From Insentient Allostasis to Adaptive Bodily Selfhood: Conscious vs Unconscious Instrumental Interoceptive Inference

1. Introduction

The pre-reflective experience of being a body is a fundamental component of consciousness and selfhood (Blanke & Metzinger, 2009; Damasio, 2021; Díaz, 2018; Gallagher & Zahavi, 2021; James, 1890; Legrand, 2007; Merleau-Ponty, 2005). A very influential proposal to account for this experience is provided by Seth and Tsakiris (2018). From the perspective of active inference, they propose that this foundational form of bodily self-consciousness arises from the neurocomputational mechanism of instrumental interoceptive inference (hereafter “III”). That is, from the content of the nervous system’s interoceptive predictions about the sensory consequences of “intero-actions” (autonomic adjustments) aimed at maintaining physiological integrity. Because these predictions are control-oriented—not about discovering inner states, but about how well the inner milieu is regulated—III would explain both the non-object-like and subjective stability of selfhood (Seth & Tsakiris, 2018). At the bottom, it would also account for the basal sense of being alive, underpinning all affective experiences, that is, felt moods, emotions and motivations (Damasio, 1994, 1999; Fuchs, 2013) and presumably all consciousness (Carvalho & Damasio, 2021; Cea & Martínez-Pernía, 2023; Craig, 2015; Seth, 2021, 2024). Additionally, this mechanism would illuminate the continuity between life and experience (Palacios-García & Parada, 2023; Seth, 2021, 2024; Thompson, 2007), help explain why conscious AI might be impossible (Seth, 2025), and advance our understanding of disruptions of self-experience in various psychopathologies (Barrett et al., 2016; Paulus et al., 2019; Quattrocki & Friston, 2014; Seth, 2024).

However, while this approach represents a major step forward, it has an important limitation. It does not distinguish between the predictive interoceptive control that the autonomic nervous system continuously exerts to keep us alive—even during unconscious states such as dreamless sleep or under general anesthesia—and the interoceptive predictions that ground the conscious feeling of being a body while awake (and presumably also while dreaming), that arguably allows a distinctly conscious form of adaptive behavior not present during behaviorally rigid unconscious states. Even in wakefulness, there remains a distinction between the likely limited slice of predictive control that constitutes our background, diffuse feeling of being a living body, and the one that operates without any associated phenomenology.

Hence, if III is indeed the main neurocomputational internal mechanism safeguarding our physiological integrity by actively minimizing interoceptive (precision-weighted) prediction errors, then it seems reasonable to suggest that there should be at least two types: conscious and unconscious instrumental interoceptive inference (hereafter “C-III” and “U-III”). But Seth & Tsakiris, 2018 proposal is insufficient to distinguish between these, limiting its explanatory power concerning the specifically phenomenal (i.e., conscious, felt) aspect of being a body that is constantly predicting itself to regulate itself.

In the remainder of this opinion piece, I will briefly discuss various alternatives to overcome this theoretical limitation of III as an account of the phenomenology of being a body and associated adaptive behavioral capabilities. I will begin by presenting two ways in which the distinction between C-III and U-III might be rejected, indicating their important theoretical costs.

2. Rejecting the Distinction: From Unconscious Reactivity to Conscious Dreamless Sleep

First, it might be suggested that while unconscious (e.g., dreamless NREM sleep), our physiology works entirely through a reactive, non-predictive, first-order form of autonomic control, without III. If this is the case, then III only applies to allostatic regulation accompanied with consciousness, thus avoiding my criticism.

