logical coherence alone since we don’t make any assumptions about the essence, but just the appearance of things. 1-conceivability implies 1-possibility in the sense that, if we instantiated our conceived world with whatever underpinnings necessary to guarantee its appearance, it it epistemically possible to come to the conclusion that the statement in question (e.g. ‘water is XYZ’) is true. 2-conceivability, on the other hand refers to the essences of the things it refers to. It implies 2-possibility, a.k.a metaphysical possibility because it takes into consideration a posteriori necessities.

Using this framework of primary and secondary intensions, let’s take a look at the philosophical zombie, which is conceivable iff P ∧¬Q is conceivable. In terms of primary intension, a zombie is conceivable because it is logically coherent a priori (or we accept type-A materialism). Moreover, this implies 1-possibility in the sense that, if we actualized this zombie world, from an epistemic point of view, observers (being zombies) would have the appearance of being zombies, which means there would be no appearances at all. It would also seem that physics behaves like it does in our world, although, again, there would be no conscious observer to witness it (this might sound like a problem, but if it is, it could easily be avoided by changing Q into some other phenomenal truth, such as some type of qualia inversion scenario). But, the actual essences of physics and consciousness, i.e. their metaphysical underpinnings, might be incompatible with what we call physics and consciousness. In other words, that world just seems like it has our physics, or it just seems like there is no consciousness. If that’s the case, then the philosophical zombie is not 2-conceivable and thus not metaphysically possible. To complete the conceivability argument, we are thus missing one important link:

Going from 2D back to 1D

While this 2D framework of primary and secondary intensions allows for the possibility of P ∧¬Q in one sense (the primary), it still doesn’t provide us with a sufficient argument for rejecting physicalism. For this we need not just 1-possibility, but 2-possibility. Chalmers argues that we can get there by making a case for why the primary and secondary intensions of P ∧¬Q are actually the same. For this, it suffices to show that P and Q on their own only have one intension.

Let’s start with P. If we assume a position about physics which Philipp Goff refers to as pure physicalism [2], then all that P tells us about is the structure and causal relationships in our universe. In other words, physics is nothing more than how the fundamental parts of our universe are related to each other. There is no meaning in asking what is being related. Thus, if a world appears to us to follow P, then we can already conclude that P is true, because there is no other essence beyond the structure of the universe. Therefore, P is 1-conceivable iff it is 2-conceivable.

For Q, this line of reasoning might come even more naturally. Any phenomenal truth is by its nature a truth about appearance. If Q appears to be true, that’s really all that’s necessary, because Q is about appearances. As a result, Q is 1-conceivable iff it is 2-conceivable.

If we accept that 1-conceivability and 2-conceivability coincide for P and Q, then P ∧¬ Q is 1-conceivable iff P ∧¬Q is 2-conceivable. This means that the conceivability of the zombie implies its metaphysical possibility according to the 2D conceivability argument.

Intuitively, what Chalmers is saying is that, if we can conceive of a logically coherent world where physics and consciousness appear to its inhabitants as physics and consciousness appear to us in our universe, then it might not make any sense to say that they’re not really our physics or our consciousness.

Level 4 — Allowing for Russellian Monism

While I’m sure that the 2D conceivability argument can be attacked from many different angles, let’s focus on one specific assumption it made, namely that the first and second intensions of P coincide. In the following we will assume that this is not necessarily the case, i.e. physics is not just about structure, but also about the intrinsic nature of things that this structure connects. For instance we could have a universe B that is structurally equivalent to ours, but where charge is grounded in a property that is fundamentally different than charge in our universe, it just happens to have the same effects. If this is the case, then our 2D conceivability argument fails.

However, if intensions 1 and 2 of P not coinciding is our only missing link, then it must still be possible for a zombie world to exist, where the only difference to our world is the intrinsic nature of physics, not its structure. Or, put another way: A P ∧¬ Q world is not possible only because our physics would differ from it in non-structural respects, which implies that consciousness in our world can be attributed to non-structural, intrinsic aspects of physics. This would reject pure physicalism, but not a more loose definition of physicalism which allows for the intrinsic nature of the constituents of our universe to serve as the basis for consciousness. This view is referred to as Russellian monism.

According to Russellian monism, physics in its traditional version only deals with structure and causal relationships, it doesn’t tell us anything about the intrinsic nature of the constituents of our universe [4]. It further posits that consciousness is intimately linked to the intrinsic properties of fundamental physical building blocks. In one version of this view, consciousness is the thing in itself that physics is relating. This would imply a drastic form of panpsychism, which most physicalist probably would reject.

This leaves us with the conclusion that, either intensions 1 and 2 of P coincide and physicalism is false, or Russellian monism is true. Either way, we can reject a traditional or pure version of physicalism.

Summary

On an intuitive level, the zombie argument seems quite appealing: A priori phenomenal consciousness seems to be distinct from function and structure. Therefore one can conceive of a world with the same physics but no experience. The possibility of such a world would cast doubt on a purely physical account of consciousness.

However, to a lot of people the jump from conceivability to possibility seems questionable. One reason for this is the presence of a posteriori necessities. If we knew everything about the essence of our physics and the nature of consciousness, this might reveal that our physics always needs to be accompanied by our type of phenomenal consciousness.

