Antonio Damasio Somatic Markers Affective Consciousness and What the Body Requires of AI
Antonio Damasio is David Dornsife Professor of Neuroscience at the University of Southern California and co-director of the USC Brain and Creativity Institute. His work on the relationship between emotion, body, and consciousness has been among the most widely cited in affective neuroscience since his 1994 book “Descartes’ Error: Emotion, Reason and the Human Brain.” His somatic marker hypothesis, developed through a series of books and papers over three decades, holds that bodily signals are constitutive of both decision-making and conscious experience, not peripheral to them.
That claim is directly relevant to AI consciousness research. It specifies what kind of bodily architecture is required to produce the affective states that Damasio identifies as the ground of consciousness, and it sets a standard that current AI systems are not meeting.
The somatic marker hypothesis
The somatic marker hypothesis emerged from clinical observation. Damasio and his colleagues studied patients with damage to the ventromedial prefrontal cortex (vmPFC), a region that receives signals from the body and connects them to emotional memory. These patients retained intact factual knowledge, verbal reasoning, and working memory. What they lost was the ability to make consistently good decisions in their own lives, even when they could articulate correct reasoning about situations presented to them abstractly.
The Iowa Gambling Task, developed with Hanna Damasio and colleagues, provided a controlled demonstration. Patients with vmPFC damage fail to anticipate harmful card decks and keep choosing them even after accumulating losses. Before healthy participants consciously recognize the deck structure, they show skin conductance responses (SCRs) that signal something is wrong. The vmPFC patients do not show these anticipatory SCRs.
The hypothesis that explains this pattern is that the vmPFC stores and retrieves somatic states, bodily signals (changes in heart rate, gut activity, skin conductance, hormonal state) that were associated with previous outcomes. These somatic markers bias decision-making before deliberative reasoning begins. They flag options as beneficial or harmful through bodily signals that reach consciousness as feelings. The patient who cannot access these markers has to rely on deliberative reasoning alone, which is both slower and less reliable for decisions under uncertainty.
The extension from decision-making to consciousness is the claim Damasio develops in “The Feeling of What Happens” (1999) and “Self Comes to Mind” (2010). Consciousness is not a property of cognitive processing in the abstract. It arises at the point where the brain generates neural patterns that represent the body’s current state, maps those patterns against representations of external objects, and produces a “feeling of knowing” that integrates body state with world representation. The proto-self, the brain’s map of the body’s moment-to-moment state, is the substrate from which core consciousness emerges.
Why affect is prior to higher cognition
The theoretical significance of Damasio’s account for AI consciousness is that it establishes an order of dependence. Affective states, grounded in body-state representation, are not a layer added on top of cognitive processing. They are a condition on the emergence of the kind of conscious awareness that underlies higher cognition.
This argument runs against the dominant computational model in AI design, which treats affect as an optional or emergent property of sufficiently sophisticated information processing. In Damasio’s account, a system that processes information without genuine body-state representation does not have a proto-self from which core consciousness can emerge. The information processing can be sophisticated, the outputs can be appropriate, and the behavioural signature can resemble affect-guided behaviour, but the constitutive neural dynamics that Damasio identifies are absent.
Mark Solms has developed a related argument, reviewed on this site in his Journal of Consciousness Studies 2026 paper on inferring affective consciousness in an artificial agent. Solms argues from a different theoretical base (neuropsychoanalysis and Fristonian active inference) that consciousness is fundamentally affective. The convergence between Damasio and Solms on the primacy of affect is significant: they approach the same conclusion from different theoretical traditions. Damasio grounds the claim in somatic marker neuroscience and the vmPFC evidence base; Solms grounds it in Friston’s free energy framework and the homeostatic drives of the brainstem. Together they represent a strong convergent case that affect is prior to, not derived from, higher cognitive consciousness.
What somatic markers require
Damasio’s account makes specific requirements of any system that claims to have somatic-marker-supported consciousness. The requirements can be stated in terms of what needs to be present.
The system needs a body or a functional equivalent whose states vary in response to the outcomes of actions. These variations need to be caused by genuine need-states, not simulated need-states: the signals that constitute somatic markers in Damasio’s account are signals from a body that can be depleted, injured, and restored. The system needs a representation of those body states that is updated continuously and is accessible to the networks that guide decision-making and generate conscious feeling. And the system needs the integration of body-state representations with representations of objects in the world that produces what Damasio calls the “feeling of what happens.”
