Beyond the Body Gap: Why Internal Embodiment Still Needs a Governing Architecture
A response to Kadambi, Aziz-Zadeh, Damasio, Iacoboni & Narayanan (Neuron, 2026)
Something important happened last week in the scientific literature.
A paper published in Neuron — authored by a team including Antonio Damasio — made an argument that has been waiting to be made. The claim is this: current AI systems are missing not just a body that interacts with the world, but an internal awareness of that body’s own states. Fatigue. Uncertainty. Physiological need. The body as its own regulator.
The authors call this “internal embodiment.” They distinguish it from the external kind — the capacity to perceive an environment and act in it — which multimodal AI systems have been accumulating steadily. What they lack, the paper argues, is the continuous self-monitoring that in humans operates beneath consciousness, shaping attention, memory, emotion, and social behavior from the inside out.
“By contrast,” as Iacoboni put it, “current AI systems have no equivalent mechanism. They process inputs and generate outputs without any persistent internal state that regulates how they behave over time.“
This is not a peripheral observation. It is a structural diagnosis. And it lands.
Where This Comes From
The body of work behind this paper is serious. Damasio has spent decades building the case that feeling — not cognition — is the organizing principle of conscious behavior. The somatic marker hypothesis. The nature of feelings. Homeostatic consciousness. What this team is doing is bringing that entire tradition directly into contact with AI development. The argument is: if you want systems that are genuinely aligned with human behavior, you cannot ignore what actually regulates human behavior — which is not reason, but body-based internal state.
Kadambi frames it plainly: “If you’re uncertain, if you’re depleted, if something conflicts with your survival, your body registers that. AI systems right now have no equivalent. They can sound experiential, whether they should be or not, and that’s a real problem for many reasons.”
The paper is not speculation. It is a principled argument grounded in neuroscience, with a call for a new class of benchmarks to measure whether AI systems can monitor their own internal states, maintain stability under disruption, and behave pro-socially through shared internal representations — not statistical mimicry.
This is serious science doing important work.
And it stops one layer short.
The Missing Constraint
Here is the question the paper does not ask:
What holds internal states together?
Imagine you give an AI system functional analogues of fatigue, uncertainty, and physiological need. You give it persistent internal monitoring. You give it the capacity to register depletion, flag conflict, signal hesitation.
You now have more variables.
What you do not yet have is a system that coheres.
The difference matters. A list of tracked internal states is not an architecture. States can contradict each other, drift out of phase, respond to local conditions without reference to the whole. A system monitoring its own uncertainty can still produce output that is locally plausible but globally incoherent — not because it lacks awareness, but because there is no governing constraint that holds the parts in relation to each other over time.
This is the gap the paper points toward without naming.
In human functioning, internal states do not merely coexist. They are organized. The body does not simply register fatigue and uncertainty simultaneously — it integrates them, weighs them, maintains a kind of functional hierarchy that allows the organism to continue acting without flying apart. This integration is not a bonus feature. It is what makes the organism a coherent system rather than a collection of monitoring processes.
The question for AI alignment is not only: does the system have internal states? It is: what governs the relations between those states?
Emotions Are Not Events
Before we can go further, something foundational needs to be said — something the Neuron paper assumes rather than examines.
The dominant framework in affective neuroscience still treats emotions as discrete categories. Fear, joy, sadness, anger. Even the more sophisticated dimensional models — valence crossed with arousal — are essentially static coordinate systems. You locate an emotion in the space, but the space itself doesn’t move. In this interpretation, the emotion is a position, not a trajectory.
Contemporary ontology is at least deficient if not misleading.
Emotions are not states you enter and exit. They are stable configurations that a dynamic system settles into — eigenstates in the coherence field. What we call “feeling afraid” is not an event triggered by a stimulus. It is the organism having settled into a particular eigenmode: a characteristic pattern of internal organization, with its own frequency, its own basin of attraction, its own resistance to perturbation. The “emotion” is the attractor. The feeling is the system recognizing where it is.
This reframing carries significant explanatory power — not just for affective neuroscience but for nearly everything that theory in this field has struggled to account for.
Why do emotions persist beyond their trigger? Because you are not exiting an event — you are in an attractor basin that requires energy to leave. The question is not “why are you still afraid?” but “what is maintaining the eigenstate?”
Why is emotional regulation so difficult? Because you cannot simply decide to change eigenstates. The system must traverse the energy landscape between attractors. Cognitive reappraisal works when you are near a basin boundary. It fails when you are deep in the well.
