REVIEWS H) Check lor updates Theories of consciousness Anil K. SethO'-2^ and Tim Boyne23-4 Abstract | Recent years have seen a blossoming of theories about the biological and physical basis of consciousness. Good theories guide empirical research, allowing us to interpret data, develop new experimental techniques and expand our capacity to manipulate the phenomenon of interest. Indeed, it is only when couched in terms of a theory that empirical discoveries can ultimately deliver a satisfying understanding of a phenomenon. However, in the case of consciousness, it is unclear how current theories relate to each other, or whether they can be empirically distinguished. To clarify this complicated landscape, we review four prominent theoretical approaches to consciousness: higher-order theories, global workspace theories, re-entry and predictive processing theories and integrated information theory. We describe the key characteristics of each approach by identifying which aspects of consciousness they propose to explain, what their neurobiological commitments are and what empirical data are adduced in their support. We consider how some prominent empirical debates might distinguish among these theories, and we outline three ways in which theories need to be developed to deliver a mature regimen of theory-testing in the neuroscience of consciousness. There are good reasons to think that the iterative development, testing and comparison of theories of consciousness will lead to a deeper understanding of this most profound of mysteries. Neural correlates of consciousness (NCCs). The minima] set of neural events that is jointly sufficient for a conscious state. 'Department of Informatics and Sackler Centre for Consciousness Science, University of Sussex, Brighton, UK. 2Program on Brain, Mind, and Consciousness, Canadian Institute for Advanced Research (CIFAR), Toronto, Ontario, Canada. 3School of Philosophical, Historical, and International Studies, Monash University, Melbourne, Victoria, Australia. "Monash Centre for Consciousness and Contemplative Studies, Monash University, Melbourne, Victoria, Australia. Me-mail: a.k.seth® sussex.ac.uk https://doi.org/10.1038; 541585-022-00587-4 In the early decades of its resurgence, the scientific study of consciousness focused on the search for the 'neural correlates of consciousness' (NCCs). Formally, the NCCs of a conscious state are the minimal set of neural events jointly sufficient for that state; in practice, the search for the NCCs has involved seeking the brain states and processes that are most closely related to consciousness1"3. Focusing on searching for the NCCs has been useful because the notion of NCCs is relatively 'theory neutral', and thus the NCC framework provided a common language and methodology for researchers with different theoretical and even metaphysical commitments. However, the limitations of the NCC framework have become increasingly clear, as revealed for example in the challenges involved in distinguishing 'true' NCCs from the neural prerequisites and consequences of consciousness4"7. In response to these limitations, there has been a steadily increasing focus on the development of theories of consciousness (ToCs). With a ToC in hand, we would be able to go beyond an NCC-based methodology and move towards models of consciousness that deliver explanatory insight. Indeed, having an empirically validated ToC should be the primary goal of consciousness science8'9. Whereas the NCC approach prioritizes the search for correlations between brain activity and consciousness, a theoretical approach instead focuses on identifying explanatory links between neural mechanisms and aspects of consciousness10. That being said, theorists often employ different conceptions of what it would take to secure an explanatory link between neural activity and consciousness. Some assume that a satisfactory ToC should and can close the 'explanatory gap' (BOX 1), and that it will be possible to render the relationship between neural activity and consciousness as transparent as the relationship between water's chemical structure and its gross behavioural profile11. Others doubt or remain agnostic as to whether the explanatory gap will ever be fully closed, but nonetheless hope for a framework that might explain certain aspects of consciousness and, in so doing, reduce or eliminate the sense of mystery surrounding its biophysical basis1213. Still others argue that explanatory gap intuitions are misleading, and should not be taken seriously by the science of consciousness1415. There is now a wide range of candidate ToCs (TABLE 1). Notably, instead of ToCs progressively being 'ruled out' as empirical data accumulate, they seem to be proliferating. This proliferation has led to both attempts to integrate existing theories with each other16 and the development of'adversarial collaborations', in which proponents of competing theories agree in advance about whether the outcome of a proposed experiment will support or undermine their preferred theory . However, there are significant challenges to both theory integration and adversarial collaboration, as we discuss. NATURE REVIEWS | NEUROSCIENCE VOLUME 23 | JULY 2022 | 439 REVIEWS Box 11 Theories