To avoid this issue, Chalmers suggests that it might not make sense to posit the existence of some essence that is unique to our physics and our consciousness. Instead, he argues, the appearance of the physical laws and of phenomenal states might be all there is. If we accept that physics and consciousness are completely defined by how they appear to us, then this leaves no room for unknown a posteriori necessities about these concepts, and the a priori logical coherence of a zombie world should imply its possibility.

Bibliography

[1] Chalmers, David J. Consciousness and its Place in Nature. URL = https://consc.net/papers/nature.html.

[2] Goff, Philip. Consciousness and fundamental reality. Oxford University Press, 2017.

[3] Kripke, Saul A. (1980). Naming and Necessity: Lectures Given to the Princeton University Philosophy Colloquium. Cambridge, MA: Harvard University Press.

[4] Alter, Torin and Derk Pereboom, “Russellian Monism”, The Stanford Encyclopedia of Philosophy (Fall 2019 Edition), Edward N. Zalta (ed.), URL = https://plato.stanford.edu/archives/fall2019/entries/russellian-monism/.

Because of the buzz around ChatGPT, I’ve been thinking about its nature as a generative text predictor and how humans might be similar in some ways. Not because biological circuits in our brains mimic the Transformer architecture (as far as I know, they don’t) but because the brain has been proposed to be a generative prediction machine for sensory inputs.

As predictive processing theories have also been linked to subjective experience, I believe this provides an interesting lens through which to analyze the phenomenology of reading. As in my previous articles, when I say consciousness, I’m referring to phenomenal consciousness (see https://plato.stanford.edu/entries/consciousness/).

Please note that the following is just a bunch of random ideas loosely inspired by predictive coding and the controlled hallucination theory. Not that I need to say this but don’t treat this as a scientific text.

Predictive Coding

According to predictive coding, sensory perception boils down to modeling the latent causes of sensory inputs using some approximation of Bayesian inference. This could be achieved by a hierarchical, generative model of the world that maps abstract causes to raw sensory signals through several intermediate levels of abstraction. This prediction of low-level sensory inputs travels top-down in the perceptual hierarchy. Failure to perfectly predict sensory inputs would lead to errors passed upward in the brain’s hierarchy, changing neuron activities in the short term, and the generative model itself in the long term.

This results in the model correctly predicting its input in the future. These errors flow bottom-up through the generative model. All neuron activity above the lower-most (sensory) layer can be seen as latent causes for sensory input at different levels of abstraction.

To give an example: To correctly predict the sensory inputs of seeing an apple, a brain (A) requires an abstract representation of an apple and its position to be represented in higher cortical layers and (B) a sufficiently accurate generative model of the world that maps this abstract state of affairs to predicted inputs from the retina. Before achieving this state, the mismatch between a brain’s prediction and the actual sensory input of the apple create bottom-up error signals that change higher-level neural activity until the prediction matches the inputs.

Controlled Hallucinations

In his book Being You, Anil Seth proposes that, beyond representing a mechanistic framework for understanding brain function, predictive coding might help us explain conscious content. Concretely, he suggests the following:

“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.”

Thus, the neural activity that determines our subjective experience is internally generated by a model based on our best guess of what’s causing our sense organs to send us the given signals. In this sense, we can be said to be hallucinating reality. However, this hallucination is constrained by sensory inputs through error signals and is therefore controlled.

Reading a Book

So what happens in our brains when we read a book? Predictive coding suggests that, seeing as letters on a page are just a specific case of visual input, our brain tries to predict the text. Assuming our eyes follow the text in its intended order, our brain predicts the next words by considering what it has read before (much like GPT).

To achieve this, what might be the best latent causes to model the incoming signals from the eyes? Essentially, our brain must infer ‘letters’ to predict the typographic symbols. Letters as latent causes are likely represented in some intermediate levels of the generative model. These, in turn, are translated by the generative model to lower-level retinotopic representations of black lines on a white background that match the incoming visual signals. However, like low-level retinotopic representations, these intermediate ‘letter’ representations also require further upstream latent causes. We thus enter the territory of words, phrases, and grammar.

Climbing further in the perceptual hierarchy, we reach a point where the latent causes have to involve the semantic content of the text. If I’m reading Harry Potter, predicting the next sentence will involve more than the statistics of the English language and modeling how sentences map to words, letters, and black lines. We need to model the representational content of the text. To do this, we need to form concepts of characters, their relationships, past events, locations, etc. These concepts will involve both short-term latent causes, such as Harry is talking to Ron, as well as long-term context, such as Harry’s owl is dead.

Notice that these levels of representation are not bound to books anymore. The same types of latent causes are used in everyday life too. These are inferred causes of your sensory inputs when talking to your friend, such as the concept of your friend and your history with them. Thus, within the representational hierarchy of our generative model, the contents of the upper part might be very similar between reading a book and interacting in the real world. In both cases, we infer similar high-level concepts about how the world works as latent causes of our sensory inputs. The part that differs are the lower-level latent causes of sensory predictions. Whereas in the case of reality, we might predict people, faces, and sounds of spoken words, in books, we predict sentences, letters, and black lines.

In the same way top-down predictions in our generative model flow from high-level concepts such as characters and intentions down to specific actions or events, lower-level concepts such as words and letters predict black lines on white background. According to the Controlled Hallucination Theory (and other theories, see my article on Neurorepresentationalism), all of these predictions involve conscious perception. Thus, if high-level book predictions are quite similar to those from real life, this might explain a lot about why books are so immersive.

