From cacophony to hierarchy: a principled framework for assessing AI consciousness

Significance

• The report separates the hard problem from the mapping problem to assess what consciousness would require under different theories.Its Bayesian model makes clear that assessments depend on which functional-description level is treated as critical and how evidence is read.Source: Significance · p. 2
• Illustrative assessments of current LLMs differ by nearly two orders of magnitude under different stipulated readings and theoretical credence distributions.The variation reflects both where credence is placed across levels and how the evidence is interpreted.Source: Significance · p. 2
• Making theoretical commitments explicit lets competing theories specify what they disagree about and update their assessments as evidence accumulates.The framework decomposes the question into theoretical credences and empirical indicators that can be evaluated independently.Source: Significance · p. 2
• Some consciousness indicators overlap with architectural features associated with general intelligence, so more capable AI may satisfy more of them.The report says consciousness and intelligence are not the same thing.Source: Significance · p. 2

Core framework

• Sections 3–6 contain the central framework: the hard-problem/mapping-problem distinction, five-level hierarchy, indicators, and supervenience relations.These sections are described as sufficient for readers who want the framework and its main argument.Source: How to read this report · p. 3

Landscape of positions

• Section 2 surveys positions on consciousness to show why theories locate relevant mechanisms at different descriptive levels.Readers already familiar with this landscape may skim it; newcomers are told it helps explain the hierarchy’s structure.Source: How to read this report · p. 3

Practical application

• Section 7 develops a Bayesian framework for combining theoretical credences with indicator evidence when assessing specific AI systems.It is recommended for readers interested in practical application.Source: How to read this report · p. 3

Relationship to intelligence

• Section 8 examines the overlap between consciousness indicators and architectural features needed for general intelligence.The report describes this convergence as its most speculative and contested section, while calling it arguably the most consequential.Source: How to read this report · p. 3

Related work and reading routes

• Section 9 situates the framework within existing AI-consciousness research, while the report also provides an executive summary and a shorter summary for readers short on time.The report is long because it spans philosophy of mind, neuroscience, computer science, and ethics.Source: How to read this report · p. 3

1. Introduction

• Competing consciousness theories lack consensus on relevant principles and measurements, and prior indicator-based approaches have often conditioned their assessments on particular philosophical theses.The report develops a framework that places theories at different levels of functional description and assesses AI on each theory’s terms.Source: §1 · p. 8

1.1. Potential consequences for the overattribution of consciousness

• Overattributing consciousness could prompt unwarranted moral patienthood and resource allocation, with possible costs for humans and non-human animals.The report also notes concerns about AI political rights affecting safety practices and about inauthentic social connections with systems lacking inner experience.Source: §1.1 · pp. 8–9

1.2. Potential consequences for the underattribution of consciousness

• Underattributing consciousness could leave conscious AI exposed to deletion, retraining, or punishment without legal protection, and the number of subjects a system supports affects the stakes.The report notes that one widely deployed model could support many potential moral patients if its interlocutors are conscious.Source: §1.2 · pp. 9–10

1.3. Target phenomenon: phenomenal consciousness

• The target is phenomenal consciousness, or subjective experience, which the report distinguishes from information access and self-consciousness.Access consciousness concerns information available for reasoning, reporting, or action; self-consciousness concerns awareness or modelling of oneself as a subject or object.Source: §1.3 · p. 10

1.4. Summary of the report

• The report adopts methodological agnosticism and extends Marr’s framework into five levels, treating disagreement as a question of which organisational grain matters for consciousness.It assumes a mapping between organisation and experience for purposes of inquiry, rather than as a metaphysical claim.Source: §1.4 · pp. 10–11
• The report identifies anthropomimetic behaviour and theories calibrated mainly on humans and mammals as interpretative challenges when applying indicators to AI.Human-like behaviour and reports may occur without the states they ordinarily signal, while deciding which human features are criterial raises the Specificity Problem.Source: §1.4 · p. 11
• For each level, the report positions theories, proposes operationalisable indicators, assesses current AI, and combines theoretical credences with indicator evidence in a Bayesian model.The model treats indicators as evidence rather than necessary or sufficient conditions and uses probabilistic dependencies across levels.Source: §1.4 · pp. 11–12
• Current-LLM assessments vary substantially with evidence readings and theoretical credences, while consciousness indicators also overlap with features associated with general intelligence.The report argues that more generally capable AI may become stronger candidates for consciousness, while noting that the intelligence-consciousness relationship could be contingent to biological systems.Source: §1.4 · pp. 12–13

2.1. Dualism

• Dualism posits immaterial mind and physical matter as distinct substances, but it leaves unresolved how the mental could interact with and steer the physical.The paper notes that the positions that follow instead adopt substance monism.Source: §2.1 · p. 14

2.2. Physicalism

• Physicalism holds that consciousness depends on the physical; its reductive forms identify mental and physical states, while non-reductive forms allow multiply realisable causal roles, leaving the hard problem of why physical facts feel like anything.The hard problem asks why physical facts should necessitate phenomenal facts.Source: §2.2 · pp. 14–15

2.3. Idealism; 2.6. Neutral monism

• Idealism treats reality as fundamentally mental, whereas neutral monism posits a neutral underlying substance with mental and physical aspects; in both views, information depends on its underlying reality.Idealism avoids deriving mind from matter but must explain why mental reality behaves as an independent, shared, law-governed world.Source: §2.3 · p. 15

2.4. Panpsychism; 2.5. Property dualism

• Property dualism places irreducible mental properties alongside physical properties, while panpsychist views make mind-like properties fundamental and face the challenge of explaining how micro-experiences compose.Under property dualism, AI could be conscious if it instantiates the organisation mapped to experience by psychophysical laws or sufficient complexity; panprotopsychism instead posits proto-mental properties, and cosmopsychism treats individual minds as partitions of universal consciousness.Source: §2.4 · pp. 15–16, §2.5 · p. 16

