Abstract
Illusionism is a general philosophical framework in which specific theories of consciousness can be constructed without having to invoke a magical mind essence. The advantages of illusionism are not widely recognized, perhaps because scholars tend to think only of the most extreme forms and miss the range of possibilities. The brain's internal models are never fully accurate, nothing is exactly as the brain represents it, and therefore some element of illusionism is almost certainly necessary for any working theory of consciousness or of any other property that is accessed through introspection. Here I describe the illusionist framework and propose six specific theories. One purpose of this article is to demonstrate the range of possibilities in a domain that is not yet sufficiently explored. The second purpose is to argue that even existing, popular theories, such as the integrated information theory or the global workspace theory, can be transformed and greatly strengthened by adding an illusionist layer. The third purpose is to argue that when illusionist logic is used, even very disparate theories of consciousness that emerge from unrelated conceptual origins begin to converge onto a deeper, unified understanding.
Significance Statement
Illusionism is a philosophical framework in which consciousness is not exactly what we think it is. The brain's internal models are never perfectly accurate, and therefore nothing is exactly as the brain represents it. The conceptual advantages of illusionism are not widely recognized, perhaps because scholars tend to think only of the most extreme forms and miss the range of possibilities. This article describes the illusionist framework and proposes six specific theories within it. One purpose is to demonstrate the range of possibilities in a domain that is not yet sufficiently explored. Another purpose is to show that with illusionism, even very different theories begin to converge toward a deeper, unified understanding of consciousness.
For Dan Dennett, who put illusionism on the map.
Introduction
The term “illusionism” gives the impression of a theory that dismisses the entire phenomenon of consciousness. Indeed, some forms of illusionism do exactly that, while other forms are more cautious (Gazzaniga, 1970; Nisbett and Wilson, 1977; Dennett, 1991; Blackmore, 2003; Frankish, 2016). I argue that illusionism should be seen as an umbrella term. It is not a specific theory of consciousness, so much as a general philosophical framework in which many theories can potentially be constructed. To my knowledge, however, the only cognitive neuroscience theory in the literature that is overtly illusionist is the Attention Schema Theory (AST), proposed in 2011 (Graziano and Kastner, 2011; Graziano, 2013; Webb and Graziano, 2015; Graziano et al., 2020). It would be beneficial to build other illusionist theories to serve as healthy competition to AST and to show the range and power of the general approach. Moreover, many already existing theories of consciousness can be seamlessly retrofitted with an illusionist layer, potentially greatly strengthening them. The powerful conceptual advantages of illusionism are not widely recognized, perhaps because scholars tend to think only of the most extreme form and miss the range of possibilities.
Broadly speaking, in illusionism, consciousness is not exactly what we think it is (Gazzaniga, 1970; Nisbett and Wilson, 1977; Dennett, 1991; Blackmore, 2003; Frankish, 2016). The brain's internal models are never perfectly accurate, and therefore absolutely nothing is exactly as the brain represents it. In the illusionist perspective, the reason why explanations of consciousness have been elusive is that scholars are mistakenly attempting to explain the subtleties of something that does not have the form they think it does.
The modern study of consciousness tends to focus on how the brain generates an intangible essence of experience. Yet the quality of experience was not always the target of consciousness research. More than a hundred years ago, James (1890) coined the term “stream of consciousness.” To him, consciousness was a stream of content—thoughts, decisions, sensory impressions, and memories. In principle, anything that had that stream of content was conscious. Then, in the middle twentieth century, computers began to encroach on the domain of consciousness. In 1950, for example, Turing (1950) wrote his famous paper on whether computers can think. Bit by bit, computers began to take on many of the components that are supposed to make up the stream of consciousness. They could process sensory input, store and retrieve memory, make decisions, and can now do even more. I suggest that this gradual encroachment of technology and computation over the decades threatened people and led to a rising antitechnology discomfort. That discomfort may have motivated scholars to define something special about the conscious mind that can be dissociated from mere computable content. Nagel (1974) wrote about experience, the “what-it-is-like” part, and Chalmers (1995) wrote about the “hard problem” as distinct from the “easy problem.” When you look at an apple, you do not just process the shape and color like a computer. You have a special, intangible essence, an internal experience. By now, almost all theories of consciousness accept, as a fundamental assumption, the existence of the what-it-feels-like essence. (For an extensive survey of existing theories, see Blackmore and Troscianko, 2018; Doerig et al., 2021) In that framework, the task of a theory is to explain how phenomenal experience is generated, or at least to specify the conditions under which it exists.
In illusionism, however, the question of consciousness is different. Imagine that a house is photographed and the picture is run through too many photocopies, resulting in a black-and-white, blurry, almost unrecognizable representation. To understand that picture, our first job is to avoid interpreting it too literally and mistakenly assuming that there exists an actual, 3-inch-tall, two-dimensional house with smudgy boundaries. We must be sleuths and identify the original building that is the subject of that representation, and we must understand the steps by which that building was rendered into a distorted representation. Just so, imagine an unknown something—whether substance or process—that is rendered into a representation by the brain's imprecise information embeddings, then run through higher semantic manipulation, and then run through speech production. It comes out the other end of this chain of depictions and copies of depictions as the claim that we have an ethereal magic of consciousness inside of us. As scientists, as consciousness researchers, can we figure out what the original item is? Can we figure out its specific adaptive uses in the brain and the steps by which it became represented in an imprecise or schematic manner? And just how different is the real item from the consciousness that we claim to have? In this context, it has been suggested that the term “caricature” might be a more apt descriptor of consciousness than “illusion” (Graziano, 2019). To most people who first encounter the phrase, “consciousness is an illusion,” the term seems to imply that there is nothing present and the brain has made a mistake—in effect, the picture of the house is a fake and no house exists. This is probably not exactly what most scholars mean when they use the term illusionism, and yet it tends to be the message that readers take. Caricature may be a better term, because it more clearly implies that there is, actually, a real subject of the representation, that our understanding of that item is distorted rather than entirely wrong, and that the whole phenomenon is not just a brain mistake but serves a useful purpose.
