Wanja Wiese

How should we assess which artificial systems could be conscious? Given uncertainty about the nature and distribution of consciousness, it is promising to look for indicators of consciousness that provide evidence for (or against) consciousness in artificial systems. A challenge is that there are hard cases in which the evidence pulls into different directions. In particular, cognitive and behavioural similarities between artificial and biological systems may speak for the hypothesis that a given artificial system is conscious; differences regarding the underlying mechanisms and substrates may speak against it. This paper introduces a taxonomy of indicators of consciousness and distinguishes between approaches that manage uncertainty about indicators (reaching rational verdicts in the light of uncertainty) and approaches that seek to reduce uncertainty (improving our understanding of what counts as evidence). I argue that hard cases of possible artificial systems require that we reduce uncertainty, before we can rationally infer the presence or absence of consciousness. Furthermore, I discuss ways in which a reduction of uncertainty may be achieved.

Can machines ever become conscious? A central debate in this context concerns the question whether consciousness requires biological states. Within this debate, there exists a fundamental dispute between two widely endorsed views: biological naturalism and computational functionalism about consciousness. Specifically, whereas biological naturalists hold that consciousness requires biological states, computational functionalists fiercly deny this. This fundamental dispute has hitherto remained unresolved, and we discuss several reasons why it is unlikely that it will be resolved anytime soon. Accordingly, we suggest that both sides will likely remain in a dialectical stalemate regarding the fundamental dispute and, by extension, are unlikely to reach a consensus on positive verdicts about consciousness in AI. However, we suggest that progress can be made by investigating potential non-biological constraints on implementation. Such constraints can help reach a consensus between both sides to the effect that certain types of artificial systems are not conscious, if they fail to satisfy relevant constraints. If correct, a consensus on negative verdicts about consciousness in certain challenging types of AI is possible, even if both sides remain in conflict about positive verdicts about AI consciousness.

Consciousness science is brimming with theories attempting to explain what consciousness is and how it can be identified in natural and artificial systems. These theories often make different and conflicting claims, evincing that the road to a mature theory of consciousness is still long and convoluted. In this paper, we propose that a promising way forward that at the same time facilitates addressing urgent questions about consciousness is what we call a building block approach. Building blocks are the simpler elements making up a theory, such as key notions or assumptions that contribute to explain important aspects of the phenomenon of interest. A building block approach recommends shifting the focus to these simpler elements of theories of consciousness while minimizing commitment to specific theoretical proposals. Here, we demonstrate how building block approaches can contribute to the assessment, integration and unification of diverse theories in consciousness science, by discussing what roles two potential building blocks, namely, attention and information generation, (should) play in theories of consciousness.

I suggest that conscious perception can usefully be regarded as a form of augmented reality (AR). The AR metaphor can be motivated by considering perception within the framework of predictive processing, and it is a fruitful metaphor for at least two reasons. On the one hand, it highlights some key features of conscious perception and may therefore facilitate a better understanding of conscious perception. On the other hand, the metaphor yields suggestions and new questions that are of direct relevance to the mind-technology problem (Clowes, Gärtner, and Hipólito 2021b).

How can the free energy principle contribute to research on neural correlates of consciousness, and to the scientific study of consciousness more generally? Under the free energy principle, neural correlates should be defined in terms of neural dynamics, not neural states, and should be complemented by research on computational correlates of consciousness – defined in terms of probabilities encoded by neural states. We argue that these restrictions brighten the prospects of a computational explanation of consciousness, by addressing two central problems. The first is to account for consciousness in the absence of sensory stimulation and behaviour. The second is to allow for the possibility of systems that implement computations associated with consciousness, without being conscious, which requires differentiating between computational systems that merely simulate conscious beings and computational systems that are conscious in and of themselves. Given the notion of computation entailed by the free energy principle, we derive constraints on the ascription of consciousness in controversial cases (e.g., in the absence of sensory stimulation and behaviour). We show that this also has implications for what it means to be, as opposed to merely simulate a conscious system.