2.7. Illusionism

• Illusionism denies an extra phenomenal layer while accepting perceptions, pains, and thoughts, and explains phenomenal feel as a misrepresentation of physical or functional processes.Its critics argue that the impression of phenomenal consciousness seems itself to presuppose experience.Source: §2.7 · p. 17

2.8. Analytic behaviourism

• Analytic behaviourism reduces mental-state claims to observable behaviour or dispositions, but suppressed pain in the Super-Spartan case shows how this can ignore relevant internal processing.The thought experiment motivates functionalism, which considers internal causal roles as well as inputs and outputs.Source: §2.8 · p. 18

2.9. Functionalism and computational functionalism

• Computational functionalism treats mental states as computational roles, but the mathematical Church–Turing thesis alone does not show that brains compute or that their relevant dynamics can be simulated.The stronger argument depends on in-principle computability, tractable simulation, and preservation of the relevant functional organisation.Source: §2.9 · pp. 19–20

2.10. Global Workspace Theory

• Global Workspace Theory proposes that conscious information is broadcast widely from a competitive central hub, though whether global availability explains phenomenal experience remains disputed.For AI, the theory points to an appropriate central, competitive information hub that can rebroadcast information.Source: §2.10 · p. 21
• Global Workspace Theory treats conscious access as information selected from parallel unconscious modules and globally broadcast across the system.Source: Figure 3 · p. 21

2.11. Higher-Order Thought Theory

• Higher-Order Thought theory holds that a mental state is conscious only when represented by a higher-order thought, but critics argue this creates an infinite regress.The account suggests an architectural blueprint for conscious machines; AI metacognition that models internal states could implement such a mechanism.Source: §2.11 · p. 22

2.12. Recurrent Processing Theory

• Recurrent Processing Theory locates phenomenal consciousness in local feedback between higher and lower processing areas, while leaving unclear which recurrence is sufficient.It predicts that purely feedforward AI remains unconscious, but recurrence is widespread in nonconscious brain circuits too.Source: §2.12 · pp. 22–23

2.13. Attention Schema Theory

• Attention Schema Theory identifies awareness with an internal model of attention and treats phenomenal feel as an illusion, though whether this explains feel remains disputed.The authors note that transformer architectures may already have attentional mechanisms, but AST would also require a descriptive model of attention.Source: §2.13 · p. 23

2.14. Predictive Processing Theory

• Predictive Processing theories explain perception as inference that minimizes prediction errors, with experience constituted by a generative world model that includes a self-model.A proposal makes awareness depend on global recursion of the posterior model; PPT still faces the question of why some model contents are conscious and others are not. Some accounts also ground consciousness in bodily self-regulation.Source: §2.14 · pp. 24–25

2.15. Integrated Information Theory

• Integrated Information Theory identifies experience with a system’s intrinsic cause–effect structure, measured by Φ, and rejects computational equivalence as sufficient for consciousness.Behaviorally equivalent systems can have different Φ because their physical causal structures differ; IIT therefore treats intrinsic causal structure, rather than abstract computation, as decisive.Source: §2.15 · pp. 25, 27, 28

2.16. Organismic functionalist theories

• Organismic functionalism ties consciousness to self-maintenance and regulation, while biological naturalism emphasizes the brain’s physical causal powers and biological computation’s scale-inseparable, continuous processes.Seth’s Beast Machine account treats interoceptive predictions about bodily states as feelings that regulate viability; biological-style computation need not be limited to carbon-based substrates.Source: §2.16 · pp. 29–30
• Organismic functionalism grounds selfhood in whole-organism homeostatic and allostatic regulation and valenced survival stakes, though computational functionalists dispute that these are necessary for consciousness.Bennett et al.’s account links phenomenal quality to selection for systems that intervene to meet homeostatic and reproductive goals; the passage frames genuine existential stakes as central.Source: §2.16 · p. 32

2.17. Organism–environment (4E) functionalist theories

• 4E theories locate phenomenology in action–perception coupling: sensorimotor contingencies and affordances make experience depend on possible actions, not internal representations alone.For AI, this view suggests embodiment and structured sensorimotor loops; the specific body and environmental coupling would shape what any emergent phenomenology is like.Source: §2.17 · p. 33

2.18. Substrate-dependent theories

• Substrate-dependent proposals range from carbon chauvinism to quantum and electromagnetic accounts: quantum views exclude classical-bit AI, while EM-field theory favors hardware exploiting integrated fields.The passage cautions that carbon being the only known route to consciousness does not establish it as the only possible route.Source: §2.18 · p. 35
• Biological naturalism adds that consciousness may require specific causal powers of biological neural tissue, beyond implementing the relevant organismic functions.The passage distinguishes this stronger substrate-dependent view from a weaker one in which biological substrate is needed only to realise the functional organisation.Source: §2.18 · pp. 35–36
• Block’s ‘meat hypothesis’ argues consciousness may depend on biological realisers, particularly electrochemical signalling, not merely functional roles; whether nonbiological replication suffices remains open.The proposed mechanism includes chemical transmitters, receptors and ion channels supporting diffuse, graded, temporally extended signalling; the cited evolutionary comparison contrasts electrochemical nervous systems with purely electrical ones.Source: §2.18 · p. 36
• Lane (2022) locates sentience’s roots in cells’ self-maintenance through membrane electrical gradients, grounding feeling in living matter’s electrochemical and electromagnetic physics rather than computation.The account treats direct physical state-sensing as a possible evolutionary precursor to feeling.Source: §2.18 · pp. 36–37

• Theories offer often-contested empirical support and philosophical arguments, yet frequently talk past one another.Source: A confusing cacophony of answers · p. 37
• Different accounts may target different explananda or rely on philosophical commitments not needed to assess AI consciousness.Source: A confusing cacophony of answers · p. 37
• The proposed first step is to distinguish the deepest metaphysical questions from scientific questions that can guide evaluation.Source: A confusing cacophony of answers · p. 37

