NOTE-1057: Safron, Klimaj & Sheikhbahaee — IWMT and the Human Consciousness Hypothesis

Contribution

A short position paper for the AAAI Spring Symposium. It summarises Integrated World Modeling Theory (IWMT) and lines it up with the Human Consciousness Hypothesis (HCH), which it attributes to the California Institute for Machine Consciousness (CIMC). Its contribution is the alignment and a compact statement of IWMT’s commitments, above all that consciousness needs embodied grounding. It offers no argument for IWMT beyond what the earlier papers give, and those were not read.

Key insight

On IWMT, an experience is the brain’s best current estimate of the state of an embodied sensorium, like a generated image but made of sensations, and it is conscious because it is coherent in space, time and cause and can both shape and be shaped by the agent’s acting. Integration alone and broadcast alone are each said to fall short; what they lack is a coherent model tied to a body acting in a world.

Assumptions

• HCH’s three principles (p. 345): Genesis (consciousness emerges early as a learning algorithm), Coherence (a “cortical conductor” that minimises constraint violations) and Second-order perception (perception of perceptual processes, in synchrony with them). They are stated, not argued, and not referenced.
• Marr’s levels, with a phenomenological level added (Figure 1).
• Kant’s categories as preconditions of experience: coherence “with respect to space (as organized locality), time (as proportional changes in space), and cause” (p. 346).
• The free-energy principle and active inference, with IIT and GNWT taken as partly right (p. 345).
• A minimal self. “some degree of basic self-modeling and agency may be required for coherent subjectivity”, though it “may potentially be extremely minimal” (p. 346).

Key results

There are no new results. By section:

• Abstract and introduction (p. 345). IWMT’s central proposal is that phenomenal consciousness is what it feels like to be “the spatiotemporally and causally coherent functioning of a probabilistic generative world model for an embodied agent”. “Self-organizing harmonic modes” (SOHMs), synchronous neural complexes, implement it as maximum a posteriori estimates of sensorium states. “Critically, IWMT requires (body-)world models to be capable of both informing and being informed by action-perception cycles at behaviorally relevant timescales.” It ends that consciousness could be realised in artificial systems “with appropriate recurrent dynamics and sufficient degrees (and kinds) of embodied grounding”.
• Integrating theories (p. 345). IWMT draws mainly on GNWT, IIT and the free-energy principle, and claims compatibility with recurrent processing, predictive processing, dynamic core, temporo-spatial, higher-order and attention-schema theories, “While not necessarily endorsing all claims”.
• Consciousness as coherent modelling of self and world (p. 346). World models let a system predict and simulate beyond what it observes. “IWMT suggests that without sufficiently rich physical or virtual embodiments, systems such as large language models are not likely to be able to generate the kinds of world modeling underlying robust general(izing) intelligence, nor consciousness”. Experience is generated “either perception or imagination when respectively coupled to or decoupled from sense-data”. No pre-existing theory alone is said to give sufficient conditions.
• Coherence and minimal selfhood (p. 346). Coherence in space, time and cause, and perhaps integration into a self-model, are necessary for “any kind of subjective experience whatsoever”.
• Embodiment (p. 346). “IWMT is distinct from the theories it draws upon in that it is fundamentally cybernetic and embodied in its approach.” Against IIT: systems could have “arbitrarily large amounts of integrated information and still not be conscious without embodied grounding”. Against GNWT: systems “could instantiate functional global workspaces without entailing any kind of subjective experience” unless the workspace models a coherent, embodied world. With the enactivists, consciousness is “inseparable from embodied cybernetic action–perception cycles”, but IWMT “diverges from strong enactivism in retaining brain-internal representations”.
• Learning to model the world (pp. 346–347). The section opens with the question: “What are the necessary and sufficient conditions for coherently generating a stream of experience, organized according to coherently structured perspectival reference frames?” The answer given: consciousness is a mathematical object, an iteratively computed MAP estimate of system–world relations, informally “akin to a” deep fake of sensorium states, including interoceptive ones. The model is “organized by egocentric perspective”. It is called an adaptation and a “digital twin”.
• Substrates (p. 347). SOHMs emerge as metastable attractors and implement turbo coding by loopy message passing. Experiences are “causally efficacious with respect to agency”. Alpha-band complexes in posterior cortex bind sensory states, theta-band coupling brings in frontal and hippocampal systems. IIT’s posterior hot zone and GNWT’s frontal network are both “proposed to be correct”, for phenomenal consciousness and access respectively.
• Figure 1 (p. 348). A map of the brain at phenomenological, computational, algorithmic and implementational levels, with machine learning architectures on the algorithmic level and beta, alpha and theta complexes on the implementational level. Frontal cortex “might only indirectly contribute to consciousness”.
• Machine consciousness (pp. 347–349). Against “wetness” arguments (Koch, Seth), “a simulation capturing the emergent properties of interacting water molecules would recapitulate the phenomenology of wetness within the simulated world” (p. 349). Yet realisation “may practically (and perhaps also in principle) require the parallelism and physical recurrence of neuromorphic and/or analogue computing systems”, and “non-recurrent networks (such as transformer-based large language models) lack the functional closure, robustness, and efficiency required for realizing the computational properties that generate consciousness” (p. 349). A feedforward workspace “would be divorced from the environmental couplings and historicity by which adaptive predictive (world-)models are learned” (p. 349).
• Conclusion (p. 349). “Thus, there appears to be substantial convergence between IWMT and the HCH.” Genesis and Coherence match IWMT’s early, self-organised coherence; second-order perception matches deep temporal models in which high-level beliefs stabilise low-level ones, including body and attention schemas.