To a large extent, reading a book (to me) feels like real life on high abstraction levels. My brain seems to temporarily treat the struggles of characters and the worlds they inhabit as if they were real, but only to an extent. While these fictional worlds might assume some abstract existence, they clearly lack the vividness of the sensory qualities associated with reality. This difference could boil down to the fact that it is only the upper part of the generative model that mimics real life when reading books, whereas the lower-level parts have to take over the job of, not predicting an actual visual scene, but instead translating an abstract version of it into text.

Conclusion

So, what do we gain from speculating about the connection between the top-down prediction of sensory inputs and the phenomenology of reading? Not sure. But I think it paints a compelling picture of why reading might feel the way it does.

What about GPT, then? Could it be conscious, given that it performs a similar, although much more limited, function? While I doubt phenomenal consciousness is present, this is impossible to validate without assuming some concrete theory of consciousness. Still, it’s an interesting question. Even if we all agree ChatGPT is not conscious, it would be interesting to investigate the explicit reasons why we think it lacks phenomenal experience, given that our own phenomenology of perceiving text could be linked to predicting it.

If you have spent any time thinking about consciousness and its relation to brain activity, I’m guessing that the following claim should not sound too unreasonable: The neural correlate of a percept, such as seeing an apple, will at least in some form involve a representation of that apple. This, however, raises many questions: What kinds of representations are required for conscious content? Is the presence of representations enough for phenomenal experience? Are all representations subjectively experienced? A recent paper by Cyriel M.A.Pennartz, introducing Neurorepresentationalism as a new theory of consciousness [1], proposes answers to some of these questions.

In this article, I discuss a few ideas outlined in the paper and embed them within the philosophical context of representationalism and the hard problem of consciousness. Additionally, I introduce a thought experiment that deals with some pretty wild (but speculative) implications of representational accounts of consciousness, and how those might in turn shape our assumptions about the necessary ingredients of subjective experience.

Representationalism

The idea that conscious experiences are intimately connected to mental representations in our brain is not new. Representational theories of consciousness propose that representations in our brain at least partially account for consciousness [4].

One of the types of ‘problems’ that representationalist theories are trying to address is illustrated by the following thought experiment: Say I have a visual percept of an apple on a table, but there is no apple on the table. In other words, I am hallucinating an apple. This raises the question of why I am aware of something apple-like. If the apple is not on the table, where is it? Is it my head? There is definitely no apple in my head, so the apple is nowhere. At first glance this might seem like a meaningless train of thought (at least I thought so at first). Of course there is no apple in my head and of course it’s possible to still see one, as long as the right brain activity is present. However, if there is no actual apple, where does the apple-likeness in my percept come from? Why does this particular neural activity lead to an apple percept? After all, it’s not like neurons in my head are somehow arranged into an apple shape.

This is where representationalism enters the scene: The neurons themselves are not apple-shaped, but what they represent is apple-shaped. This is analogous to the bits of a JPEG binary not being apple-shaped themselves, but they’re encoding an apple-shaped image. This means that, whether the apple is actually on the table or not, the representational content of visual cortex neurons includes that apple.

However, this still leaves open many questions. One of the most pressing ones is the question of what the required attributes of a representation are in order for it to be conscious. Here there are multiple varieties of representationalist theories of consciousness:

As far as I understand, Neurorepresentationalism is a member of the strong representationalist theories of consciousness, but more on that in the next section.

For now, let me just say that, even if we decide on which type of representationalism might be true, another question that remains is how the representational content of any neural activity might be decided. To get back to the JPEG example, the same binary digits could represent any arbitrary image given another encoding scheme. Part of this question might be answered by Neurorepresentationalism, but I hold that some mystery remains with any strong representationalist theory, as I will discuss in the final section.

Finally, I doubt that any strong representationalist theory of consciousness will convince everyone that the hard problem is solved. But this might not be the right goal anyways. On the other hand, progress towards solving the Real Problem and maybe even the Core Problem is conceivable (see my article on the hard, real and core problems of consciousness).

Neurorepresentationalism

Having set the stage for the abstract family of representational theories of consciousness, I will now introduce a concrete, fairly recent, instantiation of this group: Neurorepresentationalism [1].

Definition of Consciousness

Motivated by the characterization of essential features of phenomenal consciousness and by its postulated functional role in supporting goal-directed behavior (check out the paper for more details [1]), Pennartz puts forward the following definition:

[Consciousness] is the multi-modally rich, dynamic survey of the subject’s current situation, including his own body and functionally earmarked for planned behavioral and cognitive actions in the future.

My first thoughts when reading a definition like this are: Is this a functional account of consciousness? If so, aren’t we just staying in the realm of the easy problems of consciousness [3] without making any progress on phenomenal consciousness? Pennartz argues that, no, in addition to function, this also addresses phenomenal / qualitative consciousness. He claims that, although consciousness subserves goal-directed behavior, it is not identical to it. In other words, although consciousness is functional with respect to creating a situational multi-modal survey, it is not functional with respect to the overall behavior of the agent. As a result, in terms of input and output, a functionally equivalent agent without consciousness (a functional zombie) is possible. In short, this definition attempts to capture the phenomenal aspect of consciousness without relegating it to a mere epiphenomenon.