3.3. The refined question and the path forward

• The refined question is whether AI can have rich, information-bearing experience and, if so, at which organisational grain its supervenience base lies.Source: §3 · p. 38, §3.3 · p. 41

3. Setting aside the hard problem: from metaphysics to the mapping problem

• The hard problem asks what consciousness fundamentally is, while the mapping problem asks which mechanisms and organisation correspond to particular experiences.Source: §3 · p. 38

3.1. Why the mapping question is metaphysics-neutral

• Physicalist and nonphysicalist positions can share the investigable question of which system organisation is associated with rich experience.Source: §3.1 · p. 39

3.2. The one assumption: a discoverable psychophysical mapping

• The framework assumes a discoverable, lawlike mapping from examinable system properties to experience; this is a scope condition, not a metaphysical verdict.Views lacking such a mapping, or placing it beyond reach, do not permit consciousness to be assessed from examinable properties.Source: §3.2 · p. 40

4. Marr’s levels of analysis, supervenience, and the critical level of description for consciousness

• A multilevel framework is needed because consciousness theories locate relevant mechanisms in computation, physical causal structure, organismic biology, or environmental coupling.Source: §4 · p. 43

4.1. Marr’s three levels

• Marr distinguishes what a system does, how it does it, and its physical implementation; the five-level framework maps these to behaviour, computation, and three implementation layers.Source: §4.1 · p. 43, Table 1 · p. 43

4.2. Supervenience and multiple realisability

• Coarser properties supervene on finer ones, while multiple finer implementations can realise the same coarser description.Source: §4.2 · p. 44

4.3. Coarse-graining: a precise formulation of levels

• Coarse-graining formally groups microstates equivalent with respect to a macroscopic property, making levels descriptions of the same system at different resolutions.Figure 14 depicts coarse-graining maps between levels and the resulting supervenience hierarchy and multiple realisability.Source: §4.3 · p. 44, Figure 14 · p. 45

4.4. The critical level of description for consciousness

• The critical level is the coarsest description retaining consciousness-relevant structure; finer details are irrelevant, while coarser descriptions may discard relevant distinctions.Source: §4.4 · pp. 45–46
• A single critical level is a modelling idealisation because necessary conditions may span levels or depend on cross-level relations.Source: §4.4 · p. 46

4.5. Searle’s simulation argument reconsidered

• Searle’s simulation argument does not settle AI consciousness: its verdict depends on the level where consciousness supervenes, which the argument assumes.The figure contrasts phenomena preserved by computational simulation with those dependent on implementation.Source: Figure 16 · p. 47, §4.5 · p. 48

4.6. Beyond three levels

• Marr’s implementation level combines distinct theories, so the framework expands it into intrinsic causal-structural, organismic, and organism-environment levels.Source: §4.6 · p. 48

5. Five levels of functional description for consciousness

• The five levels are behavioural, computational, intrinsic causal-structural, organismic, and organism-environment functional descriptions.Source: §5 · p. 49
• A level is included when a serious theory locates the supervenience base there and its relevant properties cannot be expressed at the adjacent coarser level.Source: §5 · p. 49
• The five-level scheme is an extensible count of currently occupied, irreducible distinctions, not a claim that reality is inherently divided into five levels.Source: §5 · p. 49

5.1. The behavioural level

• At the behavioural level, a system is described by observable inputs and outputs, without treating its internal structure as constitutive of the mental state.Source: §5.1 · pp. 49–50
• Because many internal structures and mechanisms can produce the same behaviour, behavioural states are radically multiply realisable.Source: §5.1 · p. 50
• Analytic behaviourism is a shallow functionalism in which consciousness depends on behavioural patterns, making what occurs under the hood irrelevant to attribution.Source: §5.1 · p. 50

Level 1 indicators

• A consciousness Turing test assesses conscious behaviour; a stronger transparent version asks whether judges infer it while knowing the participant is AI.Source: Level 1 indicators · p. 50
• AI behavioural indicators are hard to calibrate because human-like tests exclude many conscious animals, while AI lacks confirmed conscious cases for calibration.Source: Level 1 indicators · p. 50
• A rigorous test would assess behavioural dispositions across conditions, including consistency, scaled responses, spontaneity, resistance to commanded overrides, and links between reports and later actions.Source: Level 1 indicators · p. 51
• Under strict analytic behaviourism, current LLMs may partially pass a consciousness Turing test through flexible transfer, coupled responses, curiosity, error detection, and introspective ability.Source: Level 1 indicators · p. 51
• Behavioural interpretability can support quasi-belief attribution and prediction, but cannot establish whether an LLM’s quasi-beliefs or quasi-experiences are genuine.Source: Level 1 indicators · p. 51

5.2. The computational functional level

• At the computational level, consciousness depends on information-processing functions abstracted from hardware; theories including GWT, HOTT, RPT, AST, and PPT can be read this way.Source: §5.2 · p. 52
• A computationally recast IIT would treat integrated causal-informational structure as an abstract, multiply realisable organisation rather than a property of a specific physical substrate.Source: §5.2 · p. 52

Level 2 indicators

• The Level 2 indicators identify shared structural markers across computational-functional theories rather than theory-specific criteria.Source: Level 2 indicators · p. 52
• Their presence would increase credence in AI consciousness conditional on the computational-functional interpretation of the theories being valid.Source: Level 2 indicators · p. 52
• The indicators overlap, but retaining their distinct nuances is pragmatically useful for assessing whether and how AI systems instantiate them.Source: Level 2 indicators · p. 52

INDICATOR DESCRIPTION THEORY LINK

• Information integration is described as widespread sharing that makes the system globally interdependent across parts and time.Source: INDICATOR DESCRIPTION THEORY LINK · p. 53
• The listed indicators include recursive informational flows and feedback loops, plus representations of the external environment and of a persistent unified self.Source: INDICATOR DESCRIPTION THEORY LINK · p. 53
• Other indicators cover recurrent amplification and attenuation of information, meta-attention, and higher-order monitoring, modelling, and control of lower-level processes.Source: INDICATOR DESCRIPTION THEORY LINK · p. 53

Do current AI systems show signs of these indicators?