In this article, I will first describe a framework for thinking about the problem of consciousness from a representation point of view, emphasizing information flow through neural systems. I will then propose and discuss six different hypotheses about “Item 1,” the unknown thing from which human claims about consciousness stem. Each of these hypotheses represents a different illusionist theory. Some are extreme forms that are unlikely to be correct, but I present them to show the range of possibilities. Some are closely related to common theories in the current literature. For example, Hypothesis 3 is an illusionist version of the integrated information theory (IIT; Tononi, 2008; Tononi et al., 2016). Pairing IIT with illusionism may seem like odd bedfellows to those who know, and yet the hypothesis makes some interesting rational sense, showing just how versatile and useful illusionism can be. Likewise, Hypothesis 4 is an illusionist version of the global workspace theory (GW; Baars, 1988; Dehaene, 2014; Mashour et al., 2020). Hypothesis 5 is AST (Graziano and Kastner, 2011; Graziano, 2013; Webb and Graziano, 2015; Graziano et al., 2020).
The most remarkable feature of these illusionist theories may be that so many of them converge. Despite their different origins, Hypotheses 3, 4, and 5 are less like alternatives and more like keyhole perspectives on one underlying theory. In the final section of the paper, I describe how the illusionist approach may open a window on a deeper, unified framework for consciousness.
From Vision to Speech
In this section, I analyze a particular instance of verbal behavior. The goal is to lay out a story that is uncontroversial to neuroscientists, at least in its general outlines. I will then shift to the topic of consciousness and ask how far one can proceed using the same concepts.
In Figure 1, Bob tells us, “There's an apple in front of me. It's red.” This speech output is labeled Item 4. One can work backward from Item 4 to provide an overarching explanation for the behavior. Speech muscles are controlled by spiking neuronal activity mainly in the hypoglossal nerve, though other nerves contribute. Those nerve outputs are under the control of a broad speech network including the facial nucleus, motor cortex, Broca's area, Wernicke's area, and so on (Hagoort, 2017). The details are not important to the present argument, and Figure 1 schematically highlights only one area. What is important here is that neurons in a speech-related network are active in a complex pattern, causing Bob's speech output.
Highly schematized account of the transformation from visual input to verbal output. Light reflecting from an apple (Item 1) is transduced into neural signals. The neural information is rendered into an embedding or a model (Item 2) which consists of a complex pattern of activity among neurons. The visual model influences a higher cognitive network (Item 3) that incorporates semantic embeddings. The higher cognitive network influences a speech network (Item 4) that constructs the correct neuronal output signals to activate speech muscles and produce Bob's verbal utterance.
The thought behind the utterance is probably not constructed in speech-specific areas, but instead in deeper layers that might be loosely called higher cognitive areas. These areas are shown schematically in Figure 1 as a parietofrontal network, though presumably the cognitive network is larger and richer. It is labeled here as Item 3. In the modern vocabulary of machine learning, one might say that a set of semantic concepts relating to apples and color have embeddings in this cognitive network. An embedding is a pattern of activity among neurons, a code that in this case stands for semantic concepts. That pattern influences the pattern of activity in the speech network, which in turn drives a pattern of muscle activity, which makes vocal sounds come out of the mouth.
Yet Bob's realization about the apple is not just a semantic concept originating in his higher cognition. Instead, his higher cognitive network is informed by his visual system, labeled here as Item 2. The pattern of activity among its neurons represents visual events, in this case a red apple in a location in front of Bob. Here I will also call that visual embedding a model in keeping with previous terminology. The visual system has constructed a bundle of information, a model, that says in effect, not in words of course but in neurally encoded information, “There's an apple, it's round, it's red, it has a dent here, it has a brown spot there,” and so on.
Ultimately, the stimulus information in the visual system derives from an object in the environment, labeled here as Item 1. Light reflects off the apple, enters the eyes, causes neuronal activity in rods and cones, and initiates a cascade of information into the visual network. It is worth making a point that may seem trivial but will return later. There must be an interface, a set of detectors, to take light from the environment and turn it into an information code; otherwise the brain would not know about the visual stimulus. Bob cannot form semantic concepts about it or talk about it, if it has not been turned into an information embedding. A second point worth making is that Item 2, the visual model, is low level in the sense that it is automatic. Higher cognition receives the visual information but has relatively little internal control over it. Bob cannot choose to make his visual system model a banana when he's looking at an apple. Of course, delusions and hallucinations are possible, but most of the time the model provides a constrained picture of reality. A third point worth making is that the visual model is not fully accurate. It is necessarily a simplification. A neural embedding is an efficient, reduced way of representing something. For example, the apple does not really have a color. It has a complicated spectrum of reflected light. The eyes and brain compress that property of the world into a low-dimensional space that we call color. All models in the brain are schematizations and simplifications of the real items that they represent.
Figure 1 therefore depicts a neuroscientist's narrative, from visual input to motor output. There should be nothing controversial about this account. Even if some of the brain details turn out to be wrong as neuroscience progresses, and even though the reality is more complex and overlapped than depicted here, the general outline is almost certainly correct. There are inputs, information is encoded in progressively deeper layers, and there are outputs, resulting in Bob's specific claim. Equipped with this framework, I will now approach the topic of consciousness.
From Self Model to Speech
In Figure 2, Bob makes a different statement. He says, “Not only is there an apple, but I have a subjective experience of it. There is something it feels like when I see the round shape and the red color. I have phenomenal, conscious experience.” Here I will follow the same conceptual framework as in Figure 1, working backward item by item, to outline how Bob, and philosophers, and me, and the readers of this article, can make that kind of claim.
Highly schematized account of the transformation from an unknown Item 1 to the verbal claim that conscious experience is present. In Hypothesis 1, Item 1 is a magical mind essence. In Hypothesis 2, Item 1 does not exist. In Hypothesis 3, Item 1 is the integration of information. In Hypothesis 4, Item 1 is Item 3, in a recursive loop. In Hypothesis 5, Item 1 is attentional enhancement. In Hypothesis 6, Item 1 is the deep, selective processing of information in the brain, which is associated with integrated information, the global workspace, and attention.
Bob's speech output, Item 4, is caused by patterns of neural activity in his speech network. The activity in the speech network is in turn driven by patterns of activity in his higher cognitive network, Item 3. Yet Bob's realization about subjective experience is not just a passing semantic concept generated by higher cognition. Instead, the higher cognitive network must be informed by a more automatic, lower-level model, an Item 2, that encodes the presence of subjective experience.