The goal of this short chapter, aimed at philosophers, is to provide an overview and brief explanation of some central concepts involved in predictive processing (PP). Even those who consider themselves experts on the topic may find it helpful to see how the central terms are used in this collection. To keep things simple, we will first informally define a set of features important to predictive processing, supplemented by some short explanations and an alphabetic glossary. The features described here are not shared in all PP accounts. Some may not be necessary for an individual model; others may be contested. Indeed, not even all authors of this collection will accept all of them. To make this transparent, we have encouraged contributors to indicate briefly which of the features are necessary to support the arguments they provide, and which (if any) are incompatible with their account. For the sake of clarity, we provide the complete list here, very roughly ordered by how central we take them to be for “Vanilla PP” (i.e., a formulation of predictive processing that will probably be accepted by most researchers working on this topic). More detailed explanations will be given below. Note that these features do not specify individually necessary and jointly sufficient conditions for the application of the concept of “predictive processing”. All we currently have is a semantic cluster, with perhaps some overlapping sets of jointly sufficient criteria. The framework is still developing, and it is difficult, maybe impossible, to provide theory-neutral explanations of all PP ideas without already introducing strong background assumptions.

Predictive processing is an influential theoretical framework for understanding human and animal cognition. In the context of predictive processing, learning is often reduced to optimizing the parameters of a generative model with a predefined structure. This is known as Bayesian parameter learning. However, to provide a comprehensive account of learning, one must also explain how the brain learns the structure of its generative model. This second kind of learning is known as structure learning. Structure learning would involve true structural changes in generative models. The purpose of the current paper is to describe the processes involved upstream of these structural changes. To do this, we first highlight the remarkable compatibility between predictive processing and the processing fluency theory. More precisely, we argue that predictive processing is able to account for all the main theoretical constructs associated with the notion of processing fluency (i.e., the fluency heuristic, naïve theory, the discrepancy-attribution hypothesis, absolute fluency, expected fluency, and relative fluency). We then use this predictive processing account of processing fluency to show how the brain could infer whether it needs a structural change for learning the causal regularities at play in the environment. Finally, we speculate on how this inference might indirectly trigger structural changes when necessary.

This chapter explores to what extent some core ideas of predictive processing can be applied to the phenomenology of time consciousness. The focus is on the experienced continuity of consciously perceived, temporally extended phenomena (such as enduring processes and successions of events). The main claim is that the hierarchy of representations posited by hierarchical predictive processing models can contribute to a deepened understanding of the continuity of consciousness. Computationally, such models show that sequences of events can be represented as states of a hierarchy of dynamical systems. Phenomenologically, they suggest a more fine-grained analysis of the perceptual contents of the specious present, in terms of a hierarchy of temporal wholes. Visual perception of static scenes not only contains perceived objects and regions but also spatial gist; similarly, auditory perception of temporal sequences, such as melodies, involves not only perceiving individual notes but also slightly more abstract features (temporal gist), which have longer temporal durations (e.g., emotional character or rhythm). Further investigations into these elusive contents of conscious perception may be facilitated by findings regarding its neural underpinnings. Predictive processing models suggest that sensorimotor areas may influence these contents.

This paper sketches an account that explains the elusive subjective quality of 'enduring substantiality' of the phenomenal self. It integrates a recent predictive processing account of the self by Chris Letheby and Philip Gerrans with key ideas of Michael Graziano's attention schema theory of consciousness. Similarly to the attention schema theory, the present account posits an internal model of ongoing attentional processing that supports attentional control. In terms of predictive processing, it is a dynamic model of precision estimates that represents the salience of features, objects, and internal and external processes. The apparent substantiality of the self is then explained by a structural feature of this salience model: it binds different dimensions of salience by representations of higher-order dimensions of salience. The result of this salience binding is a phenomenally transparent 'salience object', which is represented as a cause of widespread changes in precision estimates, and results in a form of epistemic self-control. Phenomenologically, this goes along with the experience of the self as an attentional agent, and in this way creates the origin of our consciously experienced first-person perspective: I attend, I am.