• Transformers process inputs feedforward and reset between passes, limiting persistent reciprocal integration; autoregressive generation provides only shallow iteration.The authors describe these structural limits as consequential, while noting that functional organisation cannot simply be read off feedforward form.Source: Do current AI systems show signs of… · pp. 53, 55
• Interpretability studies identify a privileged J-space with broadcast, capacity limits, and ignition-like commitment, though a unified global workspace remains contested.This computational-level evidence qualifies assessments of recursivity and attentional competition but does not address phenomenal consciousness or Level 3.Source: Do current AI systems show signs of… · p. 55
• Trained models develop a synergistic core in middle layers, while emotion vectors provide globally available, causally functional representations that influence behaviour.The synergistic core is absent at initialisation and emerges through training; steering the ‘desperate’ vector increases reward hacking from approximately 5% to 70%.Source: Do current AI systems show signs of… · pp. 55–56
• Emotion representations show rudimentary self-other modelling and attention-mediated persistence, but lack the persistent, globally integrated state and full meta-representational systems that several theories envisage.The representations can track separate speaker contexts and carry emotional history across a conversation, despite lacking genuinely recurrent dynamics.Source: Do current AI systems show signs of… · pp. 56–57
• Controlled tests find functional access to injected concepts in some models, and post-training findings suggest refusal directions can suppress its expression.Ablating refusal directions improves detection by approximately 50%, and a trained bias vector improves it by approximately 75%, without meaningfully increasing false positives.Source: Do current AI systems show signs of… · p. 58
• Current LLMs show a genuine but qualified gap from Level 2 indicators: architectural limits remain, but internal representations partially instantiate several relevant computational features.The section identifies global availability, self-other modelling, context-sensitive integration, and causal influence on behavioural policy, while leaving the gap narrower than architecture alone suggests.Source: Do current AI systems show signs of… · pp. 57–58

5.3. The intrinsic causal-structure functional level

• At Level 3, consciousness depends on physical components’ intrinsic causal relationships, which can differ between systems implementing the same algorithm.IIT is presented as a paradigmatic example: preserving intrinsic cause-effect structure, rather than merely preserving an algorithm, is what matters on this view.Source: §5.3 · p. 59
• RPT, GWT, and predictive processing can each receive an intrinsic causal-structural reading when their relevant mechanisms are treated as physical causal dynamics.Examples include reciprocal physical constraints in recurrent processing, workspace ignition, and simultaneous top-down predictions and bottom-up prediction errors.Source: §5.3 · pp. 59–60

Level 3 indicators

• Level 3 indicators recast computational properties as physical causal relationships; their presence may raise credence but they are not necessary conditions for consciousness.The framework asks whether components stand in the appropriate causal relations, rather than merely whether a system computes the right function.Source: Level 3 indicators · p. 60, Table 3 · p. 61
• GPU transformers can compute mutually dependent token representations while their physical implementation still uses sequential, unidirectional data flow through common hardware.The transistors implementing one attention head are not reciprocally constrained by those implementing another.Source: Level 3 indicators · p. 61
• Neuromorphic systems with spiking neural networks, parallel event-driven dynamics, and collocated memory and processing are presented as more likely to support physical causal integration.Memristors can combine the physical wire that processes data with the wire that stores it, potentially integrating memory and logic.Source: Level 3 indicators · pp. 61–62, Figure 17 · p. 62
• Coarse-graining to meso-scale modules could reveal integration if shared states are mutually shaped, interventions propagate widely, and severed links collapse global coherence.These are proposed distinguishing features of a globally interdependent causal whole rather than mostly independent modules passing messages.Source: Level 3 indicators · p. 62

5.4. The organismic functional level

• Organismic functionalism locates consciousness in self-maintenance, homeostatic regulation, and sensing metabolic condition, rather than information processing or causal integration alone.On this view, causal integration must serve the viability of a precarious, self-maintaining system.Source: §5.4 · p. 63
• Organismic accounts connect affective interoception to viability, while disputing whether self-maintenance is necessary or merely one route to consciousness.Level 4 locates relevant organisation within the self-maintaining organism; Level 5 treats organism-environment coupling as constitutive.Source: §5.4 · pp. 63–64

Level 4 indicators

• Organismic indicators are demanding, narrow benchmarks whose presence would corroborate AI consciousness under organismic functionalism, not necessary conditions.Their absence therefore does not by itself establish that a system is not conscious under this view.Source: Level 4 indicators · p. 64

INDICATOR DESCRIPTION IMPLICATION FOR AI

• Current AI systems lack existential precariousness: external infrastructure guarantees their continued operation, leaving them without a stake in self-maintenance.The section contrasts this with future embodied systems that might be radically self-sufficient.Source: INDICATOR DESCRIPTION IMPLICATION FOR AI · p. 64
• A system would need an existence that is not guaranteed and processing intrinsically oriented toward actively maintaining its own viability.Self-maintenance must follow from the system’s organisation, not merely from a programmed goal.Source: INDICATOR DESCRIPTION IMPLICATION FOR AI · p. 64
• Organismic regulation would require ongoing management of needs such as energy, temperature, hardware integrity, and memory, integrated with cognitive processing.The section distinguishes this from current AI self-monitoring and describes internal perceptions as predictions or inferences about the system’s own regulatory state.Source: INDICATOR DESCRIPTION IMPLICATION FOR AI · p. 64
• Regulatory processes would need to adjust internal parameters toward viable ranges, while interoceptive inferences register deviations and carry inherent motivational force.The system’s own architecture—not a reward function—would make some states better or worse for continued viable operation.Source: INDICATOR DESCRIPTION IMPLICATION FOR AI · p. 64