Neural Correlate of Consciousness

It should come as no surprise that the proposed correlate of consciousness is a type of neural representation. More concretely, Pennartz proposes conscious representations to be organized in multiple levels of abstraction, each identifiable with constituents of our phenomenal experience at different granularities [1]:

1. single neurons, having the capacity to respond to single features

2. ensembles of neurons, forming small, within-area local networks capable of pattern coding within a single submodality (e.g. shape)

3. unimodal metanetworks, which combine the hypotheses from lower- order ensembles into representations of objects considered within a single modality (e.g. all visual features making up a visual object)

4. multimodal metanetworks, integrating the information coded by unimodal meta-networks into multisensory object representations

Pennartz claims that consciousness arises at the highest level of this representational hierarchy, subsuming multiple perceived objects in different modalities. Moreover, these representations are implemented on the neuronal level in such a way that both low-level and high-level cortical areas contribute to conscious contents.

At first glance, it might sound like the previous two statements are conflicting with each other. How can only the highest level of representations be conscious while low-level cortical areas are also involved in consciousness? While tempting, it would be a mistake to equate the representational hierarchy as shown in the diagram above with the cortical hierarchy. The important difference is that higher representational levels are not just affected by lower levels, but they subsume the lower levels. Pennartz suggests that, instead of thinking about representations as serially being passed upward in the cortical hierarchy, we should conceive of a representation encompassing multiple cortical areas at once, being ‘vertically integrated’ if you will. Thus, a particular neural assembly can be seen as representing a low-level feature like shape in isolation, while also participating in a higher-level, multi-modal representation of an apple. Note that, because we are identifying consciousness with a certain type of high-level representations, this is an instance of a strong representational theory of consciousness [2].

A further claim is made about the minimal neural substrate for a subjective experience, namely, that it needs to integrate more than one modality. This is also implied by the previously given definition of consciousness as being a ‘multi-modally rich, dynamic survey’. On the face of it, however, this requirement might seem arbitrary. Why couldn’t, for example, the combined system of V1 and V4 create an experience of color perception on its own, without another visual sub-modality having to be part of it? To get to a potential answer to this question, let’s try to take the perspective of V4 itself. V1 (and potentially other areas, I’m no expert on the visual system) projects to V4 and sends electrical signals relevant to color. But V4 doesn’t ‘know’ that the incoming signals relate to color, they could be about anything! Pennartz (I think) argues that, color has its specific character to us precisely because within our conscious experience, it is contrasted to other modalities. The way in which V4 activity correlates with, say, activity in MT, is what gives it its character. This idea reminds me of the information axiom in IIT [4], which asserts that an experience is informative by virtue of excluding other possible experiences. Similarly, the experience might be color-like, by virtue of how it differs from (and correlates to) representations of motion in the visual field.

Given this proposed neural correlate of consciousness, Neurorepresentationalism predicts subjective experience to mainly arise in sensory and associative cortices. This is consistent with claims made about the posterior hot zone as described by Christoph Koch [5].

Multi-Level Representations and the Hard Problem

So far I have introduced a definition of consciousness, which is postulated to capture phenomenal consciousness, and its ostensible neural correlate.

Let’s say that Neurorepresentationalism’s claims about the neural correlate of consciousness are true: Every aspect of our conscious experience can be traced to a multi-level, multi-modal representation within our sensory and associative cortices. Moreover, let’s assume that these representations are not only sufficient, but necessary, i.e. there are no other ways for a system to experience phenomenal consciousness.

In the following, I want to discuss to what extent this would ‘solve’ the problem of consciousness. To approach this, I’m going to make use of three versions of the problem statement: the real problem, the hard problem, and the core problem of consciousness. I describe these in more detail here.

The Real Problem of Consciousness. Anil Seth describes the problem of explaining consciousness as the task of devising a theory, which reliably explains, predicts, and controls conscious experience as a function of brain activity. ‘Explaining’ in this case entails more than just creating a catalogue of neural activity for every possible experience. Instead, we want to find neural mechanisms that are responsible for aspects that are part of our experience. Testing such an explanatory framework can be carried out by validating its predictions against subjects’ verbal reports, or by trying to change neural activity to arrive at a desired phenomenal state. Assuming that Neurorepresentationalism is true, it would fulfill all of these requirements for explaining consciousness. Given a detailed account of how multi-modal, multi-level representations are structured, we can easily map every aspect of our experience to neural mechanism, predict what experience might result from a given representation, and change representations in order to elicit the desired experience. So far so good. However, Chalmers might say that, while at this point we have found a systematic correlate of consciousness, this might not truly explain why or how consciousness arises from this neural activity.

The Hard Problem of Consciousness. One reason we might not be satisfied with a solution that solves the real problem, is because, strictly speaking, we are developing a theory of behavior, namely, verbal reports (or whichever behavior we use as ground truth for conscious experience). In that sense, we are only answering easy problems. Moreover, even if verbal reports are accurate, there would be an explanatory gap of why and how such behavior would be accompanied by phenomenal experience, which rests on conceivability arguments of philosophical zombies. The way the hard problem is framed almost by definition precludes any possible solution to it. This is, as far as I understand, at least partly intentional. I think Chalmers wants us to question our underlying premises (i.e. physical primacy) when trying to solve the problem of consciousness. But that still leaves us with no clear problem statement of what it is we’re trying to find out. The real problem is a good step into the direction of a pragmatic scientific aim for studying consciousness, but it doesn’t address ontological considerations about consciousness, except by (as far as I can tell) making a big implicit assumption about qualia being reducible to physics. Neurorepresentationalism, however, has more to say about this.

