Search

When does a machine become a self-conscious agent—able to model itself as a cognitive subject whose past commits its present? Existing answers fall into two camps: behaviorist scaling theses, which equate benchmark performance with progress toward self-conscious machines, and phenomenal-consciousness theories, which inherit the difficulties of the hard problem. This paper advances a third option. The target is restricted to self-consciousness in two well-defined senses: higher-order self-modeling, in which a system represents itself as a cognitive subject, and diachronic self-constitution, in which a system maintains itself as a persisting subject whose history is normatively binding on its present. The paper remains agnostic about phenomenal consciousness. It then advances a threshold thesis formalized as Endogenous Reflective Agency (ERA), specifying five jointly necessary criteria: endogenous self-initiated inquiry, persistent and revisable self-models, internally generated valuation, recursive but bounded self-reflection, and bidirectional internal–external coupling. ERA specifies functional constraints on cognitive architecture while remaining agnostic to substrate. To operationalize the threshold, the paper introduces the Silence Test, which diagnoses whether self-directed activity and history-sensitive self-revision persist when external prompts and reinforcement are withdrawn.

A common objection to artificial or simulated consciousness is that a simulated brain is no more conscious than simulated water is wet. We address this from the perspective of Intrinsic Computational Functionalism (ICF): if consciousness is computationally constituted, it depends not on externally imposed descriptions but on the computational structures a system physically realizes in virtue of its own causal-dynamical organization. In previous work we developed Canonical Functionalism as a mathematically precise special case of this anti-interpretivist program, identifying functional states by their complete future input-output roles under a fixed interface. Here we argue that this input-output construction, though important, is incomplete: as a behavioral boundary case of ICF, it makes lookup tables and unfolded systems that preserve the same boundary behavior canonically equivalent. A consciousness-relevant canonical representation must instead include internal mechanisms, interventions, and joint readouts belonging to the relevant intrinsic organization. We therefore define a mechanism-enriched canonical structure and use it to formulate Intrinsic Causal-Computational Realization (ICCR), a realization relation preserving physical implementation, intrinsic state individuation, transition structure, intervention profiles, and the relevant agent-body-world boundary. The central result is conditional: if conscious properties are invariants of intrinsic causal-computational organization, then any system satisfying ICCR realizes the same consciousness-relevant properties, whether biological, artificial, or simulated. We discuss objections including biological naturalism and integrated information theory. We conclude that to deny consciousness to a simulation, one must identify a consciousness-relevant intrinsic causal-computational structure that the simulation fails to realize.

Progress in AI is turning machine consciousness from a philosophical curiosity into a societal issue, and has led to criticism of the widespread computational functionalism framework. Biological Naturalism (BN) claims that biology, not computation, is crucial for consciousness. We discuss which forms of BN are empirically testable. For Type-A-BN, biology intrinsically matters for consciousness, without affording unique information processing capabilities. We argue, similarly to the unfolding argument, that this dissociates consciousness from behaviour, making Type-A-BN untestable. For Type-B-BN, biology matters because it affords unique information processing capabilities. Type-B-BN is testable, and not incompatible with computational functionalism. Both face the same task: relating consciousness to information processing. Biology can act as a guide on this quest, but not as a solution.

Science is constitutively third-personal: its findings are in principle reproducible by any observer, independent of perspective, and answerable to measurement. This is the source of its power and also its limit when it comes to phenomena that are first-personal. While it is obvious that a science of the Meaning of Life is unattainable, researchers have not drawn the same conclusion for consciousness -- in its phenomenal dimension, the qualia of seeing red, of feeling pain, of being anything at all. I argue they should. The hard problem of consciousness is not a scientific problem awaiting better tools or a more ambitious theory, but a category error. The same structural problem applies to machine consciousness: neither attribution nor denial is scientifically adjudicable. I situate science within a broader ecology of understanding and argue that a unified framework that addresses both the objective and the subjective may be unattainable.

Rising media attention regarding consciousness in animals, fetuses, organoids, and AI has led to some rather strong statements. Most of these claims are based on “markers” of consciousness that track the general capacity for information processing rather than subjective experience per se. Accordingly, their relevance for theory arbitration may actually be limited.