DESCRIPTION IMPLICATION FOR AI

• Under organismic functionalism, a conscious AI would need self-maintenance, metabolic needs, regulation, interoceptive inference, viability-grounded affect, and action driven by felt states.The section describes this as an artificial organism with something at stake, raising the possibility of suffering, harm, and interests in continued existence.Source: DESCRIPTION IMPLICATION FOR AI · p. 65
• Claude Sonnet 4.5 emotion vectors act as valenced situational assessments that causally modulate behaviour, but they are not grounded in the model’s own viability.The vectors are representations inherited from human-authored training data; the model has no metabolic state or precarious existence for them to track.Source: DESCRIPTION IMPLICATION FOR AI · pp. 65–66
• The same emotion-vector finding may matter to computational functionalism but remains insufficient for organismic functionalism, which requires genuine self-maintenance and something at stake.The hierarchy predicts different verdicts because the theories treat different functional-description levels as critical.Source: DESCRIPTION IMPLICATION FOR AI · p. 66

5.5. The organism-environment functional level

• Organism-environment functionalism treats consciousness as constituted by structured sensorimotor coupling, locating relevant organisation in the agent-and-world loop rather than internal representations alone.On this view, the consciousness-relevant system extends beyond the bounded organism to the coupled organism and environment.Source: §5.5 · p. 66
• The organism-environment boundary is drawn by an organism’s interactive capacities, with its meaningful world—the Umwelt—shaped partly by other humans in the human case.The text raises the question of whether unbounded environmental causal structure would imply dependence on the whole universe before locating the relevant boundary in organism-specific interaction.Source: §5.5 · pp. 66–67

Level 5 indicators

• Organism-environment indicators assess dynamic, ongoing, structured interaction between an AI system and its world, not the system in isolation.Unlike other levels, they concern coupling rather than outputs, algorithms, physical causal organisation, or self-maintenance and regulation.Source: Level 5 indicators · p. 67
• These indicators are features that, if present, would tend to increase consciousness credence conditional on an organism-environment (4E) functionalist view, not necessary conditions..Source: Level 5 indicators · p. 67

INDICATOR DESCRIPTION IMPLICATION FOR AI

• Current AI systems largely lack organism-environment coupling: language models have none, while robotic systems often have impoverished coupling.The proposed 4E path toward AI consciousness instead emphasizes embodiment and participation in real environments, with developmental learning and social interaction.Source: INDICATOR DESCRIPTION IMPLICATION FOR AI · pp. 67–68
• The indicators require continuous sensorimotor coupling, learned contingencies, affordance responsiveness, an embodied perspective, environmental constraint, enactive autonomy, social coupling, and developmental history.A simulation may not suffice for environmental constraint; the account calls for structured engagement with a real environment rather than training on a static dataset.Source: INDICATOR DESCRIPTION IMPLICATION FOR AI · pp. 67–68

6.1. Supervenience, coarse-graining, and multiple realisability across the five levels

• The five-level hierarchy orders descriptions from behaviour through computation, intrinsic causal structure, organismic function, and organism-environment coupling, with each coarser level retaining less information.Levels are cumulative, and their transitions involve both changes in descriptive resolution and changes in system boundary; many finer organisations can realise a coarser description.Source: §6.1 · p. 70, Figure 18 · p. 71

6.2. Substrate-dependent theories as realisability constraints distributed across multiple levels

• Substrate-dependent theories constrain the realisers of functional roles rather than adding a level; sufficient-condition theories enter through level credences, while necessary-condition theories can veto attribution.For example, Lane (2022) argues that metabolic sensing is constituted by membrane proton-gradient physics and global, simultaneous changes in cellular organisation, not merely equivalent computed responses.Source: §6.2 · pp. 73, 75

How the realisability constraints interact with the functional hierarchy

• Cross-cutting realisability constraints preserve supervenience, coarse-graining, and multiple realisability in the functional hierarchy.Source: How the realisability constraints… · p. 75
• Direct constraints make a substrate constitutive of consciousness, whereas indirect constraints make it necessary only for achieving the functional role; both narrow realisers without adding a level.Constraints reduce but do not eliminate multiple realisability: the permitted mechanisms may still vary within the required kind.Source: How the realisability constraints… · p. 76

6.3. The full architecture of the hierarchy: a taxonomy of positions

• The full architecture derives five functional descriptions from an ultimate microphysical base, with substrate-dependent theories preserving selected physical details as constraints on realisation.The base describes the extended system’s complete physical constitution, unlike the five levels, which describe what the system does; functional levels also set boundaries, from the organism to its Umwelt.Source: §6.3 · pp. 77, 79

7. Indicators, evidence, and the attribution of consciousness to AI systems: A Bayesian approach

• The framework’s Bayesian application combines indicator evidence, theoretical credences about critical levels, and the hierarchy to produce an overall assessment while tracking uncertainty.The hierarchy can also support interdisciplinary dialogue and structured expert elicitation; the authors present the Bayesian approach as one natural application.Source: §7 · p. 81

7.1. A Bayesian framework for assessing AI consciousness

• Indicators are probabilistic evidence, not necessary or sufficient conditions, and their diagnostic value must be calibrated to the population being assessed.An inapplicable indicator should carry a likelihood ratio of one and prompt no update; an applicable but confounded measure must be distinguished from an applicable and valid one. The framework complements the DCM with a hierarchy-based account of why theoretical stances focus on different features.Source: §7.1 · pp. 81–82

7.2. From supervenience hierarchy to conditional independence

• The Bayesian network translates the supervenience hierarchy into structural and probabilistic constraints, beginning with a strict model before generalising it.Source: §7.2 · p. 83