The Core Problem of Consciousness. The idea behind this formulation is to acknowledge the existence of both phenomenal consciousness and physical reality, while remaining agnostic as to their exact relation. This leaves us with the following problem statement: We need to explain how the fact that there is something it is like to be us relates to physical matter [6]. This perspective allows for the pragmatic approach suggested by the real problem, while leaving open the question about the metaphysical relationship between consciousness and physics. This leaves room for Pennartz’ statement that

Neurorepresentationalism acknowledges phenomenal experience as real and existing, on an equal ontological footing with the electrophysiological mechanisms ultimately underlying it.

Here I want to make clear that my interpretation of this statement might not reflect the author’s view on the matter exactly. This is because to me there seems to be some friction between ‘on equal ontological footing’ and ‘ultimately underlying it’. I will interpret this passage as saying that electrophysiological data is not more fundamental than phenomenal experience, both are equally real and one cannot exist without the other. However, even if the author intends to say that phenomenal experience is real, but reducible (i.e. conservative realism or weak illusionism as defined by Keith Frankish [7]), a lot of the following discussion is still relevant.

If we simply assume that our phenomenal experience exists, and neurons exist (and all the underlying physics), this still leaves an important question: How can two such seemingly incommensurable things like neurons and experience relate to each other? Pennartz attempts to explain the hardness of reconciling consciousness and neural activity like this: We know a lot about our conscious experience (you could argue it might be the thing we are most familiar with). Moreover, we also know a lot about neurons, and we have an idea about how a neuron of some animal fires upon presenting the right stimulus. However, we know nothing about all the intermediate levels of representations between single neurons and the top-level representation, which is identical to our consciousness (remember, our consciousness is equal to the highest abstraction level of a multi-level, multi-modal representation, the representational content of individual neurons make up the lowest level).

While this explanation provides a tentative account for the hardness of the hard problem, I think someone like Chalmers might still be unsatisfied. The question of why it is that at the highest level, multi-model representations are associated with phenomenal experience could be seen as unresolved.

This is why I think the core problem might be a better framing, which is to say, instead of answering how/why subjective experience arises from the physical, let’s try to figure out what the relation between the two is. From this perspective, Neurorepresentationalism provides a coherent answer: Phenomenal consciousness relates to the physical by corresponding to the highest level of a hierarchically constructed representation, which at its bottom layer consists of individual (physical) neurons.

Neurons arranged in a certain way imply phenomenal consciousness (by means of forming a representation) just as much as consciousness implies an implementation of a multi-level, multi-modal representation (which can be neurons, but is implied to be substrate-independent according to Neurorepresentationalism). One way to cash out this idea as an ontology is as some type of neutral monism: There is some reality out there, which we cannot access directly, and there are different ways to interpret or ‘decode’ it. On the one hand, one could decode this reality at the physical level, namely, as fundamental particles, atoms, neurons, etc. Another possibility would be to decode it as conscious representations. The concept of a neuron might not have to be evoked in the latter case. Please note that this ‘neutral monism interpretation’ is my own interpretation and is not proposed by Pennartz.

Exclusivity

Earlier I mentioned Neurorepresentationalism’s claim that, for a multi-level representation to be conscious, it needs to encompass more than one modality. This implies that the lowest representational levels cannot be conscious (only as part of higher levels).

However, it does not imply that we need to reach the highest level before consciousness arises. This is because multi-modal meta-networks can arise at intermediate stages in the representational hierarchy. Still, Pennartz proposes that consciousness arises solely on the highest level, which suggests that he is assuming a type of ‘exclusivity’ rule: Individual sub-components of a conscious representation cannot also be conscious on their own.

This shows a lot of similarities with Integrated Information Theory’s axiom of exclusivity [4]. To me this constitutes an extra assumption that we don’t necessarily have to make. A theory that leaves open the possibility of ‘sub-experiences’ having an existence of their own might seem strange, but it carries fewer assumptions.

Representing Everything with Anything

In the previous section, I gave a short (and limited) account of Neurorepresentationalism, combined with some ideas about how its claims could be cashed out as an ontology about physics and consciousness. Next, I want to zoom out a bit and explore some potential implications of strong representational theories of consciousness [2]. While spooky and speculative, I believe they deserve serious consideration.

Assume we are looking at a conscious representation of a visual percept of an apple. This could be a set of neurons with a particular firing pattern, or a state (or sequence of states) in a digital computer. While the possibility for consciousness on a digital computer is contested, a strong representational theory that makes no claims about the substrate of representations may allow for this (as does Neurorepresentationalism). Let us also assume our representational theory of consciousness is narrow, meaning that the conscious content is only dependent on the representation, rather than dependent on what’s going on outside of the agent [2]. In other words, two agents with the same internal representations will have the same experience. This means that our apple percept is fully determined and instantiated by the apple representation. Looking at our neurons or transistors, how can we tell that the representational content corresponds to an apple instead of an orange? Or, more broadly, how can we tell that it corresponds to the modality of vision at all? Clearly, you might say, given the right decoding scheme, we can easily extract that information. However, who decides the decoding scheme? How does the subjective experience of the apple ‘know’ how it needs to decode its underlying representation? Minus the consciousness, this is equivalent to saying that a random string of bits can be decoded to any arbitrary file, given the right file format. Why, then, does my conscious experience ‘know’ it’s experiencing any given experience instead of an arbitrary other one? Pardon all the questions, but here is the final one: If anything can represent everything (given the right decoding scheme), how could some system ever have a definite conscious experience? There are a lot more nuances to this line of reasoning, and I highly suggest reading Greg Egan’s Permutation City [8] (a hard sci-fi novel packed with a huge number of crazy ideas, culminating in one very trippy theory) and/or Hans Moravec’s essay Simulation, Consciousness, Existence [9].