One might ask, why do we need a model, a bundle of information that “tells” Bob that he is having a conscious experience? Why can't he just have the conscious experience? The reason is that it is not possible for Bob to construct semantic content, or to say it out loud, if his neural networks do not encode the relevant information. In analogy, if Bob believes there is an apple in front of him, and says there is, then he must have neurally encoded information in his brain that represents that apple. Just so, if Bob believes he has conscious experience, if he says he has it, then whatever conscious experience is or is not, Bob must have a neural embedding that represents it.
We do not know the details of this hidden model that tells Bob about the presence of consciousness in him. In Figure 2, I have centered it on the temporoparietal junction for reasons that emerged from my lab's empirical work (Kelly et al., 2014; Webb et al., 2016a; Wilterson et al., 2021), but the proposed location is not important to the present argument and it could be somewhere else or distributed across many areas. Whatever system constructs it, the information in the model depicts Bob's mental state of experienceness and that information becomes available to the higher cognitive network and from there to the speech network.
To be clear, the account so far is not a theory of consciousness. It is, instead, a framework for thinking about how people semantically know they have consciousness and how they can verbally claim to have it. It is a recognition that because we can talk about consciousness and because we have semantic content about consciousness, the brain must have an information embedding that depicts consciousness. We have yet to address the question: what is the event or object that informs the model, that informs higher cognition, that leads Bob to say he has a subjective, phenomenal experience? What is Item 1? Below I offer six hypotheses.
Hypothesis 1: The Magical Mind and the Problem of Arrow B
In a traditional account, Bob has a “feeling” or experience associated with visually processing the apple—he has phenomenology. There are many proposed theories of how this phenomenal essence is generated by the brain (for review, see Blackmore and Troscianko, 2018; Doerig et al., 2021). All of them face a challenge that must be overcome.
Figure 3 illustrates the challenge. Theories of consciousness almost always focus on how the activity of neurons can produce a conscious feeling, a problem sometimes called the explanatory gap or the hard problem (Levine, 1983; Chalmers, 1995). I call it Arrow A. There is, however, a second, rarely considered explanatory gap. How does the feeling, once it is produced, cause a specific information embedding in Bob's neural networks? I call this second process Arrow B (Graziano, 2013). The concept of Arrow B is in many ways similar to the “hard question” posed by Dennett, “And then what happens?” (Dennett, 2018). The feeling itself, the intangible, what-it-feels-like essence that emerges, would need to physically impact the activity of neurons, revving up this neuron and inhibiting that one, in such a specific pattern, that it imprints exactly the right information to represent itself. Otherwise, Bob's brain would not be able to contain the semantic idea that he has it, or construct the words to report that he has it. To implement Arrow B, neurons would need to have a type of receptor mechanism, never before discovered, that responds to the presence of subjective conscious feeling. To be clear, the challenge is not for these hypothetical neurons to respond to electric fields, or to oscillations, or to quantum states, or to the extent of integration of information, or to whatever physical mechanism one hypothesizes to ultimately cause the conscious feeling. Instead, these hypothetical neurons would need to respond to the emergent feeling itself—the what-it-is-like part. Like the rods and cones in the retina that turn light into neural activity, here you would need neurons that turn conscious feeling into a pattern of neural signals. These consciousness Geiger counters embedded in the brain begin to sound uncomfortably like pseudoscience.
Arrow A and Arrow B. Arrow A is the first explanatory gap. How can physical neurons give rise to an ethereal feeling? Arrow B is the second, equally important, explanatory gap. Once the feeling is generated, how does it then physically impact neuronal activity, such that it can be transduced into a neural embedding or a model, such that higher cognition can form the semantic idea that it is present, such that people can say that they have it?
As shown in Figure 4, without the neurons that have consciousness receptors, without Arrow B, the causal link is broken and the conscious feeling that emerges from the brain cannot be the consciousness that we say we have. Without Arrow B, even if Bob's brain generates a magnificent conscious experience through Arrow A, it would not be the reason why Bob constructs the semantic proposition that he has consciousness, why he is so certain that he has consciousness, or why he says he has consciousness. Even the way he describes consciousness—the specific properties he says that it has—cannot be motivated by the actual properties of the consciousness in his head, as long as that causal link is broken.
The problem of the magical mind theory is therefore not so much Arrow A as it is Arrow B—the second explanatory gap. If a theory does not explain Arrow B—how neurons physically detect consciousness and thereby turn it into an informational embedding—then you do not have a working theory because it does not explain the known behavior.
It is powerfully tempting to say, “Bob believes he has consciousness, and claims to have consciousness, because he actually does have it.” Who can argue with that? But spend half a minute thinking about it, and you should realize that it does not work. Consider an analogy. Your brain has a complex pattern of blood flow, but as far as is currently known, does not translate that pattern into an information embedding. Or at least, if it does so, that model is not accessible to higher cognition. You cannot introspect and say, “My brain has a really nice blood pattern right now.” You do not have that automatic, internal model, moment-by-moment informing your higher cognition. Because blood flow is a physical process that could, in principle, be detected and transduced into a pattern of neural activity, this type of introspective insight could in principle have evolved. It appears not to have, since we don't have it. The lesson of the blood flow example is that, just because your brain has something in it, does not mean you know it, believe it, think it, or claim it. You can do those things only if the item in question is physically translated into a neuronal embedding, such that specific information is encoded in a specific way in the right neural networks. It does not work to say, “Bob believes and claims he has conscious experience because he does have it.” That argument is neither meaningful nor valid. It is magical thinking. It treats Bob's claim that he has consciousness as if it follows different rules from every other claim that Bob makes.