The downside of this suggestion is that it creates an unparsimonious (even ad hoc) divide between two types of autonomic control: one governed by the principles of active inference and the other not. This is contrary to the theoretical value of explaining bodily experience, affect and physiological regulation by common principles (Seth, 2021, 2024; Seth & Tsakiris, 2018). It is also in tension with the hypothesis that all allostatic regulation engages autonomic reflex arcs to exert changes that fulfill descending interoceptive predictions (Gu et al., 2013), which extends previous proposals that motor action consists in self-fulfilling proprioceptive predictions “enslaving” classical reflex arcs (Adams et al., 2013). This provides a very elegant, parsimonious account integrating both action-perception and autonomic-interoceptive cycles. In contrast, positing that unconscious regulation relies on non-predictive, merely reactive reflexes not based on the principles of active inference seems a theoretical cost that would be preferable to avoid.

One might reply that this alternative’s unparsimonious divide is no worse than that between C-III and U-III, and thus theoretical disunity in explaining (neuro)physiological regulation is acceptable. Yet this reply fails. The C-III/U-III distinction can be drawn within the scope of active inference, by appealing to differentiating neurocomputational properties (see below), as customary in explaining different phenomena of interest using this framework. The reactive alternative, by contrast, demands a sharper split, carving out a domain where active inference no longer applies, thereby undermining it as an all-encompassing framework to explain both life and mind (Kirchhoff et al., 2018; Ramstead et al., 2020).

On the other extreme, it might be claimed that the distinction between C-III & U-III is not needed because the predictive physiological regulation that occurs during purportedly unconscious states such as dreamless sleep, might indeed be associated with consciousness. In other words, this alternative proposes that there is only one type of III because there is really no unconscious bodily regulation. The relevant states can be understood as episodes of “pure awareness” (Metzinger, 2024; Srinivasan, 2020) or “contentless sleep experiences” (Alcaraz-Sánchez et al., 2022; Windt et al., 2016), possibly equivalent to the “basal, formless, and ever-present sense of simply ‘being’ an embodied, living organism” that Seth (2021, p. 182) considers to be the foundational bedrock of experience, and that he claims is explainable by III. Thus, even in dreamless sleep there might be an objectless, contentless pre-reflective experience of pure awareness, of just being alive (Barile, 2023; Cea & Martínez-Pernía, 2023; Craig, 2015), that III may account for.

The problem with this option is that it contradicts the prevailing consensus in neuroscience and physiology that humans undergo well-documented states of unconsciousness. These include dreamless sleep—especially (but not always) in NREM stages (Siclari et al., 2018)—along with pharmacologically induced general anesthesia (Mashour, 2024) and clinical disorders of consciousness such as coma and the unresponsive wakefulness syndrome (UWS) (Laureys et al., 2010). Interpreting these conditions as episodes of pure awareness or just feeling alive rather than genuine absence of experience would represent a substantial departure from established scientific understanding and may weaken the explanatory credibility of interoceptive inference.

That said, it is worth noting that their operational classification as unconscious is primarily based on non-reportability and lack of introspective access, which do not strictly entail the complete absence of phenomenal qualities. Accordingly, the prevailing scientific consensus may partly reflect current empirical, methodological, and even theoretical limitations, and could be revised as new evidence, methods, and conceptual frameworks emerge.

In sum, instrumental interoceptive inference (III) can be preserved as the main neurocomputational mechanism that explains the basal feeling of being a body—without the need to distinguish between conscious and unconscious types—but at the price of either creating an unparsimonious theoretical divide between predictive and non-predictive autonomic control; or by positing consciousness in states that neuroscience and physiology largely consider unconscious.

3. Differentiating C-III and U-III Neurocomputationally

In this section, I will offer an alternative that avoids the previous options and instead grants the distinction between conscious and unconscious forms of instrumental interoceptive inference (i.e., C-III & U-III). Based on recent literature on affect, consciousness and active inference, I propose computational properties that C-III may additionally possess compared to U-III.