Paweł Gładziejewski has recently argued that the framework of predictive processing postulates genuine representations. His focus is on establishing that certain structures posited by PP actually play a representational role. The goal of this paper is to promote this discussion by exploring the contents of representations posited by PP. Gładziejewski already points out that structural theories of representational content can successfully be applied to PP. Here, I propose to make the treatment slightly more rigorous by invoking Francis Egan’s distinction between mathematical and cognitive contents. Applying this distinction to representational contents in PP, I first show that cognitive contents in PP are determined by mathematical contents, at least in the sense that computational descriptions in PP put constraints on ascriptions of cognitive contents. After that, I explore to what extent these constraints are specific. I argue that the general mathematical contents posited by PP do not constrain ascriptions of cognitive content in a specific way. However, there are at least three aspects of PP that constrain ascriptions of cognitive contents in more specific ways: formal PP models posit specific mathematical contents that define more specific constraints; PP entails claims about how computational mechanisms underpin cognitive phenomena ; the processing hierarchy posited by PP goes along with more specific constraints.

The problem of phenomenal unity consists in providing a phenomenological characterization of the difference between phenomenally unified and disunified conscious experiences. Potential solutions to PPU are faced with an important challenge. I show that this challenge can be conceived as a phenomenological dual to what is known as Bradley’s regress. This perspective facilitates progress on PPU by finding duals to possible solutions to Bradley’s regress and makes it intelligible why many characterize phenomenal unity in terms of the existence of a single global conscious state. I call this latter view the “single state conception”. SSC is superficially attractive, because it seems to provide a solution to the phenomenological dual to Bradley’s regress, but should still be rejected, because it does not solve PPU; instead, it creates more problems; these problems can be avoided by alternative conceptions of phenomenal unity.

The Predictive Mind by Jakob Hohwy is the first monograph to address the philosophical significance of what Hohwy calls the prediction error minimization framework. The central claim of the book is that, on a conceptual level, perception, action, and cognition can be understood by reference to a single principle: prediction error minimization. The corresponding empirical hypothesis is that the brain implements a hierarchical generative model that generates predictions about sensory inputs and their hidden causes. When sensory signals arrive, only their divergence from the predictions has to be further processed. The general strategy of using predictions derived from generative models to compress and transmit information is also known as predictive coding.Perception is thus not construed as a purely bottom–up process, but.

Are there logically possible types of conscious experience that are nomologically impossible, given independently justified assumptions about the neural underpinnings of consciousness in human beings? In one sense, this is trivial: just consider the fact that the types of perceptual experiences we can have are limited by our sensory organs. But there may be non-trivial types of conscious experience that are impossible. For instance, if there is a basic type of self-consciousness, corresponding to a phenomenal property that is nomologically necessary for consciousness, then experiences lacking this phenomenal property will be (nomologically) impossible. More generally, it may be that there are causal dependencies between the neural mechanisms that are required to instantiate distinct phenomenal properties (in human beings). If this is the case, instantiating one of these phenomenal properties without certain others may be impossible, which means there are non-trivial cases of nomologically impossible types of conscious experience. This paper clarifies this hypothesis, outlines a general methodology for its investigation, and relates it to research on radical disruptions of self-consciousness.

According to active inference, action is enabled by a top-down modulation of sensory signals. Computational models of this mechanism complement ideomotor theories of action representation. Such theories postulate common neural representations for action and perception, without specifying how action is enabled by such representations. In active inference, motor commands are replaced by proprioceptive predictions. In order to initiate action through such predictions, sensory prediction errors have to be attenuated. This paper argues that such top-down modulation involves systematic misrepresentations. More specifically, the paper first argues for the following conditional claim. If active inference provides an accurate computational description of how action is enabled in the brain, then action is enabled by systematic misrepresentations. Furthermore, it is argued that an inference to the best explanation provides reason for believing the antecedent is true: Firstly, active inference provides a crucial extension to ideomotor theories. Secondly, active inference explains otherwise puzzling phenomena related to sensory attenuation, e.g. in force-matching or self-tickling paradigms. Taken together, these reasons support the claim that action is indeed enabled by systematic misrepresentations. The claim casts doubt on the assumption that representations are systematically beneficial to the extent that they are true: if the argument in this paper is sound, systematically beneficial misrepresentations may lie at the heart of our neural architecture.