Could the magical mind theories be correct anyway? Suppose, for the moment, that Figure 3 is an accurate outline. By way of Arrow A, neurons produce a conscious feeling, a subjective experience, as almost all accounts assume. Suppose that some as-yet unknown mechanism for Arrow B exists. Conscious feeling itself somehow impacts neural networks, creating a specific information code, or model, informing Bob's cognitive machinery and his speech network, such that he can say that a conscious feeling is present. Assume for the moment that this account is correct. Even so, the feeling that Bob “knows” he has and says he has is entirely dependent on the information embedding—on Item 2. Whatever the model describes, that is what Bob is certain he has. There is no reason to suppose that the model is accurate. In fact, the opposite—it appears that all models in the brain are schematic, simplified, and inaccurate. Therefore, even if we accept the magical mind hypothesis, it is still the case that the consciousness arising from the brain is different from the kind of consciousness that Bob describes having. The conscious experience that you “know” you have, that philosophers are certain they have, that people tell each other they have and write books about having, is a phenomenon as it is described by a sketchy informational embedding. The nature of the real item, the real consciousness, remains unknown. The only description we have of it comes from an indirect route—through an imprecise model that informs a higher cognitive network.
The causality is broken without Arrow B. If the brain has no mechanism for detecting conscious feeling, such that it can no longer be represented in a neural code, then people have no way to think about or talk about that consciousness. Without Arrow B, when people form the semantic idea that they have conscious experience, when they say they have conscious experience, and when they describe specific properties and features of conscious feeling, none of those events are based on an actual conscious feeling generated by the brain. Theories that propose an Arrow A but ignore Arrow B are not explanatory.
Therefore, even if we try to make the traditional account work, the real consciousness that Bob actually has could just as well be something unexpected, different in its intrinsic properties from what either Bob or we believe it is—as different as, for example, the real city of New York is from a paper map of the city. Even the traditional account, when held to a basic standard of logic, turns into an illusionist account.
It is an uncomfortable insight, but the fact that Bob and you and I all “know” that we are having a phenomenological conscious experience right now is not evidence that we have it. Instead, it is strong evidence that our brains have an information embedding, a model, that depicts something about ourselves. The depiction may be so schematic and distorted that we are left not understanding the item that is being modeled.
If the magical mind theories are so irrational and riddled with not one but two fatal gaps, then why are they so popular? There is a reason why the magical mind theories are universally compelling and capture the majority of the field of study. They are verbalizations of, or at least they are consistent with, our automatic internal models. They resonate with our intuitions. People's convictions fit easily into that way of thinking and rebel at the alternative. There may also be a defensive component: magical mind theories help to protect the mystery of the human spirit against materialism and maybe especially against the uneasy growth of technology. Artificial deep neural networks in particular, which are capable of such spectacular successes by imitating aspects of the brain, turn out to be linear algebra machines, and it violates our natural intuitions—our self models—to think that the human brain is a collection of biologically-instantiated algebra machines. Scholars are therefore powerfully, ideologically motivated to search for magical theories of consciousness. I will go so far as to say that, in my view, the field of consciousness studies has become a Trojan horse for the entry of overt spiritualism into the discussions, the conferences, the journals, and the funding opportunities of professional science.
Hypothesis 2: Hard Illusionism
Here I will describe the most extreme form of illusionism. Though I think it is unlikely to be correct, it helps define the conceptual boundaries.
The brain builds models of real things. The visual system builds models of objects in the world. The auditory system builds models of sounds. The body schema is a model of the physical body. These bundles of information represent things in the world that are useful to monitor and predict. But suppose that the brain anomalously builds a model that depicts Bob as having a conscious experience, and causes Bob's higher cognitive machinery to contain the semantic information that he has a conscious experience, and causes Bob to say that he has a conscious experience—but the model does not actually correspond to anything real? What if there is no Item 1? What if Bob's consciousness is entirely a fiction, depicted by an errant model? Maybe Bob learned this false model through cultural upbringing. Everyone else thinks they have consciousness, so Bob does too. In that case, consciousness is not just a false construct, but also limited to humans.
This hard illusionist account is not inconsistent with Bob's behavior, and therefore it should be considered as a possibility. However, for my part, I find it difficult to see how such a model would have emerged. Evolution shapes the brain to construct models that can monitor and predict real items in the world. It seems unlikely that evolution would have shaped the brain to waste its energy building models of things that do not actually exist. As far as is known, a similar model is present in almost all human brains across cultures. If it is an anomalous model of nothing at all, then the sheer scale and consistency of the bamboozlement seems improbable.
A less extreme form of illusionism might be more likely. Suppose there is an Item 1—there is a real thing that is being modeled—but the model is schematic and imperfect. Therefore, the consciousness that Bob describes having, the phenomenology or the intangible experience that he believes he has is somewhat different from the actuality of Item 1. This distinction between a model in the brain and the item that it models is necessarily true, as all neuroscientists should know. Visual models, somatosensory models, cognitive models, self models, and all models in the brain are reductions or informational embeddings that simplify and compress. Subtle illusionism is necessarily true of all models in the brain and must be true of consciousness as well. The following hypotheses are consistent with this subtle illusionism.
Hypothesis 3: The Illusionist Version of the Integrated Information Theory
In the integrated information theory (IIT), consciousness occurs when a system contains information that is in a state of integration (Tononi, 2008; Tononi et al., 2016). A mathematical metric for integration, phi, has been proposed. Whether any infinitesimal degree of integrated information can result in consciousness, or whether a threshold must be reached, is debated. Although the theory contains many facets, I will not discuss them here.
By the terminology outlined in this article, IIT is a magical mind theory. First, it assumes that a subjective, phenomenal experience, as we report it, exists. Second, it proposes a condition under which that conscious experience emerges; it proposes an Arrow A. Third, the theory lacks a specific proposal for Arrow B, or for how the conscious feeling, having emerged, has the physical ability to imprint itself on neuronal activity, such that it can become represented in Bob's semantic content and verbal output.
Here I will outline a simple addition to IIT that converts it from a magical mind theory to an illusionist theory (i-IIT). It is unlikely that the proponents of IIT would agree with i-IIT, because the illusionist version specifically removes the magicalism that makes IIT intuitively comfortable. However, i-IIT retains the central observation that the property we call conscious experience has similarities to the state of integration of information. It also has the advantage of solving several problems of IIT that have caused some scholars to call it nonscientific (IIT-Concerned et al., 2023).