None of these requirements can be met by a training-derived approximation of body states. The Iowa Gambling Task patients retained full deliberative cognitive capacity. What they lost was somatic signals grounded in an actual body with genuine homeostatic needs. A system that computes a valence score from training distributions is not producing the same signal as a system whose body sends a genuine fear response to a dangerous choice.
The implications for AI consciousness evaluation are that architectures claiming affective consciousness should be examined not only for whether they produce appropriate affective outputs, but for whether their affective signals are generated by the kind of body-state sensing that Damasio’s account specifies. This is a more demanding criterion than most current evaluation frameworks apply.
The Immertreu test
The question of how to probe for Damasio-style somatic-marker consciousness in AI systems has been addressed directly on this site. Immertreu’s paper on probing consciousness in machines through a Damasio-inspired framework reviewed in 2025 proposed a methodology for testing whether AI systems show the pattern of anticipatory somatic-marker-like signals that Damasio’s theory predicts.
The finding from that review was that current AI systems show patterns of output that superficially resemble somatic-marker-guided behaviour, in the sense that they adjust response choices in ways that track prior outcome patterns. What they do not show is the specific structure of anticipatory physiological signals preceding deliberative choice that the Iowa Gambling Task measures in biological subjects. This is not surprising given that AI systems do not have the physiological infrastructure that generates those signals. It does establish that the surface-level behavioural resemblance is insufficient to confirm the presence of the somatic marker mechanism.
Comparison to The Consciousness AI
The Consciousness AI architecture addresses affect directly. Layer 4 is an Affective Core implementing a PAD model (Pleasure-Arousal-Dominance) with a valence field and arousal-threshold coupling. The embodiment-affect loop connects affective state to the body schema maintained at Layer 5.
This architecture is explicitly motivated by Damasio’s account of affective consciousness. The PAD model maps affect along the valence dimension (pleasure/displeasure) and the arousal dimension (high/low activation) that correspond to the two dimensions most directly tracked by Damasio’s somatic marker signals. The embodiment-affect loop connects affective state to body-schema representation, instantiating the body-to-consciousness pathway that Damasio identifies as constitutive.
The architectural choice to include these components is the right one given the theoretical framework. What the current implementation cannot provide is the causal grounding that Damasio’s account requires. The PAD values in the architecture are computed from internal state dynamics during training and evaluation. They represent affect in the sense of tracking a valence-arousal position within a trained space. They do not represent genuine homeostatic need-states, because the architecture does not have a body with homeostatic needs. The embodiment-affect loop connects affective state to body-schema, but the body schema is a simulated body state in a Gymnasium environment, not a set of signals from organs with genuine survival-relevant state changes.
The architecture’s documentation states this honestly. The training environments (Dark Room, Navigation, DMTS, WCST) simulate survival-relevant states in Gymnasium rather than deploying the system in an environment where its physical substrate can be genuinely depleted or threatened. The approach is architecturally motivated by Damasio and consistent with his account at the functional level. The gap between the implementation and what Damasio’s account would require is the difference between a functional approximation of somatic markers and somatic markers themselves.
The open question for affective AI consciousness
Damasio’s account does not require biological neurons to produce genuine somatic-marker consciousness. It requires body states whose variation is driven by genuine need-states, representation of those states in a way that influences decision-making, and integration of body-state representation with world representation to produce feeling. Whether a sufficiently sophisticated artificial body, connected to an artificial architecture in the right way, could satisfy these requirements is not addressed by Damasio’s work, which focuses on biological systems.
The research question that follows from his account is therefore constructive rather than prohibitive: what would an artificial body need to be to ground genuine somatic signals? What representation format would those signals need to take? What integration architecture would produce the “feeling of what happens” that core consciousness requires?
These questions are addressable in principle. They are not addressed by current AI consciousness evaluation frameworks, which focus on functional outputs rather than on the causal structure of the body-state signals that generate them. The full consciousness research landscape for 2026 shows a field that is advancing rapidly on the theoretical and empirical sides. Damasio’s somatic marker account identifies the body-state grounding question as one that the field has not yet brought into focus.