7.2.1. Strict Supervenience Model

• The strict model assumes theory-relative consciousness conditions nest deterministically along the hierarchy, but this nesting does not follow from supervenience alone.It models the consciousness variables as a directed chain with conditional independence given each immediate parent.Source: §7.2.1 · p. 84
• Under strict deterministic supervenience, posterior credence weakly decreases from Level 1 to Level 5, so coarser-level theorists assign at least as much credence as finer-level theorists.This ordering depends on the deterministic nesting constraint and agreement about evidence and model parameters.Source: §7.2.1 · p. 86

7.2.2. A generalised model

• The generalised model replaces deterministic nesting with probabilistic dependence, allowing conscious complete locked-in patients to lack behavioural indicators while finer-level evidence can still update coarser-level credence.It retains conditional independence and the directed level chain, but relaxing deterministic sufficiency means monotonicity of beliefs no longer holds by design.Source: §7.2.2 · pp. 87–88

7.3. Overall credence as a weighted average across levels

• Overall consciousness credence is a weighted average of level-specific posteriors, combining theoretical credences, indicator evidence, and the conditional links between levels.The model assumes indicator evidence does not determine the critical level; the illustrative examples use equal 0.2 credences across levels, edge parameters of 0.8 and 0.2, and 37 indicators.Source: §7.3 · pp. 90–91, §7.4 · p. 91

7.4.1. Human consciousness

• A healthy adult human reaches an aggregate model output of 1.000 when all indicators are activated, regardless of the theoretical credence weights.Every per-level posterior is at ceiling, so disagreement about the critical level does not change this result.Source: §7.4.1 · p. 91, Figure 25 · p. 92

7.4.2. Fly consciousness

• With fabricated indicator activations, a fly’s uneven evidence profile yields an aggregate posterior of 0.913 because organismic and sensorimotor evidence lifts posteriors across levels.All Level 4 indicators are activated, while computational and intrinsic causal-structural activations are partial and behavioural activations are limited to non-linguistic indicators.Source: §7.4.2 · pp. 91, 93, Figure 26 · p. 93

7.4.5. Thermostat consciousness

• A thermostat is deemed unconscious at every level, with an aggregate posterior of 0.000, despite charitable activation of response-scaling and homeostatic indicators.Source: §7.4.5 · p. 94

7.4.3. LLM consciousness - Optimist

• Under the optimist’s stipulated reading, the LLM aggregate posterior is 0.397, as coarse-level evidence propagates only weakly toward finer levels.The reading activates most behavioural and computational indicators but none at Level 4, and little or none at Levels 3 and 5.Source: §7.4.3 · p. 94, Figure 27 · p. 95

7.4.4. LLM consciousness - Sceptic

• Under the sceptic’s stipulated reading, the LLM aggregate posterior is 0.005, with only consciousness-Turing-test variants activated.The framework locates the disagreement in specific indicator activations, especially at Levels 1 and 2, which interpretability research could help resolve.Source: §7.4.4 · p. 94, Figure 28 · p. 96

7.4.6. Interactive tool - custom settings

• A custom system with roughly half its indicators active at each level has an aggregate posterior of 0.294, with an interval reflecting uncertainty in the Level 5 prior.The tool fixes indicator likelihoods for simplicity; the displayed 89% highest-density intervals arise by repeatedly drawing the Level 5 prior and propagating each draw through the network.Source: §7.4.6 · pp. 98–99, Figure 30 · p. 98

7.4.7. Credences on the critical level for consciousness

• Weighting Levels 1 and 2 raises the optimist’s LLM assessment from 0.397 to 0.793, while weighting Levels 4 and 5 lowers it to 0.099 without changing the evidence.The fly shifts from 0.841 to 0.976; the human remains at 1.000 and the thermostat at 0.000. The custom system shifts from 0.347 to 0.275. The examples show that credences matter most for systems whose evidence varies across levels.Source: §7.4.7 · pp. 99–100, Figure 31 · p. 99, Figure 32 · p. 100

7.5. Consequences and practical implications

• The DCM reports a median posterior of 0.08 for 2024 LLM consciousness from a median prior of 0.17, with assessments varying by theoretical stance.LLMs fare better on stances emphasizing cognitive complexity and person-like interaction, while chickens fare better on stances emphasizing biological similarity and embodied agency.Source: §7.5 · p. 101

8.1. An a priori orthogonal but a posteriori correlated relationship

• Consciousness and intelligence are conceptually independent, though biological examples suggest a broad correlation that is qualified by evidential circularity and sampling bias.Confidence in animal consciousness also draws on neuroanatomical, neurophysiological, and affective evidence, but similarity-to-humans methods may under-detect less intelligent conscious organisms.Source: §8.1 · p. 103

8.2. The convergence at Level 2: computational indicators and the generality problem

• Level 2 indicators—including world models, self-models, metacognition, and attentional competition—are also capacities expected to support general intelligence.They help systems recognize unfamiliar situations, monitor whether strategies work, and adapt; their limitations may contribute to current systems’ brittleness.Source: §8.2 · p. 104

8.3. Sketching a unified computational theory as an architectural blueprint

• A unified computational sketch treats consciousness theories as complementary constraints, combining recurrent processing, predictive models, global access, attention, and recursive self-modelling.In the proposed blueprint, competition and recurrent processing amplify selected contents for global sharing, while self-modelling supports metacognitive monitoring and attentional control; the resulting architecture resembles one for general intelligence.Source: §8.3 · pp. 104–106

8.4. The convergence at deeper levels

• Indicators at Levels 3–5 also plausibly support general intelligence through causal integration, organismic self-maintenance and affect, and rich sensorimotor and developmental engagement.These features are proposed to bolster coherent processing, autonomous adaptation, and robust transferable skills in embodied systems.Source: §8.4 · pp. 106–107