Suppose the level of integration of information within the brain's neural networks is indeed a coherent, measurable property (Item 1). If that property is physically measurable, then it can in principle be detected by neurons and turned into a model (Item 2) that depicts the current state of integration. That model is, as always, schematic. It is a reduction or simplification. The model informs higher cognition (Item 3), which can inform the speech network (Item 4). In this hypothesis, Bob says that he has a subjective, intangible, phenomenal experience, because that is the schematic, reduced way in which Item 1 has been represented. In this modified version of IIT, one could not accurately say that integrated information is consciousness. Instead, a more accurate and careful description would be that integrated information is Item 1 in the chain of events that leads people to claim they have consciousness. In this hypothesis, consciousness resembles integrated information because consciousness is a distorted representation of integrated information.
One advantage of i-IIT is that it does not have a magical, intangible essence at the heart of it. The hard problem is dissolved. The subjective, phenomenal experience that Bob describes is a distorted construct. A schematic model represents integrated information in that distorted manner and therefore that is the property that Bob claims to have. Without the model, integrated information by itself would have no relationship to consciousness.
A second advantage of i-IIT is that it solves the problem of panpsychism (Pennartz, 2022). Since all things in the universe contain some degree of integrated information, therefore, in at least some interpretations of IIT, all things have consciousness. From a spiritual, religious, or wish-fulfillment perspective, panpsychism may be a desirable feature—and many scholars are especially drawn to this aspect of IIT. But for a scientific theory of consciousness, the claim that all things in the universe are conscious is unhelpful as well as untestable. It renders consciousness an essentially meaningless concept. In i-IIT, the problem of panpsychism disappears. Many objects in the universe contain integrated information, but without the associated neural embeddings and network processing, those objects have no relationship to consciousness.
For all its advantages, i-IIT does have potential limitations. First, it is not clear if the state of integration of information in the brain can be detected by neurons, such that it can be rendered into a model. Second, the adaptive use of the model is not clear. The brain builds models of items that are useful to monitor and predict. For example, the brain builds a model of the arm—the arm schema—because monitoring and predicting the state of the arm is useful in the control of movement. But what is the selective pressure for the brain to “know” about its own state of integrated information? Why would that model be constructed? It is possible that a continuously computed model of the state of integration of information in one's own brain would serve an as-yet undetermined, useful cognitive function. But the utility argument may be stronger for the hypotheses proposed next.
Hypothesis 4: The Illusionist Version of the Global Workspace Theory
In the global workspace theory (GW), a select set of information in the brain reaches a state of signal enhancement, sometimes called ignition, in which the information is able to be broadcast widely to affect many systems around the brain (Baars, 1988; Dehaene, 2014; Mashour et al., 2020). This select set of information is said to be in the global workspace and by being in that state, has entered consciousness. In the global neuronal workspace theory (GNW), the workspace is centered specifically on the brain's parietal-frontal networks (Dehaene, 2014).
GW, at least in its simplest form, is a magical mind theory. Once the apple information reaches Bob's global workspace, it enters his consciousness. In this way, GW proposes Arrow A that gives rise to a mental experience. But where is Arrow B? How does Bob's brain acquire the semantic content, or make the verbal claim, that he is having a subjective experience? Another way to put the difficulty is that GW corresponds to the condition shown in Figure 1. Information about the apple is processed in visual cortex, enhanced by attention mechanisms to the extent that it reaches the parietofrontal networks, and from there it can impact downstream systems such as speech networks. This diagram shows how Bob processes visual information and can therefore say, “There is an apple and it is red.” However, GW does not encompass the condition shown in Figure 2. The information in the global workspace is all about the apple. Where did Bob's information about experienceness come from? How does he know to say, “And I have a subjective experience of the apple”?
These difficulties are immediately solved if we convert GW into an illusionist account. Suppose that in addition to having a global workspace, Bob's brain also constructs a model of its own global workspace. Since the global workspace has outputs that affect other parts of the brain, it is possible in principle for a system in the brain to gain information about and build a model of the global workspace. The model is, as always, schematic. It is a reduction or simplification. Bob says that he has a subjective, intangible, phenomenal experience, a theater of consciousness into which the apple or other objects can enter, and the theater of consciousness empowers him to choose and act and speak. Bob “knows” this is true, and says it, because that is the schematic, reduced way in which that model represents his global workspace. In this illusionist version of GW (i-GW), the global workspace is Item 1. The model, Item 2, is responsible for adding the more magical, phenomenological spin to the account that Bob ultimately gives.
A curious feature of i-GW is that Item 3, the higher cognitive network, is also the global workspace. Item 1 (the item being modeled) is the global workspace; Item 2 (the model) is a depiction of the global workspace; and Item 3 (the cognitive network that receives information from lower-level models) is, once again, the global workspace. In a somewhat mind-bending recursive twist, the global workspace contains simplified information about itself. It contains information about the apple and about how the global workspace contains the apple. The global workspace influences other systems such as speech output and it contains information about how it influences output systems. This interesting recursive structure explains how Bob can report on the apple (since the visual apple information has entered the global workspace) and yet also, separately, report on his state of experience of the apple (since the information about “experienceness” has also entered the global workspace). It is the key to understanding how Bob can say, “There is an apple,” as shown in Figure 1, and yet also be able to say, “I have a subjective experience of it,” as shown in Figure 2. Without that recursive property—if the global workspace did not contain a model of itself—we would be creatures that can talk about apples and other specific external content, but would never know about or talk about our own consciousness. With it, we become the philosopher creatures that we are, who can take notice, ponder, and argue about consciousness.
What is the adaptive advantage of building a model of the global workspace? Why would it have evolved in the first place? The answer may be that any controller performs better if it has a model of the thing it controls. In the end, the brain is a controller of behavior, and the global workspace is almost 100% determinative of behavior. Items that get into the global workspace can influence behavior, and items that do not rise to the level of the global workspace have at most a peripheral impact on behavior. Therefore, to guide one's own behavior, it would be useful to build a predictive model of one's own global workspace. It is, in effect, a model of me, of the most central, determinative parts of my brain's functioning, a model of what it means for me to fully know things in the moment and to make decisions and choose actions. The model, being schematic and incomplete, does not depict me as a complex biological machine. It instead depicts me as a magical essence inhabiting my body, capable of subjective, phenomenal experience and of a conscious will. The model tells me that I have consciousness because that is a useful, simplified way of keeping track of my global workspace.