Claims

id	claim	strength	support
C1	Phenomenal consciousness is the coherent functioning of an embodied agent’s generative world model	asserted; argued in earlier papers not read	pp. 345–346
C2	Embodied grounding is necessary: integration or a workspace without it does not suffice	asserted	p. 346
C3	The world model may run decoupled from sense data, as imagination	asserted	p. 346
C4	Spatiotemporal and causal coherence, and perhaps a minimal self-model, are necessary for any experience	asserted, after Kant	p. 346
C5	SOHMs at alpha and theta frequencies are the substrate; IIT and GNWT each describe one part	speculative	p. 347
C6	Transformer language models lack the recurrence, closure and embodiment consciousness needs	asserted, citing Safron 2022	pp. 346, 349
C7	IWMT and HCH substantially converge	asserted; HCH is not referenced	p. 349

Concepts

• Integrated world modelling: consciousness as coherent, embodied self-and-world modelling, across Marr’s levels.
• Self-organizing harmonic modes (SOHMs): synchronous neural complexes, metastable attractors that act as integrated cores and workspaces.
• Embodied grounding: the world model’s capacity to inform and be informed by action–perception cycles at behavioural timescales, and the history of such coupling by which it was learned.
• HCH: Genesis, Coherence and Second-order perception (CIMC).

Connections

It restates Safron 2020a, 2020b, 2021a, 2022 and 2023, and draws on Thompson & Varela’s radical embodiment, Metzinger’s minimal phenomenal experience, Cleeremans’s radical plasticity thesis and Wiese’s minimal unifying model. The record holds Safron’s editorial on world models (LIT-195, NOTE-136), which already cites IWMT as a possible functional account. Laukkonen et al. (LIT-1363) and Whyte et al. (LIT-1364) both cite Safron’s IWMT papers, the latter only to set them aside as conceptual models. It cites COGITATE (Ferrante et al. 2025, LIT-973 in the record), and Seth & Friston 2016 for interoceptive inference in feeling.

Bearing on the record

• CLAIM-413 (grounding), and a correction to the exchange. The exchange cited this paper as requiring “embodiment and sensorimotor coupling, where the owner’s thesis does not require current grounding, so the two part on that point”. The reading changes where they part. IWMT does not require current sensory coupling: the model generates “either perception or imagination when respectively coupled to or decoupled from sense-data”. It requires a capacity to inform and be informed by action–perception cycles, and the “historicity” of coupling by which the model was learned. So it agrees with the claim’s occurrence half for current grounding, and parts from it on developmental grounding and on the capacity for coupling, the second and third of A6’s four independence theses, about which the claim says it is silent. IWMT is then not an objection to the claim as stated. It is the enactive objection the claim leaves open, made specific. It also allows “virtual” embodiment, which the exchange did not mention.
• QUESTION-254. The paper asks nearly the question itself: “the necessary and sufficient conditions for coherently generating a stream of experience, organized according to coherently structured perspectival reference frames” (p. 346). Its answer is the IWMT bundle: coherence in space, time and cause; egocentric organisation; embodied grounding; recurrence. The separations it draws from IIT and GNWT have the form the question asks for (integration, or broadcast, without the bundle is not enough). But the bundle is asserted, not derived, and embodiment is defined by degree (“sufficient degrees (and kinds)”), so it does not give the verdict on a firm or a language model that follows from the property rather than from intuition. On language models it gives one anyway.
• CLAIM-412 (constitution). IWMT’s phenomenal consciousness, what it is like to be the coherent functioning of a world model, is an identity of the claim’s kind, with a different organization: a world model of an embodied sensorium, not interpretive organization in TERM-073’s sense. It is a rival candidate, and it builds in the embodiment the claim does not assume.
• TERM-072 (perspective). IWMT makes egocentric, perspectival organization part of experience and suggests a minimal self-model may be required. That is an organizational and perhaps a represented perspective, against Laukkonen et al., who make a first-person perspective optional (NOTE-1063).
• THEORY-063 (unity without a self-model). IWMT hedges against its second clause: “some degree of basic self-modeling and agency may be required” (p. 346). It offers no case.
• QUESTION-255 (unity). Unity on IWMT is coherence of one estimate, the MAP state. That is the single-assignment reading, and the paper does not discuss conflict held within one experience.
• CLAIM-411. IWMT accepts that much modelling is not conscious (not every image in Figure 1 is “meant to indicate conscious experiences”), and its condition is coherence plus embodiment. It does not test that condition against blindsight or masked priming.
• THEORY-767. Interoceptive inference enters the sensorium estimate, after Seth & Friston 2016; nothing on ownership.
• No instruction for machine-learning practice. Its remarks on JEPA, transformers and neuromorphic hardware are about what might be conscious, not how to build models.

Limitations

• A symposium summary: no data, no derivation and no new argument. Its claims rest on Safron’s earlier papers, which were not read.
• HCH is described without a reference, so its principles cannot be checked against their source here.
• “Embodied grounding” is never given a threshold or a test; it comes in “sufficient degrees (and kinds)”.
• The paper argues both that a simulation would recapitulate phenomenology within the simulated world and that consciousness may “in principle” require physical recurrence. It does not reconcile the two.
• The extracted text of the PDF is out of order because of the two-column layout and an abstract that runs into the introduction; one reference (Ferrante et al.) is broken across a column.

Open questions

• What would show that a system lacks “sufficient” embodied grounding, rather than that it lacks experience for some other reason?
• Does developmental embodiment suffice, so that a world model learned in a body and then cut off stays conscious? IWMT’s “historicity” suggests yes, which is CLAIM-413’s case.
• Where in HCH’s own statement are the three principles argued?