of consciousness and the 'hard problem' In the 1990s, David Chalmers famously distinguished between the 'hard' and 'easy' problems of consciousness164. The easy problems are concerned with the functions and behaviours associated with consciousness, whereas the hard problem concerns the experiential (phenomenal, subjective) dimensions of consciousness. What makes the hard problem hard is the 'explanatory gap'165 — the intuition that there seems to be no prospect of a fully reductive explanation of experience in physical or functional terms. Some theories of consciousness (ToCs) (for example, integrated information theory (NT) and certain versions of higher-order theory (HOT)) address the hard problem directly. Other theories (for example, global workspace theories (GWTs)) focus on the functional and behavioural properties associated with consciousness; although they can be viewed as addressing the hard problem, this is not the primary goal of their proponents. A third strategy (adopted by some predictive processing theorists) aims to provide a framework in which various questions about the phenomenal properties of consciousness can be addressed, without attempting to account for the existence of phenomenology as such67 — an approach sometimes called the 'real problem'13166. A critical question in this area is whether the hard problem is indeed a genuine challenge that ought to be addressed by a science of consciousness, or whether it ought to be dissolved rather than solved. Those who take the latter view often argue that the appearance of a distinctively hard problem derives from the peculiar features of the concepts ('phenomenal concepts') that we employ in representing our own conscious states167 168. A related view is illusionism, according to which we do not actually have phenomenal states but merely represent ourselves as having such states1415. Whatever the merits of these proposals, it seems likely that the grip of the hard problem may loosen as our capacity to explain, predict and control both phenomenological and functional properties of consciousness expands166 169. In this Review, we consider how a range of ToCs relate to each other and to empirical data, and we identify some promising avenues by which theory development and empirical research can jointly support each other in the search for a satisfying scientific account of conscious experience. Our attention is restricted to theories that are either themselves expressed in neuro-biological terms or are plausibly taken to entail claims that can be expressed in neurobiological terms. (As we will see, some neurobiological' ToCs are expressed in the abstract language of functional relations or information theory, and qualify as neurobiological' only because the abstract features that they appeal to are associated with particular neural mechanisms.) We also consider only neuroscientific theories that are consistent with known physical theory, and we also leave to one side theories that link consciousness directly to quantum mechanical processes (for examples, see REFS1819). Preliminaries One of the main reasons why ToCs 'talk' past each other is that they often have different explanatory targets. We therefore begin by considering what a comprehensive ToC should aim to account for, noting that even this issue is contested, with theorists often disagreeing about what kinds of phenomena a ToC should explain. The heart of the problem of consciousness is the issue of experience' or 'subjective awareness'. Although no non-circular definition of these terms can be provided, the target phenomenon can be illuminated through some intuitive distinctions. There is something it is like' for an organism to be conscious20, and what it is like to be in one state of consciousness differs from what it is like to be in another state of consciousness. A comprehensive ToC will explain why some organisms or systems are Explanatory gap intuitions Intuitions that there is no prospect of a fully satisfying explanation of consciousness in physical, mechanistic terms. Adversarial collaborations Research projects in which proponents of different theories together design an experiment to distinguish their preferred theories, and agree in advance about how the outcome will favour one theory over the other(s). Global states Relating to an organism's overall state of consciousness, usually linked to arousal and behavioural responsiveness, and associated with the 'level' of consciousness. Local states Relating to particular conscious mental states, such as a conscious perception, emotion or thought. Local states are also often called conscious contents. conscious whereas others are not, and it will also explain why states of consciousness differ from each other in the ways that they do. States of consciousness can be grouped into two classes: global states and local states. Global states concern an organism's overall subjective profile and are associated with changes in arousal and behavioural responsiveness. Familiar global states include wakefulness, dreaming, sedation, the minimally conscious state, and (perhaps) the psychedelic state. These global states are sometimes called 'levels' of consciousness, but we prefer the term global