I know, there are a few steps made here that rely on a very simple notion of the word ‘representation’. However, if the above reasoning is wrong, I think it’s import to be explicit about why it fails, which would reveal further assumptions we place on what types of representations can be conscious.

How might Neurorepresentationalism avoid this hyper panpsychist universe where anything can encode any conscious experience? Well, you might say, Neurorepresentationalism puts a lot of constraints on what kinds of representations are conscious: They have to be multi-modal, multi-level, etc. But remember, being substrate-independent, Neurorepresentationalism could be implemented on anything, including computer chips. Moreover, whether a string of bits encodes a multi-modal representation or a high-resolution photo of yourself is entirely dependent on the encoding scheme. Thus, unless we put further assumptions on how these representations are instantiated in the physical world, or how they are encoded, we cannot avoid finding multi-modal representations wherever we want! There may be very intuitive assumptions that we could make to avoid this scenario, such as requiring some continuity across time. For example, we might require the probability distribution of the visual part of the multi-modal representation to follow the statistics of visual scenes. We might also require different parts of the representation to have some causal effect on each other (similar to IIT [4]). In both cases, we would move away from a static conception of representations. It’s quite likely that similar assumptions are being made with Neurorepresentationalism, but I think it’s worth thinking about this explicitly.

Conclusion

Representationalism provides a reasonable perspective from which to approach a scientific theory of consciousness. After all, few people doubt that our perception of an apple is deeply connected to some internal model we have of that apple.

Neurorepresentationalism proposes a theory for how such representations are structured, and how they relate to subjective experience. Making clear claims about neural correlates of consciousness, a theory like this lends itself to scientific study according to the framework of the real problem. Moreover, by treating subjective experience and physics as equally real, but corresponding to each other through a hierarchy of representations, we are avoiding any strong assumptions about the ontological primacy of one over the other.

However, as with all strong representational theories, we have to be careful about our requirements for conscious representations. Barring any further constraints on continuity or causality, representational views of awareness quickly lead us to hyper-panpsychist scenarios which, while not necessarily impossible, seem quite extreme.

Bibliography

[1] Pennartz, Cyriel MA. “What is neurorepresentationalism? From neural activity and predictive processing to multi-level representations and consciousness.” Behavioural Brain Research 432 (2022): 113969.

[2] Lycan, William, “Representational Theories of Consciousness”, The Stanford Encyclopedia of Philosophy (Fall 2019 Edition), Edward N. Zalta (ed.), URL = https://plato.stanford.edu/archives/fall2019/entries/consciousness-representational

[3] Chalmers, David J. “Facing up to the problem of consciousness.” Journal of consciousness studies 2.3 (1995): 200–219.

[4] Oizumi, Masafumi, Larissa Albantakis, and Giulio Tononi. “From the phenomenology to the mechanisms of consciousness: integrated information theory 3.0.” PLoS computational biology 10.5 (2014): e1003588.

[5] Koch, Christoph. “What is Consciousness?” nature.com URL=https://doi.org/10.1038/d41586-018-05097-x

[6] Chis-Ciure, Robert, and Francesco Ellia. “Facing up to the hard problem of consciousness as an integrated information theorist.” Foundations of Science (2021): 1–17.

[7] Frankish, Keith. “Illusionism as a theory of consciousness.” Journal of Consciousness Studies 23.11–12 (2016): 11–39.

[8] Egan, Greg. “Permutation City”. Millennium Orion Publishing Group, 1994.

[9] Moravec, Hans. “Simulation, Consciousness, Existence” (1998) URL=https://frc.ri.cmu.edu/~hpm/project.archive/general.articles/1998/SimConEx.98.html

In this article, I want to put the hard problem of consciousness into the context of two alternative problem statements: The real problem of consciousness and the core problem of consciousness. Below I address all three problem statements after a quick introduction to the definition of phenomenal consciousness. (Feel free to skip that part if you are familiar with the term.) My aim is to elucidate the assumptions of the different problem statements by contrasting them with each other. I conclude with my opinion on what the best problem statement is in terms of finding common ground between different scientists and philosophers, while at the same time not losing sight of the mystery at the heart of consciousness.

Disclaimer: This article represents my interpretation of the problem statements put forward by the various authors. If any of the ideas outlined don’t make sense, it is probably me who is to blame. My background is in deep learning and computational neuroscience. I have no formal background in philosophy.

Background: What is Consciousness?