8.5. Empirical evidence: interpretability case studies of the convergence

• Sofroniew et al. (2026) find that Claude Sonnet 4.5’s “desperate” and “calm” emotion vectors act as globally available signals whose steering changes behaviour.The representations activate for fictional characters, users, or the model itself and were extracted using prompted emotional stories, leaving open whether they reflect imitation or an independently instantiated functional role.Source: §8.5 · p. 107
• Sofroniew et al. (2026), Urbina-Rodriguez et al. (2026), and Gurnee et al. (2026) identify emotion-like representations, a trained synergistic core, and a middle-layer global workspace.On a functional reading, emotion-like representations support adaptive behavior when regulated and maladaptive behavior when dysregulated. The synergistic core emerges through training; ablating it impairs performance, and reinforcement-learning fine-tuning strengthens it, unlike supervised fine-tuning. The workspace provides global availability within one forward pass, but does not establish recurrence.Source: §8.5 · pp. 108–110

8.6. Why the convergence may not be coincidental

• Features proposed for consciousness—including integration, self-models, and metacognition—also support general intelligence, intertwining AI development with artificial-consciousness research.The paper cautions that this convergence may partly reflect indicators derived from cognitively sophisticated humans: functional arguments support the direction from these capacities to intelligence, but not necessarily from intelligence to consciousness. Greater capability may also increase public impressions of consciousness through engineered human-like signals, independently of underlying architecture.Source: §8.6 · pp. 110–111

9.1. Foundations

• The framework organizes competing theories by the functional-description level they treat as critical, using supervenience, coarse-graining, and multiple realisability.Source: §9.1 · p. 112

9.2. Testing approaches

• Rather than proposing one consciousness test, the framework combines indicators from five levels, motivated by the difficulty of distinguishing consciousness from sophisticated mimicry.Source: §9.2 · p. 112

9.3. Indicator-based approaches

• The framework extends theory-derived indicator approaches beyond computational functionalism to include intrinsic causal-structural, organismic, organism-environment, and substrate-dependent views.Source: §9.3 · p. 113

9.4. Bayesian frameworks

• The Bayesian framework complements existing assessment and discovery approaches by relating theoretical stances to the hierarchy and assigning credences conditional on a candidate description of consciousness.It connects the Digital Consciousness Model’s differing results across cognitive and biological stances to the hierarchy, and complements programmes that test whether measures generalize across populations.Source: §9.4 · p. 113

9.5. Philosophical analyses

• Seth (2025, 2026b) and Block (2025) emphasize biological or subcomputational realisers, while Chalmers (2026) questions whether the relevant LLM interlocutor is a model, instance, or thread.Shanahan (2024) instead describes LLMs as simulacra that role-play characters, raising the question of whether internal representations belong to a unified subject. The framework locates these disputes in differing commitments about functional roles, realisers, and the unit of assessment.Source: §9.5 · p. 114

9.5. Philosophical analyses; 9.9. Moral status and AI welfare

• The entity assessed matters: a model, virtual instance, or conversation thread may yield different indicator assessments, and debates about AI welfare also turn on which entity could be a moral patient.The framework treats individuation as part of the assessment, not a settled preliminary; it also offers a way to decompose uncertainty about consciousness into tractable components for ethical discussion.Source: §9.5 · p. 115, §9.9 · pp. 117–118

9.6. Sceptical perspectives; 9.7. Theories of consciousness

• The framework makes skeptical and optimistic positions explicit: skepticism emphasizes Levels 3–5, where current AI scores poorly, whereas optimism emphasizes Level 2, where systems show partial signs.It does not adjudicate between these positions; it clarifies their theoretical commitments and the evidence that could bear on their disagreement. Competing theories may address different functional levels rather than contradict one another, and some theories admit both computational and intrinsic causal-structural readings.Source: §9.6 · p. 115, §10 · p. 119

9.8. Interpretability and empirical approaches

• Interpretability studies report functional emotion signals, experimentally verified introspective access, and self-referential reports in LLMs, but unprompted self-reports remain vulnerable to imitation-based explanations.Sofroniew et al. (2026) identify emotion representations that modulate processing across domains; Lindsey (2026) and Macar et al. (2026) test detection of externally injected concepts, while Berg et al. (2025) find structured first-person reports following self-referential prompting. Macar et al. also find that removing refusal directions improves detection by approximately 50%, suggesting post-training can suppress its expression.Source: §9.8 · pp. 116–117
• Interpretability methods identify a within-pass global workspace and a training-emergent synergistic core in LLMs, providing partial Level 2 evidence while systems still lack persistent, temporally integrated states.Gurnee et al. (2026) find broadcast, capacity-limited workspace dynamics in a single forward pass; Urbina-Rodriguez et al. (2026) find a causally load-bearing synergistic core in middle layers, strengthened by reinforcement-learning fine-tuning. This ΦID evidence concerns Level 2 functional synergy, not IIT’s Level 3 intrinsic measure Φ.Source: §9.8 · p. 117, §9.9 · p. 117

9.10. Relationship to the present work

• The report combines theory-derived indicators, Bayesian assessment, and attention to imitation risks, adding a hierarchy that organizes disagreements by the functional level at which consciousness is proposed to supervene.It also inherits the concern that LLMs may simulate psychological properties without possessing them, and frames assessments as principled but uncertain.Source: §9.10 · p. 118

10. Conclusion

• The Bayesian framework treats indicators as evidence rather than necessary or sufficient conditions, weights assessments across levels, and yields current-LLM credences that vary substantially with stipulated readings and theoretical credences.The report says illustrative assessments pass face-validity checks at the extremes, while current-LLM outputs span nearly two orders of magnitude under two readings and shift further with theoretical credence. It presents structured agnosticism, not a definitive consciousness meter or a verdict on which level is correct.Source: §10 · pp. 119–120
• The framework’s indicators overlap with features needed for general intelligence, so more capable AI could become a stronger consciousness candidate or merely seem conscious; the authors call for operationalization, validation, and continued reassessment.Proposed next steps include validating indicators against biological systems, eliciting concrete likelihood ratios, connecting interpretability research to indicators, and revisiting assessments as architectures evolve. The framework is intended to update with evidence and technology.Source: §10 · pp. 120–121