states' because it leaves open the possibility that these states cannot be given a complete ordering in terms of a single dimension, but are best conceptualized as regions within a multidimensional space21. Local states — often referred to as conscious contents' or as states having qualia' — are characterized by 'what it is like' to be in them. The local state associated with having a headache is distinct from the local state associated with smelling coffee, for what it is like to have a headache differs from what it is like to smell coffee. Local states can be described at different levels of granularity, from low-level perceptual features (for example, colour), to objects, to complete multimodal perceptual scenes. An important subset of local states underpins the experience of selfhood, which encompasses experiences of mood, emotion, volition, body ownership, explicit autobiographical memory and the like13'22"24. Although neurobiological theories tend to focus on local states with sensory and perceptual content, consciousness also includes local states with cognitive and propositional content, such as the thoughts that arise when solving a crossword puzzle. Importantly, the local states that an agent has at a particular time do not simply occur as independent elements but are, instead, bound together as components of a single conscious scene that subsumes each of the agent's local states25'26. A second distinction is between the phenomenal properties of consciousness and its functional properties. The former term refers to the experiential character of consciousness, as is suggested by the phrase 'what it is like'. The functional aspects of consciousness concern the role(s) that mental states play in the cognitive economy of an organism in virtue of being conscious. ('Function' here encompasses both teleological functions — functional roles as shaped by evolution — and dispositional functions — the role a process plays in the operation of a larger system of which it is a part; see REE27) For example, consciously seeing a coffee cup may enable a range of functions, such as the ability to behave flexibly with respect to the cup (perhaps to drink from it, or to throw it across the room), to lay down an episodic memory of the event, and to provide verbal reports about the experience. In making this distinction, we are not claiming that phenomenal and functional properties are independent (they are very likely not to be independent), merely that they provide distinct explanatory targets for ToCs. As we will see, some ToCs focus on the phenomenal features of consciousness, others focus on the functional features of consciousness and still others attempt to account for both the functional and phenomenal features of consciousness. 440| JULY 2022 I VOLUME 23 www.nature.com/nrn REVIEWS Binocular rivalry A phenomenon in which different images are presented to each eye, and conscious perception alternates between the two images. A third distinction is between two kinds of questions concerning local states (contents') that a ToC might attempt to answer. On one hand, one might ask why an agent is in a certain local state (rather than another). On the other, one might ask why a particular local state has the experiential character that it has (rather than an experiential character of some other kind). This distinction can be explained with reference to binocular rivalry, in which each eye is presented with a different stimulus (say, a house to the right eye and a face to the left eye), and the subject's visual experience alternates between the left-eye stimulus and the right-eye stimulus28. Take a particular time at which the contents of consciousness involve a house, whereas the face is not consciously perceived. Here, we can ask why the mental state corresponding to 'house' is conscious (and that of'face' is unconscious), and we can also ask why visual experiences of a house have the distinctive experiential character that they have rather than, say, the experiential character of seeing a face, hearing a bell or feeling pain. Table 11 A selection of theories of consciousness Theory Primary claim Key refs Higher-ordertheory (HOT) Consciousness depends on meta-representationsof lower-order mental states 31,46 Self-organizing meta-representational theory Consciousness is the brain's (meta-representational) theory about itself 34,140 Attended intermediate representation theory Consciousness depends on the attentional amplification of intermediate-level representations 141,142 Global workspace theories (GWTs) Consciousness depends on ignition and broadcast within a neuronal global workspace where fronto-parietal cortical regions play a central, hub-like role 47-49 Integrated information theory (NT) Consciousness is identical to the cause-effect structure of a physical substrate that specifies a maximum of irreducible integrated information 57,59,60 Information closure theory Consciousness depends on non-trivial information closure with respect to an environment at particular coarse-grained scales 143 Dynamic core theory Consciousness depends on a functional cluster of neural activity combining high levels of dynamical integration and differentiation 144 Neural Darwinism Consciousness depends