One might ask, granted that the model is useful for the brain to construct, nonetheless, how does a person become conscious of the model? In an illusionist theory, the answer is that people do not become conscious of the model. Instead, the model contains information. The information enters higher cognitive networks and affects downstream systems, causing the person to semantically believe, and claim, whatever is depicted in that model—in this case, that a state of conscious experience is present.
Hypothesis 5: The Attention Schema Theory
Each illusionist hypothesis described above is a guess about the identity of Item 1, the unknown item that is represented by a model (Item 2) in Bob's brain, leading to a semantic embedding in Bob's higher cognitive networks (Item 3), and from there to a speech output (Item 4) in which Bob claims that he has consciousness. So far, the proposed identities of Item 1 include an actual, magical mind (deemed unlikely here); nothing at all (the hard illusionist perspective, also deemed unlikely); the amount of integration of information in Bob's brain; and a global workspace. Can we find a systematic method for querying a model in the brain and identifying the thing that it models?
Take a well-understood case: the brain's model of the arm. The brain builds a body schema automatically and continuously (Head and Holmes, 1911; Graziano and Botvinick, 2002; Holmes and Spence, 2004). That model has been studied for more than a century, but how can one be sure that it actually represents the body, or that one particular part of it represents the right arm as opposed to the left leg or the head? The main reason is that when Bob closes his eyes and says, “My arm is up in the air right now,” it turns out that his arm is, usually, up in the air. The model covaries with the arm.
Consider the following simple hypothetical procedure. A scientist asks Bob to close his eyes and report what he can about his body. Bob says, “I have a central trunk and a set of appendages.” By asking Bob to introspect, the scientist is gaining information via an indirect route. The information inside a deeper model, the body schema, is passed through Bob's higher cognition and then his speech networks, to be received by the scientist. The content of the speech output may not be a perfect reflection of the deeper body schema—it may be an imperfect throughput—but it is one of the few easy ways for the scientist to listen in on Bob's body schema.
The scientist says, “Bob, tell me about one of those appendages. Pick one – I want to see if this internal model of yours matches anything externally measurable.”
Bob says, “I have an appendage attached to my right shoulder. It's about a meter long. It has a joint in the middle and little sticks at the end.”
The scientist excitedly measures Bob's right arm and says, “That description at least superficially resembles a real object. Tell me more.”
Bob says, “The appendage can move around. Now it's down at my side. Now it's up in the air.”
The scientist watches the movement of the arm and scribbles on his notepad, saying, “Yes, that internal model in your brain, as filtered through your cognition and verbal report, appears to covary with a real, measurable object – your right arm. Clearly, it's a model of your right arm.”
A second scientist, who is more skeptical, says, “But is it really a model of the arm? The arm has muscles, blood vessels, and bones. Bob, can you introspect in more detail and accurately tell me about those details of your arm?”
Bob says, “Introspectively, I don't know about those granular details. I can only tell you about the overall shape and position of this appendage.”
The skeptical scientist says, “Clearly the model lacks many of the details that the real arm has. And even worse, I can devise a situation in the lab such that the real arm is angled to the right, but Bob says it's to the left. Or, worse yet, if we cut Bob's arm off entirely, we know he's likely to have a phantom limb, i.e., the model persists even when the arm is gone. Therefore, that model doesn't always perfectly track the state of the arm. Are we sure it's a model of the arm? Maybe it's a model of something else. Is it a model of a leg? A model of somebody else's arm?”
A third scholar, who is more mystical-minded, chimes in. He says, “I think that Bob has an ethereal, invisible arm made of ghost essence. Sure, the ghost arm happens to move around roughly as his physical, objectively measurable arm moves. But the ghost arm is a separate thing. That's why it's still there after the real arm is amputated. Maybe it's a quantum vibration. Maybe it's an electric field, or a new ghost field that should be added to the standard model. The ghost arm is currently beyond the ability of science to measure, but Bob can nonetheless introspect and tell us about it – because, for reasons unknown, the ghost arm can physically impact Bob's cognitive and speech networks.”
I will risk the bold assertion that the mystical scholar is wrong. Yet the correct answer is not perfectly straightforward either. The skeptical scholar has a point—the real arm as measured objectively, and the arm that Bob describes on the basis of introspection, are not identical. How can we make sense out of the confusion? The answer is to adhere to a few principles. The brain's internal models are schematic and imperfect. Therefore, Bob's introspective account, informed by his arm schema, will not perfectly match the real arm. The match is still apparent—but it is just not perfect. The arm that Bob describes resembles the real arm in that both have a joint in the middle and a hand at the end. Moreover, the model accurately tracks the movement of the real arm a good 98% of the time. Therefore, it is a reasonable inference that when Bob closes his eyes and reports on the appendage sticking out of his right shoulder, that report derives from an internal model of Bob's actual right arm. It's not a model of a leg after all, because the properties it depicts do not resemble a leg and it does not covary with either of Bob's legs.
The logic is so obvious in the case of the arm that my explanation reads like an unnecessary belaboring of the point. Maybe the logic can be equally obvious in the case of consciousness.
As diagrammed in Figure 2, there is an unknown Item 1; it is represented by a model, Item 2; the model informs higher cognitive networks, Item 3; and higher cognition informs the language network, Item 4. When Bob verbally describes his conscious experience, he is providing information from the output end of that sequence. But how do we identify Item 1, at the start of the sequence? We’re looking for something with three diagnostic properties. First, Item 1 should have the physical ability to impact the activity patterns of neural networks. Otherwise, there would be no mechanism for detecting it and creating a model of it that can then affect cognition and speech. Second, Item 1 should resemble the consciousness that Bob describes having, at least in its superficial properties. We do not expect it to perfectly match conscious experience, because no model is a perfect description. In particular, we should expect the actual Item 1 to contain granular, physical details that are missing from Bob's account of consciousness. Despite those differences, the general, functional outlines of Item 1 should resemble the outlines of conscious experience as Bob describes it. Third, the real Item 1 should change state as Bob's account of his consciousness changes. It should covary with Bob's reported consciousness almost all of the time. If we can find something that satisfies those three constraints, then we could reasonably conclude that we have found Item 1, the real item that is the ultimate cause of Bob's claim that he has conscious experience.