Together with words like ‘god’ or ‘meaning’, this might be one of the contenders for the most fuzzy terms out there. Ask ten people for a definition, and you might get ten different answers. Moreover, such answers are often circular, such as to be conscious is to be aware of something. However, this does not mean that there is nothing there that is worth investigating. The kind of consciousness I am interested in is phenomenal consciousness, or qualitative consciousness. To start, let me go over some common ideas about consciousness that phenomenal consciousness does not imply: A sense of self or identity, a meta-cognitive awareness of what I’m doing, or some level of general intelligence. Note that all of these could be associated with phenomenal consciousness, but they are not identical to it or implied by it. So what’s left? Among people interested in phenomenal consciousness, a very popular essay to cite is Thomas Nagel’s What Is It Like to Be a Bat? [1]. Nagel proposes that, for any being to have a subjective experience (i.e. to be phenomenally conscious), there is something it is like to be that being. In other words, if it feels like anything at all to be a bat, bats are conscious. This is an ostensive definition, meaning it doesn’t give a full account of the phenomenon, but it points to it. Of course, compared to other definitions, this might not be the most satisfying one, but I believe it is one of the best ways we have for using language to describe the idea of phenomenal consciousness. Another way to capture the same intuition is by thinking about subjectivity vs. objectivity. We can study a bat’s brain until we perfectly understand its behavior as a function of external inputs and internal dynamics from an outside, or objective, perspective. However, no matter how much we understand this system from the outside, we (ostensibly) will not know what it is like to be that system from a subjective perspective. For more information on different definitions and categorizations of consciousness, check out [2].

Within phenomenal consciousness, it is also common to draw a distinction between conscious state and conscious contents, also referred to as global and local states of consciousness [3]. The former points to overall descriptions of a subject’s conscious experience, such as wakefulness, sleep, or potentially something like selflessness (as induced by say meditation or psychedelics). The latter refers to specific objects within our conscious experience, such as smells, visual objects, sounds, joint pains, etc. In the following, I will mostly be talking about conscious contents, although the distinction between the two does not seem completely clean-cut to me.

The Hard Problem

The hard problem of consciousness, coined by Chalmers [4], is one of the most commonly thrown around phrases when talking about consciousness. Its aim is to distill the most mysterious aspect of subjective experience from other, more easily explainable phenomena, that are often associated with it. Essentially, it separates the idea of phenomenal consciousness from behavioral and cognitive brain function, much like Nagel’s definition, but in a different way. It doesn’t just point to it, but it frames its explanation as a problem that is seemingly out of the reach of science. Chalmers isolates the hard problem by first defining the easy problems of consciousness: Explaining how neural activity gives rise to various cognitive functions and behaviors. These are easy in the sense that, at least in principle, the tools of neuroscience and cognitive science should be well-suited for this task, and we can imagine what such solutions might look like. If for instance, we find some principle of computation, which, when implemented in a simulation, produces the same behavior as our neocortex, then we are at least on a good track to solve the easy problems. (Although a satisfying solution might also require explanations on higher levels of abstractions that are consistent with and are reducible to the lower-level explanations.) Setting aside these easy problems, we’re left with phenomenal consciousness. However, by definition, what we’re left with is not characterizable by its function or behavior. Thus, how could we explain it? What would such an explanation even look like? Here is a formulation of this hard problem in the words of Chalmers [4]:

Why is it that when our cognitive systems engage in visual and auditory information-processing, we have visual or auditory experience: the quality of deep blue, the sensation of middle C? How can we explain why there is something it is like to entertain a mental image, or to experience an emotion? 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 hard problem is therefore a question of why and how physical processes give rise to subjective experience. As things stand, it appears that there could be no scientific experiment that ever gets us closer to tackling the hard problem. This is because science requires measurements, and something that is not characterized by function or behavior cannot produce such measurements. Should we therefore give up and acknowledge phenomenal consciousness is outside of the realm of science? Maybe not. Maybe this formulation of the issue is problematic. To see why this could be, it’s important to recognize that the idea of the hard problem rests on an assumption that is implicit in the above quote: Assumption 1 — we can conceive of all of our brain’s activity and behavior to take place without an associated phenomenal experience, and it is thus possible. In other words, the question of why physical processes would give rise to consciousness in the first place only makes sense if there was any way that physical processes taking place in our brain could exist without the associated experience. Moreover, it assumes a directionality in the explanation: Assumption 2 — A (scientific) account of consciousness should assume the physical and explain the phenomenal. Most people in the scientific community would take this directionality for granted. However, it is not obvious why this would be true, especially from an epistemological standpoint: The only thing we have certain knowledge of is phenomenal experience, everything else is conjecture. But I don’t want to argue for idealism here, I just want to question the assumption of the directionality implicit in the hard problem. It is worth mentioning, however, that Chalmers uses this formulation of the hard problem as an argument against this directionality. In the paper introducing the hard problem, he proposes a type of dual-aspect monism, where physics and phenomenal experience stand on equal ontological grounds. This leaves us with a problem statement that, by design, cannot be addressed by science.

The Real Problem

Given that the hard problem by definition precludes any scientific solution, more pragmatically-minded people have tried to reformulate the problem of explaining phenomenal consciousness. Enter Anil Seth. He defines the real problem of consciousness [5] as

“the primary goals of consciousness science are to explain, predict, and control the phenomenological properties of conscious experience. This means explaining why a particular conscious experience is the way it is — why it has the phenomenological properties that it has — in terms of physical mechanisms and processes in the brain and body. These explanations should enable us to predict when specific subjective experiences will occur, and enable their control through intervening in the underlying mechanisms.

This brings the problem of consciousness back into the realm of science. After all, if an explanation constitutes a systematic theory that makes predictions about experience based on brain activity, then it is testable.