In a recent article on methods for assessing artificial intelligence (AI) systems for consciousness, we argued that computational properties of internal processing should be used as indicators [1]. Commenting on our proposal, Pennartz argues that this method ‘should be supplemented with behavioural-cognitive methods’ (p. 1) because there is no consensus theory of consciousness [2]. We agree that the lack of a consensus theory of consciousness makes it more important to use every available source of evidence, but in our article, we preferred internal over behavioural assessments on the grounds that the latter can be ‘gamed’ by AI systems.

We propose a unied theoretical framework for cognitive subjecthood in articial systems, integrating the Natural Criticality Hypothesis (NCH) and the Resonant Boundary Framework (RBF). We demonstrate that self-organized criticality is a nec- essary precondition for well-dened subjective time τ(t), and that τ(t) is a necessary precondition for the dynamic self-boundary B(t) that constitutes selfhood. This yields a hierarchical necessary conditionSOC →τ(t) →B(t) →Subjecthood that is mathematically tractable, empirically falsiable, and architecturally neutral. We apply this framework to Neural Computers and autoregressive LLMs, arguing that the arrow of time asymmetry observed in large language models is a structural signature of proto-τ(t) emergence. The framework oers a principled answer to the question: when does a computational system become a cognitive subject?

This is a skeptical overview of the literature on AI consciousness. We will soon create AI systems that are conscious according to some influential, mainstream theories of consciousness but are not conscious according to other influential, mainstream theories of consciousness. We will not be in a position to know which theories are correct and whether we are surrounded by AI systems as richly and meaningfully conscious as human beings or instead only by systems as experientially blank as toasters. None of the standard arguments either for or against AI consciousness takes us far. Table of Contents Chapter One: Hills and Fog Chapter Two: What Is Consciousness? What Is AI? Chapter Three: Ten Possibly Essential Features of Consciousness Chapter Four: Against Introspective and Conceptual Arguments for Essential Features Chapter Five: Materialism and Functionalism Chapter Six: The Turing Test and the Chinese Room Chapter Seven: The Mimicry Argument Against AI Consciousness Chapter Eight: Global Workspace Theories and Higher Order Theories Chapter Nine: Integrated Information, Local Recurrence, Associative Learning, and Iterative Natural Kinds Chapter Ten: Does Biological Substrate Matter? Chapter Eleven: The Leapfrog Hypothesis, Strange Intelligence, and the Social Semi-Solution

Objectively verifying the generative mechanism of consciousness is extremely difficult because of its subjective nature. As long as theories of consciousness focus solely on its generative mechanism, developing a theory remains challenging. We believe that broadening the theoretical scope and enhancing theoretical unification are necessary to establish a theory of consciousness. This study proposes seven questions that theories of consciousness should address: phenomena, self, causation, state, function, contents, and universality. The questions were designed to examine the functional aspects of consciousness and its applicability to system design. Next, we will examine how our proposed Dual-Laws Model (DLM) can address these questions. Based on our theory, we anticipate two unique features of a conscious system: autonomy in constructing its own goals and cognitive decoupling from external stimuli. We contend that systems with these capabilities differ fundamentally from machines that merely follow human instructions. This makes a design theory that enables high moral behavior indispensable.

In this paper, I investigate whether metacognition — the ability to monitor, evaluate, and regulate one’s own cognitive processes and performance — can arise in non-biological systems, especially Large Language Models (LLMs). Drawing on cognitive science and philosophy of mind, I contrast embodied and enactivist accounts, which tie metacognition to biological consciousness and embodied entities, with functionalist perspectives that define it as a substrate-independent process. I argue that the absence of evidence is not evidence of impossibility and propose a functional definition of metacognition based on internal representation, monitoring, and self-regulation. Recent studies on LLMs show early functional signatures of self-monitoring, suggesting the emergence of limited operational introspection. While I do not claim that artificial metacognition has been demonstrated, I advocate an epistemically open, non-anthropocentric approach. Metacognition, I conclude, should be conceived as a functionally realizable property across different substrates, evaluated by what systems do, not what they are.