I suggest there is a well-known item that matches those characteristics. Selective attention is a complex neural process, mainly in the cortex and thalamus, in which chunks of information that are encoded in neural networks compete with each other for greater signal strength (Desimone and Duncan, 1995; Kastner and Pinsk, 2004; Beck and Kastner, 2009; Moore and Zirnsak, 2017). Attention on the apple means the apple's representation in the brain is enhanced and can have a bigger downstream effect on the rest of the brain.
Attention appears to match the three diagnostic properties described above. First, because attention is a physical process that can have a measurable effect on neurons, it can be detected or transduced into an informational embedding. The brain would be able to construct a descriptive model of its own state of attention. Second, in their superficial properties, attention and consciousness resemble each other. The similarity is not perfect. Attention, after all, involves granular physical details such as neurons, synaptic inhibition, and specific pathways through the brain. Bob, describing his consciousness, might say that it is a what-it-feels-like essence, a vividness or immediate presence, that has no obvious physical moving parts—and that once something is in his consciousness, he can then make decisions about it. Bob's description of consciousness reads like a high-level, somewhat abstracted description of the mechanistic process of attention.
As for the third diagnostic property, we now have a hundred years of data showing that selective attention covaries with people's claims of conscious experience (James, 1890; Posner, 1994; Merikle and Joordens, 1997; De Brigard and Prinz, 2010; Prinz, 2012). If Bob says he was not conscious of the highway sign, he almost certainly had no attention on it, a phenomenon called inattentional blindness (Mack and Rock, 1988; Simons and Chabris, 1999; Drew, et al., 2013). If Bob says he was conscious of the highway sign, he almost certainly had at least some attention drawn to it, if only partial and peripheral attention. Attention and conscious experience covary so closely that it is difficult, though possible, to design a laboratory condition in which they are pried apart (McCormick, 1997; Kentridge et al., 1999, 2008; Norman et al., 2013; Webb et al., 2016b; Wilterson et al., 2020).
I suggest that the relationship between attention and consciousness has been staring science in the face for decades. But for the allure of mystical thinking and prior assumptions, it would be as straightforward as the relationship between the arm and the arm schema. Others have noted that attention and consciousness tend to covary (James, 1890; Posner, 1994; Merikle and Joordens, 1997; De Brigard and Prinz, 2010; Prinz, 2012). Here I offer this simple explanatory hypothesis. Attention is Item 1. It is a process of selectively enhancing some representations over others. Bob can pay attention to color and shape, to a sound, to a thought, to a memory, and so on. The brain not only uses the process of attention but also constructs a descriptive model of its own state of attention, which is functionally useful for the control of attention. That model, the attention schema, is Item 2.
Just like a visual model depicts the apple, and the body schema depicts the body, so the attention schema depicts attention in an automatic manner. Bob does not intellectually choose to think of himself that way. He can't help it. The attention model happens automatically and continuously. It is also not a detailed or accurate account of the mechanics of attention. There is no evolutionary pressure for the brain to know about neuronal competition, synaptic inhibition, the dorsal attention cortical pathway, or other mechanistic details. What this model depicts, instead, is an amorphous, mental essence, a mind vividly possessing something in the moment. That simple depiction works well enough for keeping track of attention. The model can then inform Item 3, the higher cognitive network. Higher cognition can influence Item 4, the speech networks. As a result, Bob claims to have conscious experience.
The attention schema is named in analogy to the body schema, and the analogy may involve more than a superficial similarity. Just as the brain needs a model of the arm to help control the physical appendage of the arm, which moves around and grasps objects, so it may need a model of attention to help control the virtual appendage of attention, which also moves around and grasps objects in its own way. In that sense, the attention schema is an extension of the body schema. Connections between the body schema, embodiment in general, and consciousness have been proposed before (Damasio, 1999; Blanke and Metzinger, 2009).
It has also been noted (Graziano et al., 2020) that AST is a version of a higher order thought theory (HOT). In HOT, the brain is conscious of X if the brain generates a higher order thought about X (Rosenthal, 2006; Gennaro, 2012; LeDoux and Brown, 2017). A higher order thought, or a metacognitive thought, is, one might say, the process of thinking about thinking. AST is a form of HOT because it postulates a model of attention. Attention is a cognitive process; constructing a representation of something is a cognitive process; and therefore to construct a model of attention is to apply one cognitive process to another. In AST, the visual model of the apple is a lower or first order representation, and the model of the act of attention on the apple is a second order or higher order representation. The model of attention is, however, not higher order in the sense of an explicit, or volitional thought. It is not that people think to themselves, “Let's suppose I’m conscious today.” Instead, the model of attention is just as automatic as the body schema, or as a visual model. When Bob looks at an apple and deeper systems in his brain supply his higher cognitive networks with a picture of reality in the moment, that picture includes a model of an apple and a model of attention. The picture of reality says, “There is an apple, and I have a conscious experience of it.”
One might call AST an illusionized version of HOT, or i-HOT. In AST, a conscious experience does not emerge from having a higher order thought. Instead, all claims that a person makes depend on information in the brain, and the model of attention is the specific bundle of information that serves as the basis for the claim, “I have a conscious experience.”
Hypothesis 6: Deep Illusionism, or Many Keyhole Perspectives on a Deeper Theory
The most important suggestion in this article may be that many of the disparate hypotheses described here are not actually separate. Though they come from different points of origin, they converge on a deeper concept. The reason for that convergence is that a deep connection exists between selective attention, the global workspace, and the integration of information.
Attention and the global workspace are closely related. Attention involves the enhancement of representations in the brain, which then have a bigger effect on downstream systems. Attention can operate at many levels in the visual hierarchy, including the thalamus, primary visual cortex, and many higher layers of cortical processing (Desimone and Duncan, 1995; Kastner and Pinsk, 2004; Beck and Kastner, 2009; Moore and Zirnsak, 2017). Similarly, in GW, attentional enhancement boosts information to the highest hierarchical levels of processing in parietofrontal networks, where it can have a bigger effect on downstream systems. The global workspace and attention are inseparable phenomena. In that perspective, a model of the global workspace is really a type of model of attention, and as such, i-GW and AST are really just different versions of the same theory.
The integration of information is also closely related to attention. As known since Treisman's pioneering work on feature integration theory, information becomes integrated in the presence of attention (Treisman and Gelade, 1980). When you attend to a visual stimulus, the colors and shapes, patterns and textures are integrated into a single object. The parts click together. When no attention is directed to the object, the visual components come apart and are no longer processed in an integrated manner. The color, shape, texture, motion, and so on become computationally separated from each other.