A first key element is that C-III, in contrast to U-III, additionally involves in-context, real-time modulation of the precision weighting of interoceptive prediction errors (Solms, 2019, 2021; Solms & Friston, 2018). According to Solms, the fundamental layer of consciousness corresponds to felt arousal, understood as the most basic affective feeling of being conscious. Neurocomputationally, this would arise from the nervous system’s (mainly the ERTAS) active modulation of postsynaptic gain—that is, the adjustment of the precision weighting of interoceptive prediction errors—in real time. Complementarily, Solms proposes that the hedonic valence of affect derives from the direction of this modulation: when precision weighting increases interoceptive prediction error, it feels bad; when it reduces error, the affective tone is positive. In this way, the process of precision modulation would account for both the presence of basic affective arousal and its inherent valence.

Thus, by being integrated into this regulatory loop of precision optimization, the content of interoceptive predictions may become phenomenologically imbued with felt arousal and hedonic valence, thereby qualifying as experiential. This account is consistent with a wide body of psychological evidence indicating that the dimensions of arousal and valence are fundamental building blocks of affective-somatic phenomenology (Russell, 2003; Russell & Barrett, 1999; Yik et al., 2011). Moreover, recent work emphasizes that although core affective feeling comprising blends of arousal and valence are often reportable, they likely constitute a pre-reflective background of experience—an implicit bodily form of self-consciousness (Legrand, 2007)—that anchors conscious experience in felt vitality (Cea, 2023; Cea & Martínez-Pernía, 2023; Fuchs, 2012). By contrast, U-III would involve predictive interoceptive control without online precision optimization, enabling effective physiological regulation without phenomenology.

However, this is likely not the whole story. It has recently been argued that conscious—in contrast to unconscious—active inference involves posteriors that are disposed to guide counterfactual policy selection at the appropriate timescale and with sufficient precision (Whyte et al., 2024). This aligns with Friston’s suggestion that consciousness arises when the self-evidencing generative model is temporally and counterfactually deep (Friston, 2018). Such a view resonates with Husserl’s influential claim that all consciousness involves not only the experience of the immediate present but also of the just-elapsed and the about-to-occur (i.e., “retention” and “protention,” respectively) (Husserl, 2012), which Zahavi (2008) interprets as the structural basis of the minimal sense of selfhood or “mineness” intrinsic to experience. Gallagher (2017), following Merleau-Ponty (2005), has further emphasized that, from an enactive perspective, consciousness should be understood as an embodied “I can,” meaning a basic awareness of available possibilities for action, which presupposes a sensitivity to potential future interactions with the environment. In line with these considerations, Kiverstein (2020) argues that the temporal thickness of inferential self-modeling is fundamental for the emergence of the basic experiential selfhood that plausibly characterizes all conscious life, demarcating it from insentient biology.

These considerations are related to what has recently been identified as the intrinsically motivational aspect of the fundamental form of bodily self-consciousness explained by III (Cea & Martínez-Pernía, 2023). This intrinsic motivation refers to the basic driving of behavior that core affective feelings (i.e., arousal & valence) inherently exert in virtue of how they feel, rather than by depending on external rewards or norms. Phenomenologically, they involve a pre-reflective, arguably instinctual sense of willingness either to act (expend energy) or to remain inactive (conserve energy)—depending on arousal level—for the organism to either perpetuate or end a given feeling and the associated situation (depending on hedonic valence). This is based on the hedonic principle that subjects act to prolong or repeat pleasant states and to terminate or avoid unpleasant ones (Damasio & Damasio, 2024; Russell, 2003, 2005), while the arousal dimension corresponds to the felt potential for mobilization and work, which impels movement when heightened but favors inactivity when depleted (Liu et al., 2025; Thayer, 1990, 1996).