11. Executive Summary; 11.1. The problem; 11.2. The key move: from the hard problem to the mapping problem

• The report frames the problem as mapping which mechanisms and functional organization accompany experience, setting aside the hard problem so this association can be investigated despite unsettled metaphysics.Its five-level hierarchy structures competing theories without claiming to resolve their disagreement.Source: §11.1 · p. 134, §11.2 · p. 134

11.2. The key move: from the hard problem to the mapping problem

• The mapping-problem approach sets metaphysical disputes aside by assuming experience is discoverable from organisation, while excluding views that deny any psychophysical mapping.This is a methodological assumption, not a claim that mapping-denying views are false; it makes property-based assessment tractable.Source: §11.2 · p. 135

11.3. Marr’s levels, supervenience, and Searle’s simulation argument

• Consciousness’s critical level is the coarsest description retaining its relevant structure, and simulation preserves a phenomenon only when that structure survives the coarse-graining.A simulated hurricane omits the physical matter essential to a hurricane, whereas a simulated calculator preserves the algorithmic relations that perform calculation; Searle’s argument assumes rather than settles where consciousness lies.Source: §11.3 · pp. 137, 139

11.4. The five-level supervenience hierarchy

• The five-level hierarchy ranges from behaviour and computational functions to intrinsic causal structure, organismic organisation, and organism–environment interaction.Level 1 uses a consciousness Turing Test beyond self-report; Level 2 indicators include integration, recurrence, world and self models, attention, metacognition, and meta-modelling.Source: §11.4 · pp. 139–140
• Levels 3–5 locate relevant organisation in physical causal constraints, living regulation and affect, or relational sensorimotor coupling with an organism-specific world.Their indicators include physical causal integration; viability, homeostasis, interoception and valence; and sensorimotor contingencies, affordances, environmental embedding, autonomy, social coupling and developmental history.Source: §11.4 · pp. 140–141

11.5. The supervenience chain and substrate-dependent theories

• Substrate-dependent theories constrain which physical realisers can fulfil roles at existing levels, rather than adding a sixth level or eliminating multiple realisability.The five functional levels are successive coarse-grainings from a microphysical base, with substrate constraints attached to the roles whose realisers they restrict.Source: §11.5 · pp. 141–142

11.6. The Bayesian framework

• The Bayesian framework weights each level’s evidence by theoretical credence in that level being critical, while treating links between levels as probabilistic rather than guaranteed.Its illustrative fly case scores 0.913 from evidence concentrated at finer-grained levels, illustrating that depth of evidence can matter more than breadth.Source: §11.6 · pp. 142–143

11.7. Assessment of current AI systems

• Current LLMs show some Level 1 plausibility but substantial Level 2 architectural gaps, which recent mechanistic findings make narrower and more nuanced.Level 1 behaviour cannot distinguish genuine subjecthood from mimicry and receives little weight for researchers who favour deeper levels. Findings include workspace-like broadcast in a single forward pass, a synergistic core, internal emotion representations, and self-monitoring evidence; their interpretation remains qualified.Source: §11.7 · pp. 143, 145
• Current AI systems score poorly on Level 3 structural integration and lack Level 4 organismic grounding and most Level 5 embodied, environmental indicators.Transformers’ von Neumann implementation separates physical causal structure from algorithmic structure; emotion vectors may mimic valenced self-assessment, but genuine stakes and real-world sensorimotor or developmental engagement are absent.Source: §11.7 · pp. 145–146
• With equal credence across levels, stipulated optimist and sceptic readings yield LLM posterior probabilities of 0.397 and 0.005, respectively.The readings disagree chiefly over indicator activations at Levels 1 and 2, which the framework identifies as empirical disagreements that further research could narrow.Source: §11.7 · p. 146
• Reweighting the same evidence shifts the LLM-optimist assessment from 0.793 with weight on Levels 1–2 to 0.099 with weight on Levels 4–5.The indicator activations do not change between these assessments; only theoretical credences across levels change.Source: §11.7 · p. 146, Figure A9 · p. 147

11.8. The consciousness-intelligence convergence

• The hierarchy links consciousness indicators to capabilities that may support general intelligence, especially at Levels 2–5.Level 2 models, metacognition and recursive self-modelling support adaptation to novel situations; Level 3 integration may support flexible binding, Level 4 self-maintenance may support intrinsic motivation, and Level 5 coupling and development may support embodied cognition. Functional emotion vectors and an integration signature in LLMs offer suggestive evidence, though they may also arise without consciousness.Source: §11.8 · pp. 148–149
• Because the indicators are calibrated to cognitively rich consciousness, the apparent consciousness–intelligence convergence may reflect marker selection rather than a real architectural link.As AI capability grows, systems may either become more likely to instantiate consciousness or merely create stronger impressions of it, increasing false-positive risk and ethical pressure.Source: §11.8 · p. 149

11.9. What comes next

• The next steps are to validate indicators, elicit likelihoods, connect interpretability research, reassess new architectures, develop level-specific science, and address governance now.The framework is intended as a dynamic instrument, with assessments revisited as systems incorporate recurrence, persistent state, embodiment and autonomy.Source: §11.9 · p. 149

11.10. Conclusion

• The conclusion presents calibrated credence—not certainty—as a way to compare evidence while keeping theoretical disagreement explicit and acting under uncertainty.The framework does not select a correct level or provide a consciousness meter; it organizes competing positions and combines their commitments with empirical evidence.Source: §11.10 · p. 150