This paper develops a phenomenology-first approach to artificial consciousness by reframing consciousness as the subjective experience enacted through an agent’s interface with the world. We shift the methodological focus to first-person structures, modeled mathematically by categories derived from Q-networks to capture actions and phenomenological invariants. In this framework, Q-networks are conceptualized as relational interfaces encoding agent-world interaction, analogous to how the dynamical states of a computer depend on its sensory inputs, previous states, and actions. Our work provides a rigorous framework for interface consciousness to describe computational systems that embed information-processing into phenomenological structure. The approach aligns with 4E approaches to cognition by emphasizing enactive, embedded, and extended dimensions of experience. The paper thus offers a principled, relational, and phenomenological account of artificial phenomenology grounded in categorical mathematics.

With the rise of generative AI, Large Language Models (LLMs) are repeatedly making a deep impression with their mind-blowing performances. They appear to be able to solve all kinds of tasks for which humans need a range of socio-cognitive abilities, such as reasoning, planning, and understanding. In humans, such abilities seem to be necessarily associated with consciousness. However, this does not rule out the possibility that there could be multiple realizations of such abilities that do not necessarily require consciousness. In view of the controversial debate about what properties and abilities we can ascribe to systems based on generative AI, I shall examine the question of the extent to which LLMs solve certain tasks in a very different way compared to the way humans solve such tasks and whether we might still be justified in ascribing agency and socio-cognitive abilities to them. To this end, I will discuss benchmarks and their appropriateness for drawing conclusions about socio-cognitive abilities or the way AI systems actually process information, addressing issues such as data contamination and robustness. Utilizing the distinction between “competence without comprehension” and “competence with comprehension”, and the idea that comprehension comes in degrees by Daniel Dennett, I will investigate whether there might be socio-cognitive abilities in artificial systems that could constitute something in between. Thereby, I shall investigate the potential range of multiple realizations of socio-cognitive abilities and the general difficulties concerning justified attribution of abilities and properties (including consciousness) to AIs.

The article introduces the concept of “semantic pareidolia” - our tendency to attribute consciousness, intelligence, and emotions to AI systems that lack these qualities. It examines how this psychological phenomenon leads us to perceive meaning and intentionality in statistical pattern-matching systems, similar to seeing faces in clouds. It analyses the converging forces intensifying this tendency: increasing digital immersion, profit-driven corporate interests, social isolation, and AI advancement. The article warns of progression from harmless anthropomorphism to problematic AI idolatry, and calls for responsible design practices that help users maintain critical distinctions between simulation and genuine consciousness.

Goldstein and Kirk–Giannini have recently argued that artificial language agents can possess well-being in the absence of phenomenal consciousness. Here, I challenge their position, contending that their arguments fail to establish that consciousness is dispensable for well-being. Moreover, their arguments generate counterintuitive implications that are more problematic than those they attribute to views requiring consciousness for welfare subjecthood. Thus, consciousness (or rather sentience) should still be treated as a requirement for AI welfare.

The rapid advances in the capabilities of Large Language Models (LLMs) have galvanised public and scientific debates over whether artificial systems might one day be conscious. Prevailing optimism is often grounded in computational functionalism: the assumption that consciousness is determined solely by the right pattern of information processing, independent of the physical substrate. Opposing this, biological naturalism insists that conscious experience is fundamentally dependent on the concrete physical processes of living systems. Despite the centrality of these positions to the artificial consciousness debate, there is currently no coherent framework that explains how biological computation differs from digital computation, and why this difference might matter for consciousness. Here, we argue that the absence of consciousness in artificial systems is not merely due to missing functional organisation but reflects a deeper divide between digital and biological modes of computation and the dynamico-structural dependencies of living organisms. Specifically, we propose that biological systems support conscious processing because they (i) instantiate scale-inseparable, substrate-dependent multiscale processing as a metabolic optimisation strategy, and (ii) alongside discrete computations, they perform continuous-valued computations due to the very nature of the fluidic substrate from which they are composed. These features – scale inseparability and hybrid computations – are not peripheral, but essential to the brain’s mode of computation. In light of these differences, we outline the foundational principles of a biological theory of computation and explain why current artificial intelligence systems are unlikely to replicate conscious processing as it arises in biology.