on re-entrant interactions reflecting a history of value-dependent learning events shaped by selectionist principles 145,146 Local recurrency Consciousness depends on local recurrent or re-entrant cortical processing and promotes learning 65,71 Predictive processing Perception depends on predictive inference of the causes of sensory signals; provides a framework for systematically mapping neural mechanisms to aspects of consciousness 67,73,79 Neuro-representationalism Consciousness depends on multilevel neurally encoded predictive representations 84 Active inference Although views vary, in one version consciousness depends on temporally and counterfactually deep inference about self-generated actions 76; see also91 Beast machine theory Consciousness is grounded in allostatic control-oriented predictive inference 13,75,77. also90 Neural subjective frame Consciousness depends on neural maps of the bodily state providing a first-person perspective 24 Self comes to mind theory Consciousness depends on interactions between homeostatic routines and multilevel interoceptive maps, with affect and feeling at the core 23,147 Attention schema theory Consciousness depends on a neurally encoded model of the control of attention 148 Multiple drafts model Consciousness depends on multiple (potentially inconsistent) representations ratherthan a single, unified representation that is available to a central system 149 Sensorimotor theory Consciousness depends on mastery of the laws governing sensorimotor contingencies 88 Unlimited associative learning Consciousness depends on a form of learning which enables an organism to link motivational value with stimuli or actions that are novel, compound and non-reflex inducing 150 Dendritic integration theory Consciousness depends on integration of top-down and bottom-up signalling at a cellular level 151 Electromagnetic field theory Consciousness is identical to physically integrated, and causally active, information encoded in the brain's global electromagnetic field 152 Orchestrated objective reduction Consciousness depends on quantum computations within microtubules inside neurons 18 Our selection of theories includes those that are either neurobiological in nature or potentially expressible in neurobiological terms. NATURE REVIEWS | NEUROSCIENCE VOLUME 23 | JULY 2022 | 441 REVIEWS Notably, there may be some contents that cannot be conscious (for example, low-level processing within early sensory or regulatory systems) and others that can only be conscious (for example, globally integrated perceptual scenes). Thus, in addition to explaining why some mental contents are conscious in some contexts but not others, another challenge is to explain why some contents can never be conscious and why others can exist only as conscious. Rather than address the full range of issues that we have just identified, most ToCs aim to explain only certain aspects of consciousness, perhaps as a step on the way to becoming comprehensive. Although being restricted in some way is not itself an objection to a ToC, it does mean that the task of inter-theory comparison is less straightforward than it might otherwise be. If theories are targeting different aspects of consciousness (say, one theory is focused on the phenomenal character of consciousness and another is focused on its functional profile) then they might not be the adversaries' that they at first glance appear to be. The ToCs we review here are grouped into four categories: higher-order theories (HOTs), global workspace theories (GWTs), integrated information theory (IIT) and re-entry and predictive processing theories. Although some accounts of consciousness straddle multiple categories, and others are not plausibly subsumed under any of these categories (TABLE 1), this four-way distinction between ToCs provides a useful lens through which to view the current state of play in the science of consciousness (BOX 2; for other ways of grouping theories, see for example REF.29). In what follows, we introduce the key elements of each category, describe some notable within-category differences, and identify those aspects of consciousness most closely associated Box 2 | Other approaches: attention, learning and affect The landscape of theories of consciousness (ToCs) includes numerous other theoretical approaches in addition to those surveyed in this Review (TABLE 1). One approach focuses on attention. For example, Graziano's attention schema theory associates conscious perception with a model of the control of attention148. Another attention-based ToC is the attended intermediate representational theory. First proposed by Jackendoff141 and defended in detail by Prinz142, this theory holds that consciousness occurs when intermediate-level perceptual representations gain access to attention. Other theoretical approaches focus on learning. These include the proposal by Jablonka and Ginsburg that minimal consciousness is underpinned by a form of associative learning they term 'unlimited associative learning'. According to their proposal, this form of learning enables an organism to link motivational value with stimuli or actions that are novel, compound and non-reflex inducing150. Other learning-based theories overlap