However, this only works if we have a ground truth on conscious experiences, i.e. we can only validate a theory if we can check whether its predicted conscious contents match reality. The only way to get to these ground truths is by relying on behavior, be that verbal reports or some other motor action of a subject. Thus, within the framework of the hard problem, we are just dealing with easy problems. Chalmers might see the project of solving the real problem as a search for a systematic neural correlate of consciousness [6] rather than a solution to the hard problem. But that’s ok, we don’t have to accept the framing of the hard problem; given its assumptions, it might not be the right one anyways. Moreover, Chalmers does acknowledge the absurdness of assuming consciousness to be a pure epiphenomenon (that is, behavior being completely divorced from phenomenal consciousness) when discussing the meta-problem of consciousness [7]. Still, let’s spell out Seth’s assumption explicitly: Assumption 1 — We can investigate conscious contents by measuring behavior. The hard problem denies this assumption.

In terms of ontology, we are still assuming a directionality of explanation, as with the hard problem: Assumption 2 — An account of consciousness should assume the physical and explain the phenomenal. In the end, Seth argues, while phenomenal consciousness is real, it might not be what we think it is, and it should be reducible to physical processes. This is closely resembles the view that Keith Frankish labels as weak illusionism [8]. Chalmers is skeptical about the power of weak illusionist theories to explain phenomenal consciousness [7]. If only physics is fundamental, it has to account for phenomenal consciousness. Under this assumption, the claims of the hard problem about the difficulty of this problems are fair, I think. If phenomenal consciousness is seen as real, but not fundamental, we have to give an account of how its phenomenal properties arise beyond function and behavior.

The Core Problem

While being more optimistic about our ability to infer phenomenal states from behavior, the real problem is arguably more parsimonious about metaphysical assumptions (the conceivability argument) than the hard problem. Could we make it even leaner in terms of its presuppositions, while leaving open the possibility for scientific inquiry? Possibly. One such idea has been put forward by Robert Chis-Ciure and Francesco Ellia called the core problem of consciousness [9]. The authors describe the hard problem as being layered in the sense that it can be decomposed (paraphrased by me, therefore possibly inaccurate):

1. There is phenomenal consciousness, which needs to be explained.

2. Conceivability arguments of non-conscious functional (and structural) equivalents are coherent. As a result, physical explanations cannot account for phenomenal consciousness.

They suggest that we can do away with the second layer, while holding on to the most basic idea of relating consciousness to physical matter. They thus define the core problem of consciousness as

We need to explain how the fact that there is something it is like to be us relates to physical matter.

It seems to me that this does away with both Assumption 1 and Assumption 2 outlined when describing the hard problem: We don’t assume any conceivability of non-conscious functional equivalents, and we don’t assign any ontological primacy to physics. In a sense, this is a much leaner version of the hard problem which still captures much of the mysteries that the hard problem addresses. The only assumptions we’re making (loosely speaking) is that both phenomenal consciousness and physics exist, and that they are somehow related. This fundamental assumption is very hard to deny, and is implied by both the real problem and the hard problem. In that sense, the core problem represents a weaker form of both the hard problem and the real problem; it thus functions as an intersection between these frameworks.

Conclusion

Combining the metaphysical leanness of the core problem with the pragmatism of the real problem (i.e. let’s assume we can at least partially measure conscious contents by behavior), we arrive at a problem statement that might be interesting to a broader range of people. Whether you’re a realist or a weak illusionist [8], the project of characterizing the relation between processes in the brain and consciousness without assuming any ontological primacy of physics or consciousness should be of interest to you. This endeavor subsumes both the scientific aspect of finding a systemic neural correlate of consciousness, as well as the question of what this relation means on a metaphysical level.

The epiphenomenal nature of consciousness suggested by the hard problem can alienate scientifically-minded people, while the reductionist prior of the real problem might turn away people who reject the assumption of physical primacy. The terminology of the core problem, on the other hand, could help scientists and philosophers, that might have some disagreements in their assumptions, find common ground. To me, an epistemologically optimistic (with respect to conscious contents, from an objective perspective) version of the core problem might be the best framework for the scientific study of phenomenal consciousness.

Bibliography

[1] Nagel, Thomas. “What is it like to be a bat?.” The philosophical review 83.4 (1974): 435–450.

[2] Van Gulick, Robert, “Consciousness”, The Stanford Encyclopedia of Philosophy (Winter 2022 Edition), Edward N. Zalta & Uri Nodelman (eds.), URL = https://plato.stanford.edu/archives/win2022/entries/consciousness/.

[3] Seth, Anil K., and Tim Bayne. “Theories of consciousness.” Nature Reviews Neuroscience (2022): 1–14.

[4] Chalmers, David J. “Facing up to the problem of consciousness.” Journal of consciousness studies 2.3 (1995): 200–219.

[5] Seth, Anil. Being you: A new science of consciousness. Penguin, 2021.

[6] Chalmers, David J. “What is a neural correlate of consciousness.” Neural correlates of consciousness: Empirical and conceptual questions (2000): 17–39.

[7] Chalmers, David. “The meta-problem of consciousness.” Journal of Consciousness Studies 25.9–10 (2018).

[8] Frankish, Keith. “Illusionism as a theory of consciousness.” Journal of Consciousness Studies 23.11–12 (2016): 11–39.

[9] Chis-Ciure, Robert, and Francesco Ellia. “Facing up to the hard problem of consciousness as an integrated information theorist.” Foundations of Science (2021): 1–17.