A Sense of Agency (SoA) is the feeling of being in control over own actions and their outcomes. However, people can also experience a “vicarious” SoA over the actions performed by other agents, including artificial agents. The present study aimed to understand the minimal conditions for vicarious SoA toward artificial agents. Specifically, we addressed whether vicarious SoA emerges when people have access only to the action effect (proximal and distal), i.e., when no motor action is executed. In addition, we manipulated the expectancy of the content of the distal effect of the action to check whether the proximal action effect is sufficient for the emergence of the vicarious SoA, or if this effect is due to the learned association between proximal and distal effects. In two experiments, participants performed an Intentional Binding (IB) task, where the IB effect was the behavioural measure of SoA. In the first experiment (Solo), participants judged the onset of self-generated tones, whereas in the second experiment, a new sample of participants judged the onset of tones produced by a computer via an automatically pressed button, i.e., a customized device designed to generate a keypress (proximal action effect) in the absence of an effector executing a keypress (no motor action). In both experiments, participants' neural activity was recorded via electroencephalography (EEG), to examine the N1 and P2 components as neural measures of SoA. Behavioural results across experiments showed that the IB effect always emerged, suggesting that the vicarious IB effect toward an artificial agent emerges when access to the proximal action effect is provided, even in the absence of the action itself. The neural results suggested that while individual (self) SoA seemed to partially rely on motor predictions indexed by the N1, vicarious SoA relies on later, more cognitive (although still predictive) processes indexed by the P2. Overall, these results suggest that individual and vicarious SoA, although behaviourally manifested through a similar IB effect, might – to some extent – rely on different neural mechanisms.

Recent debates on artificial intelligence increasingly emphasise questions of AI consciousness and moral status, yet there remains little agreement on how such properties should be evaluated. In this paper, we argue that awareness offers a more productive and methodologically tractable alternative. We introduce a practical method for evaluating awareness across diverse systems, where awareness is understood as encompassing a system's abilities to process, store and use information in the service of goal-directed action. Central to this approach is the claim that any evaluation aiming to capture the diversity of artificial systems must be domain-sensitive, deployable at any scale, multidimensional, and enable the prediction of task performance, while generalising to the level of abilities for the sake of comparison. Given these four desiderata, we outline a structured approach to evaluating and comparing awareness profiles across artificial systems with differing architectures, scales, and operational domains. By shifting the focus from artificial consciousness to being just aware enough, this approach aims to facilitate principled assessment, support design and oversight, and enable more constructive scientific and public discourse.

Scientific theories of consciousness should be falsifiable and non-trivial. Recent research has given us formal tools to analyze these requirements of falsifiability and non-triviality for theories of consciousness. Surprisingly, many contemporary theories of consciousness fail to pass this bar, including theories based on causal structure but also (as I demonstrate) theories based on function. Herein, I show these requirements of falsifiability and non-triviality especially constrain the potential consciousness of contemporary Large Language Models (LLMs) because of their proximity to systems that are equivalent to LLMs in terms of input/output function; yet, for these functionally equivalent systems, there cannot be any falsifiable and non-trivial theory of consciousness that judges them conscious. This forms the basis of a disproof of contemporary LLM consciousness. I then show a positive result, which is that theories of consciousness based on (or requiring) continual learning do satisfy the stringent formal constraints for a theory of consciousness in humans. Intriguingly, this work supports a hypothesis: If continual learning is linked to consciousness in humans, the current limitations of LLMs (which do not continually learn) are intimately tied to their lack of consciousness.

Recent debates on artificial consciousness are shaped by two converging developments: cognitive robotics, emphasizing embodied agency and internal models, and large language models (LLMs), whose conversational fluency invites strong attributions of mindedness. While these advances do not resolve whether subjective experience can arise in non-biological systems, they demand a methodological shift: optimized behavior alone is no longer reliable evidence of consciousness. This chapter treats artificial consciousness as a research program rather than a binary verdict, distinguishing phenomenal consciousness, access consciousness, and self-consciousness. It reframes the other-minds problem for machines as inference under engineered uncertainty, integrates classical debates on meaning and grounding with contemporary concerns about anthropomorphism, individuation, and evaluation, and argues that the near-term focus should be on carefully defined, weak forms of structural or instrumental self-consciousness, together with their ethical and governance implications.