with some theories we have already described, such as Cleeremans' version of higher-order theory (HOT)34 140 and Lamme's local recurrency account, which holds that recurrent signalling underpins consciousness in virtue of its role in learning65. Learning-based theories are also closely related to 'selectionist' approaches, which ground consciousness in evolutionary-like dynamics within and between neuronal populations145 146. Affect-based theories emphasize the brain's role in physiological regulation as the basis for consciousness. These theories include Damasio's proposal that consciousness depends on hierarchically nested representations of the organism's physiological condition147170, and proposals that mix an affect-based emphasis with predictive processing to ground conscious experiences in control-oriented interoceptive predictions13-77-90. Some affect-based theories deny that cortical mechanisms are necessary for consciousness, instead locating the mechanisms of consciousness in the brainstem171172 (although see REF.173). Phenomenal character The experiential nature of a local state, such as the 'redness' of an experience of red or the pain of a toothache — sometimes also called qualia. Meta-representation A mental representation that has as its target another mental representation with each category. We then illustrate how these ToCs relate to each other in terms of some prominent empirical debates, and present several proposals that, we suggest, will help to drive a virtuous cycle between theory development and experimental investigation. Higher-order theories The core claim that unites all HOTs is that a mental state is conscious in virtue of being the target of a certain kind of meta-representational state. Meta-representations are not merely representations that occur higher or deeper in a processing hierarchy but are, rather, representations that have as their targets other representations (FIG. 1). For example, a representation with the content 'I have a visual experience of a moving dot' is a meta-representation, for its content concerns the agent's own representations of the world rather than the world itself. An important respect in which HOTs differ from each other concerns the account that they give of the nature and role of the meta-representations that are responsible for consciousness. Some versions of the approach identify the kinds of meta-representations that are crucial for consciousness with thoughts (or thought-like states) that have conceptual content30-32. Other varieties of HOT have been expressed in computational terms. According to the self-organizing meta-representational account, consciousness involves higher-order brain networks learning to redescribe the representations encoded in lower-order networks in a way that counts as meta-representational33'34. Alternatively, higher-order state space theory proposes that subjective reports (for example, statements such as 'I am aware of X') are metacognitive (higher-order) decisions about a generative model of perceptual content35, whereas perceptual reality monitoring posits that conscious perception arises when a higher-order network judges a first-order representation to be a reliable reflection of the external world36'37. As should be clear from the foregoing, HOTs focus on explaining why some contents are conscious whereas others are not. However, these theories are not limited to this particular focus — they also have the resources to address issues pertaining to the experiential character of local states. One prominent example concerns the (debated) intuition that the contents of perceptual experience often outstrip the information available in 'first-order' sensory representations, as is alleged to occur in the context of peripheral vision38'39. The HOT-based proposal here is that the apparently 'inflated' phenomenology of peripheral visual experience is caused by the higher-order misrepresentation of first-order states40. The HOT approach can also be extended to explain why some contents are unable to be conscious (they cannot be the targets of appropriate meta-representational states) and why some contents are necessarily conscious (they are necessarily accompanied by appropriate meta-representational states). HOTs rarely focus on global states of consciousness, but it would be natural for them to appeal to the integrity of (meta-)representational processes to account for the distinctions between global states. 