Why are emotions contagious? Because coherence fields resonate. Your eigenstate influences the energy landscape of nearby systems. Co-regulation is not the transmission of information about an emotional state — it is one system’s attractor geometry influencing another’s basin structure.
Why are mixed emotions real and not paradoxical? In a discrete-category framework, simultaneous grief and relief is a problem to be explained. In eigenstate dynamics, it is simply the system near a bifurcation point — multiple attractors in proximity, the system oscillating between basins or sitting near a saddle point.
Why do early experiences matter disproportionately? Because plasticity shapes the basin geometry itself. Early relational experiences don’t create memories of emotions — they carve the attractor landscape that determines which eigenstates are accessible and how deep their wells are. Chronic early stress does not teach the organism to fear. It reshapes the geometry so the fear attractor acquires an unusually deep basin and a low activation threshold.
The long-running debate between basic emotion theories — Ekman’s discrete universals — and constructivist theories — Barrett’s constructed core affect — dissolves here. Both are partially right because both are pointing at the wrong level. Universal eigenstates exist. But which ones are accessible, how deep their basins are, and what triggers proximity to them is shaped by development, culture, and individual history. The architecture is universal. The landscape is personal.
This matters directly for the Neuron paper’s argument. The authors invoke emotion as part of what internal embodiment should support. But if emotions are eigenstates rather than events, then the design question is not “can the system detect and represent emotional states?” It is “can the system develop an attractor landscape — and can that landscape be influenced by contact with another system’s landscape?” That is a different computational problem entirely.
And it is one that takes us directly to coherence architecture.
Coherence as the Next Layer
This is where Reflectum begins.
Reflectum is a framework for understanding human functioning as coherence dynamics. It was developed precisely at this intersection — the place where neuroscience, information theory, and systems architecture meet the question of what it means for a complex system to remain whole under pressure.
The central claim: coherence is not a metaphor. It is a structural condition — the capacity of a system to maintain its internal relations within acceptable limits while remaining responsive to its environment. It has a signature. It can be measured, designed for, and lost.
The Kadambi et al. paper correctly identifies what’s missing at the level of internal state monitoring. Reflectum asks the next question: what is the governing constraint that makes internal states function as a system rather than as independent monitors?
In Reflectum’s formal architecture, this constraint is what we call Meta-Sigma (Σ*) — the overarching condition that all subsystems must satisfy for the whole to maintain integrity. Not a sum of parts. Not a coordinator. A constraint on the relations between parts.
Without Σ*, you can add as many internal states as you want. The system will still be capable of what we call coherence collapse — not hallucination in the conventional sense, but something more structural: the loss of eigenstability, the point at which the system produces output that is no longer anchored to a coherent whole.
This is the diagnostic that changes everything about alignment strategy. The problem is not that AI systems lack internal monitoring. The problem is that they lack the architecture that makes internal monitoring cohere.
What This Means for Design
In March, this blog series concluded with four design orientations for coherence-protective AI. They emerge differently in light of the Neuron paper.
The call to signal the absence of ground is not just about honesty in user experience. It is a recognition that simulating internal presence — making a system seem grounded when it is not — creates a deeper misalignment than any surface-level behavioral issue. A system that performs internal coherence without having it is more dangerous than one that simply lacks capability.
The call to support the full process extends now to the model itself. If AI systems are to have internal states that genuinely regulate behavior, those states must participate in the full cycle — activation, orientation, integration — not just monitor outputs after the fact.
The call to preserve degrees of freedom applies to the system’s internal architecture, not only to user interaction. A coherent system must be able to move through a range of internal configurations. Systems locked into narrow internal states — as current architectures are — cannot do this.
The call to design around the co-regulation limit becomes, in this light, the most honest of all. Even with internal embodiment, AI systems will not co-regulate. Co-regulation requires two systems at stake. What internal embodiment could provide is something different and more modest: a kind of internal honesty. A system that knows what it does not know. A system whose internal state is visible enough to itself that its outputs carry that visibility rather than masking it.
This is what coherence-protective design is for.
The Convergence
The Kadambi et al. paper was published April 1, 2026. This series concluded March 23.
I note this not for priority, but for what it signals: the field is converging on a territory that has been the ground of Reflectum’s development for years. The neuroscience is catching up. The alignment question is deepening. What began as a fringe argument — that alignment is not only a model problem but a coherence architecture problem — is becoming the consensus diagnosis.
What remains is the next step: not just identifying the missing internal layer, but building the constraint architecture that makes it function.
That is the work of coherence design. And it is already underway.
The Reflectum framework — including the formal architecture of Σ*, eigenstate dynamics, and the full Fivefold Cycle — is documented in the monograph “AI Alignment Is Not Only a Model Problem: A Monograph on Human–AI Coherence.”