One could validly argue that attention, the global workspace, and integrated information are different descriptions of the same underlying object—the deep, selective processing of information in the cerebral cortex. It is like the proverbial blind men feeling the elephant, where one feels the trunk, another feels the body, and a third feels the tail. Suppose you take a photo of the elephant. Do you now have a picture of a trunk? Of a body? Of a tail? They are all components of the original object, and you will find them all represented in the picture. The “elephant” in this case is the deep selective processing of information in the cerebral cortex. The “picture” is the schematic model that the brain constructs. It should not be a surprise if scholars, trying to interpret that picture, say, “Consciousness resembles attention; but it also resembles the global workspace; and yet, again, it resembles integrated information. It also seems to covary with all three of those.”
Given all of these considerations, I propose here a final, combined, illusionist hypothesis: Deep Illusionism. In this hypothesis, Item 1 is the deep selective processing of information in the brain. The brain also builds a simplified, schematic model of that deep processing (Item 2). The model allows the brain to monitor, predict, and better control its own deep processing. The model is a source of input to higher cognition (Item 3). Higher cognition informs the linguistic network (Item 4). At the end of that game of telephone, in which the original Item 1 has been coded and recoded, what comes out the other end is Bob's account of his own conscious experience.
If this is how Bob is constructed, if this is how we are all constructed, then it is understandable if a camp of scholars insists that consciousness resembles the global workspace. After all, based on introspective access to that model, everyone believes that consciousness is like a theater of the mind, the stage for all the things we think about and choose to act on. Consciousness and the global workspace share so many properties, it is intuitive to suggest that consciousness simply is the global workspace (Baars, 1988; Dehaene, 2014; Mashour et al., 2020). Another camp of scholars might say that consciousness resembles attention. After all, consciousness is when the mind seizes possession of something in an intensified manner. It shares so many of the properties of attention that maybe consciousness is attention (James, 1890; Posner, 1994; Merikle and Joordens, 1997; De Brigard and Prinz, 2010; Prinz, 2012). A third camp of scholars might say that consciousness resembles integrated information. Everyone knows that consciousness is both rich and at the same time unified; and an integrated set of very diverse information is by definition both rich and unified. Therefore, maybe consciousness is integrated information (Tononi, 2008; Tononi et al., 2016). Yet another camp might argue that based on introspection, everyone knows that consciousness is a feeling, not a physical thing, and therefore no scientific theory can ever bridge the gap from mechanism to an intangible, phenomenal experience—the hard problem (Chalmers, 1995). The reason why these different camps arrive at their many suggestions is that they are all, in their own ways, examining the same, schematic, internal model, and they are noting different, though related features of it.
The deep illusionist hypothesis is therefore not just a theory of consciousness. It is also a fairly effective theory of the broader landscape of theories. It helps explain why scholars are prone to arrive at the proposals they do. A swarm of theories, each different from the others, some of them trending more materialistic and some more magical, all emerge from partial, clouded glimpses of the brain's picture of the elephant. They emerge from Item 3 (higher cognition) accessing Item 2 (a deeper, automatic model), which is itself a distorted, detail-poor picture of Item 1 (the deep, selective processing of information).
It is tempting to ask where, in this framework, consciousness resides. Is it Item 1? Item 2? Is Item 3 necessary for consciousness? Does consciousness emerge from the interaction between these items? What are the necessary and the sufficient conditions for consciousness? However, the question itself is poorly formulated and misses the point of the illusionist framework. The framework proposes that we are agents that say we have conscious experience; we say it based on semantic embeddings in higher cognitive networks that encode the presence of conscious experience; the semantic embeddings are based on information from an automatic, deeper model; and the deeper model provides a distorted, detail-poor description of real physical processes in the brain. One cannot point to a part of that framework and say, “That's where the consciousness essence lives,” any more than one can point to a part of a car and say, “That's where the drive-around essence lives.” Illusionism explains the behavior of the system without postulating the existence of a consciousness stuff that emerges from or resides in any part of the system.
Why a Mechanistic Theory of Consciousness Is Important
In my view, there are now only two broad classes of explanation for consciousness: magic and illusionism. Most of the field of study is mired in magic. The purpose of this article is to add at least one small push back toward rationality. It is difficult to convince others who are already invested in a way of thinking, but it may be possible to reach those from outside the field and show them that the study of consciousness has options for a clear-headed scientific approach.
That scientific approach may be important for our technological future. Everything discussed in this article could have applied to artificial as well as biological consciousness. In the illusionist framework, Bob could just as well be a deep learning neural network. If it is possible to build artificial consciousness, then it is going to happen, if it is not already here. We will have to decide what to do with a world in which consciousness is not just in us but also in our everyday technology. According to some speculations, artificial consciousness is an important step toward building prosocial AI that is more cooperative (Graziano, 2019). According to other speculations, consciousness will render AI more likely to harm us (Fast and Horvitz, 2017; Leffer, 2023; Pesen, 2023). In yet a third perspective, artificial consciousness will turn technology into an object of moral consideration, a so-called moral patient, greatly complicating how humans should behave toward it (Metzinger, 2013; Ryland, 2021; Akst, 2023).
This article is not intended to discuss the very complicated topic of artificial consciousness or its social or moral implications. My final point is simply that it is now a matter of urgency that we dispense with the magic, the mysticism, the pseudoscientific populism and the wishful thinking, and instead understand what consciousness actually is at the mechanistic level, including what benefits and risks it unlocks. Understanding consciousness is necessary for our immediate future. For the past century, the science of consciousness has focused on intellectual, philosophical and even religious goals, as researchers hoped to gain insight into the human condition. However, like the invention of stone tools, controlled fire, or writing, understanding consciousness at the level of engineering is likely to have an explosive effect in a practical sense that dwarfs the philosophical issues.
Footnotes
The author declares no competing financial interests.
Research supported by grant 24400-B1459-FA010 from AE Studios.
This is an open-access article distributed under the terms of the Creative Commons Attribution 4.0 International license, which permits unrestricted use, distribution and reproduction in any medium provided that the original work is properly attributed.