442| JULY 2022 | VOLUME 23 www.nature.com/nrn REVIEWS A particularly intriguing question is whether (and if so, how) HOTs explain the distinctive phenomenal character of various kinds of experiences. Why is the phenomenal character associated with seeing a sunset so different from the phenomenal character associated with a headache? The general shape of the higher-order response to this question is that the phenomenal character of a state is determined by the properties that the relevant meta-representational state ascribes to it. Most examples of this approach focus on visual experience40, but there have also been higher-order attempts to account for the phenomenal character of emotional states41 and metacognitive states, such as 'what it is like' to feel confident in a perceptual decision42,43. Ultimately, any fully reductive version of the higher-order approach must explain why the representation of various properties generates the phenomenology that it does (or is identical to it), and how neural activity enables the relevant properties to be represented in the first place. Higher-order accounts of consciousness are primarily accounts of what makes a mental state conscious and, as such, the approach is not committed to any particular view of the function(s) of consciousness. Indeed, some HOTs downplay the idea that consciousness has any distinctive function44. Other versions of the higher-order approach identify the functional role of consciousness with the metacognitive processes associated with confidence judgements and error monitoring45. However, whereas higher-order views allow conscious mental states to be accompanied by conscious metacognitive judgements — such as those involved in explicit performance monitoring or subjective confidence reports — most versions of this approach do not require that conscious perception is always accompanied by a corresponding conscious metacognitive state. Instead, for meta-representations to be conscious, they themselves must be the objects of a suitable meta-representational state. With respect to the neural basis of consciousness, the emphasis on meta-representation has led higher-order theorists to emphasize anterior cortical regions, especially the prefrontal cortex30, given the association of these regions with complex cognitive functions. However, although most HOTs propose that anterior involvement is implicated in consciousness, there is disagreement about precisely which anterior regions (or processes) are required46. Global workspace theories GWTs originate from 'blackboard' architectures in artificial intelligence, in which the blackboard is a centralized resource through which specialized processors share and receive information. The first GWT of consciousness47 was framed at a cognitive level. It proposed that conscious mental states are those that are 'globally available' to a wide range of cognitive processes including attention, evaluation, memory and verbal report. The core claim of GWTs is that it is the wide accessibility of information to such consumer cognitive systems that constitutes conscious experience (FIG. 2). This basic claim has since been developed into a neural theory — often referred to as the global neuronal workspace theory' — according to which sensory information Fig. 1 | Higher-order theories. The core claim in higher-order theories (HOTs) of consciousness is that mental states are conscious in virtue of being the target of specific kinds of meta-representation. For example, lower-order representations of visual signals in posterior cortex would support conscious visual perception when targeted by the right kind of higher-order meta-representation. Supportive evidence for HOTs comes from studies implicating anterior cortical areas in conscious contents, with an emphasis on prefrontal cortex — especially when performance is matched across conscious and non-conscious conditions30 100. HOTs are also indirectly supported by lesion evidence linking metacogni-tion to prefrontal areas153. These theories are challenged by evidence suggesting that anterior areas are not involved in consciousness108 154, perhaps instead being necessary only for enabling subjective report and executive control6. Figure adapted with permission from REF.46, Elsevier. gains access to consciousness when it is 'broadcast' within an anatomically widespread neuronal workspace that is implemented across higher-order cortical association areas, with a particular (although not exclusive) emphasis on the prefrontal cortex48'49. Access to the global workspace is achieved through nonlinear network 'ignition in which recurrent processing amplifies and sustains neuronal representations50. The emphasis on ignition and broadcast — as compared with meta-representation — is one way in which GWT is distinguished from the HOT approach. Like HOTs, GWTs focus on the question of what makes a representation conscious, and GWT theorists have rarely attempted to account for the phenomenal differences between distinct kinds of experiences (although see REF.51). Returning to our example of binocular rivalry, the GWT view aims to explain why, at a particular point in time, the mental state corresponding to 'house' is conscious (whereas that corresponding to 'face' is unconscious), but offers no direct account of the experiential contrast between seeing a house on the one hand and seeing a face on the other. The relative silence of GWTs on the issue of experiential character aligns with the general tendency of such theories to focus on functional, rather than phenomenal, aspects of consciousness. In fact, GWTs are often explicitly proposed as accounts of conscious access'49; that is, as accounts of why certain representations are available to be flexibly used by a wide range of consuming systems NATURE REVIEWS | NEUROSCIENCE VOLUME 23 | JULY 2022 | 443 REVIEWS Global workspace __r--7---I---. No-report paradigms Behavioural experiments in which participants do not provide subjective (verbal, behavioural) reports.