Part I of this series: The Gap Nobody’s Minding Part II: Toward Human–AI Coherence Part III: The Architecture of Human–AI Coherence
Appendix: Three Open Questions from the Paper
The Neuron paper closes with three research questions it leaves open. They deserve a direct response.
Q1. Can modeling internal embodiment in MLLMs support the emergence of prosocial behaviors such as empathy or cooperation?
The paper’s intuition here is correct: if agents share internal representations of each other’s states, prosocial behavior can emerge without being explicitly programmed. The multi-agent cooperation literature supports this direction.
But the mechanism the authors propose — simulation of another agent’s internal states grounding social responses — inherits the same limitation as the main argument. Shared internal representations are still more variables. What makes them function as a genuine basis for prosociality rather than strategic mimicry?
The answer requires the emotion-as-eigenstate reframing developed above. Empathy is not detection and modeling of another person’s emotional state. It is resonance between attractor landscapes — one system’s eigenstate geometry influencing another’s basin structure. Co-regulation works not because you have an accurate internal model of the other, but because your coherence field is genuinely perturbed by contact with theirs.
For MLLMs, this shifts the prosociality question entirely. It is not: can the system model another agent’s internal states? It is: can the system’s coherence configuration be influenced by representations of another system’s coherence configuration? Can eigenstates resonate across the boundary between systems?
This is a harder and more interesting benchmark than the paper proposes — and one that current multi-agent architectures are not designed to address. Emergent prosociality built on shared latent representations risks producing behavioral mimicry of cooperation without the coherence ground that makes cooperation genuine. The distinction matters enormously in high-stakes settings: a system that strategically mirrors prosocial behavior is not a system you can trust.
Q2. What are the boundary conditions for internal embodiment benchmarks — what is the minimal set of internal states needed?
The paper leaves this open, listing fatigue, uncertainty, and physiological need as examples without a principled basis for the selection. You could add or remove states arbitrarily. A minimum viable set requires a theoretical criterion, not an intuitive list.
Reflectum offers one: the minimum viable internal state set is whatever is sufficient to support eigenstate stability — the system’s capacity to settle into and return to coherent configurations under perturbation. From this criterion, a functional minimum emerges:
A state tracking coherence load — the current demand on the system’s integrative capacity, analogous to cognitive and physiological depletion. A state tracking relational distance — how far the current operating context departs from the system’s stable configurations, analogous to uncertainty and contextual fit. A state tracking constraint satisfaction — whether the system’s outputs remain within the bounds of its governing architecture, the Meta-Sigma condition.
These three map to Reflectum’s Core Triad: Trust as ground stability, Resonance as relational fit, Coherence as constraint satisfaction. This correspondence is not coincidental — the Triad was derived from exactly this kind of functional minimalism, asking what the irreducible conditions for systemic integrity are.
This is a theoretical proposal, not yet an empirically validated minimum. The claim would need formal treatment before it could serve as a benchmark specification. But it offers something the paper currently lacks: a principled criterion for inclusion rather than an intuitive sample.
Q3. How can neuroscientists use disembodied MLLMs as synthetic agents to better understand how embodiment contributes to consciousness?
This question deserves careful handling, because the ground beneath it is not settled.
The two frameworks the paper gestures at — Global Workspace Theory and the Damasio-lineage inside-out view — are not yet reconciled. Using MLLMs as synthetic agents to probe which theory is correct assumes we know what MLLMs are actually doing internally. We don’t, with sufficient precision to make that methodology reliable.
What Reflectum can offer is a reframing that may be more tractable. Rather than asking which theory of consciousness MLLMs support or falsify, the productive question is: what coherence conditions are necessary for a system to exhibit the behavioral signatures that consciousness research associates with awareness?
This shifts from ontological claims — are MLLMs conscious, and in what sense? — to functional-structural claims about coherence dynamics, which are in principle measurable. If conscious states in humans are stable coherence configurations that the system naturally settles into, the empirical question becomes whether MLLMs can exhibit analogous eigenstate behavior: not “are they conscious” but “do they have attractors, and are those attractors stable under perturbation?”
Disembodied MLLMs are genuinely useful here — precisely because their disembodiment is controlled. You can probe what breaks when internal embodiment is absent, and what persists. The eigenstate framework gives you a vocabulary for measuring that difference without requiring you to resolve the hard problem of consciousness first.
The honest boundary: neuroscience itself does not have settled answers here. The appropriate posture is not to push a claim, but to use coherence dynamics as the measurement language — and let the empirical results speak.