Thus, if we integrate this view with Solms’ proposal that hedonic valence signals variations in free energy/interoceptive prediction error, (i.e., meaning physiological changes either towards or away from homeostatic set points), the feeling of being a living body would inherently possess a future projection that may be characterized in counterfactual terms, say, as “if I conserve energy (don’t work) my body’s future expected free energy/interoceptive prediction error will diminish” (e.g., wanting to rest when feeling tired), or “if I expend energy (work) my body’s future expected free energy/interoceptive prediction error will diminish” (e.g., actively seeking water when feeling anxiously aroused due to thirst). In other words, a plausible hypothesis is that in virtue of how one feels as a bodily entity, one is pre-reflectively disposed to either expend energy (i.e., move, work) or conserve energy (i.e., don’t move, don't work) in order to either correct or preserve in the future a present homeostatic physiological trajectory that is either moving towards or away from vital homeostatic set points (Cea & Martínez-Pernía, 2023). Thus, in the case of having less energetic resources than needed for an ongoing or anticipated task, feeling tired or fatigued might play a role in the adaptive behavioral regulation of the body’s energy budget, precisely by its inherently motivational aspect that impels one to stop working and rest (Liu et al., 2025).

In sum, this brief review of active inference literature concerning consciousness and the future-oriented character of basic bodily self-consciousness suggests that conscious instrumental interoceptive inference (C-III)—in contrast to unconscious III (U-III)—might be characterized as III that additionally involves in-context, real-time modulation of the precision weighting of interoceptive prediction errors, coupled with temporally and counterfactually deep interoceptive self-models that might also underpin a functional contribution of minimal, bodily forms of self-consciousness in adaptive behavior in both human (Barrett, 2020; Cea, 2023) and non-human animals (Veit, 2024).

4. Having a Model vs Being a Model

This, in turn, naturally leads to the hypothesis that C-III, but not U-III, may require explicitly encoded interoceptive generative models, instead of being implemented by bodies or subsystems that simply are models of their inner hidden environments (i.e., exert predictive control without explicitly encoded modeling, see Seth and Tsakiris (2018) for this distinction). Something related is indeed discussed by Seth and Tsakiris (2018), but their focus is different. They propose that the increasing complexity of selfhood—from the simple feeling of being a bodily organism to socially mediated self-awareness—might be due to increasing model explicitness and hierarchical depth. However, their proposal leaves unanswered the question on the dividing line between basal bodily selfhood and unconscious allostatic regulation.

A natural suggestion would then be that the functional attenuation of neural activity involved in explicit interoceptive modeling leads to cessation of bodily self-consciousness. Hence, in dreamless sleep and general anesthesia, physiological allostatic regulation driven by U-III might be limited to the activity of the autonomic nervous subsystem that only is a model, instead of having (i.e., explicitly encoding) a model of the inner body. Neurobiologically, this would likely entail significant reduction of activity in the anterior insular cortex during unconscious periods of NREM sleep and general anesthesia (consistent with Betta et al., 2021; Warnaby et al., 2016). However, the hypothesis that the most basic forms of bodily selfhood mainly rely not on the cortex, but on self-models encoded in brainstem structures like the ERTAS (Solms, 2013, 2021) or the PAG (Panksepp, 1998), should also be considered. Importantly, this remains a matter of ongoing debate, with authors like Safron (2021) arguing that subcortical affective processes may not be sufficient, on their own, to generate conscious experience without cortical re-representation. The present proposal does not presuppose a resolution of this debate.

At the same time, we remain open to the possibility that speaking in terms of generative “models” might just be an epistemic instrument for explanation, prediction, simulation, and control, and not ontologically real representations encoded in neural dynamics (van Es, 2021). If this is the case, then the aforementioned difference between “having” and “being” a model may collapse into a mere difference of degree in the complexity of the interoceptive expectations enacted by visceromotor actions (Ramstead et al., 2020).

5. Concluding Remarks

In conclusion, there are good reasons to differentiate between the predictive, interoceptive-autonomic control sustaining life in unconscious states, and the allostatic regulation underpinning adaptive bodily self-consciousness in wakefulness. We proposed specific neurocomputational properties to distinguish between these, as conscious and unconscious forms of instrumental interoceptive inference. With this proposal, we hope to encourage more research on the neurocomputational differences between unconsciousness, for example, in dreamless sleep, or general anesthesia, and minimal forms of bodily self-consciousness, for example, in thought-less and imagery-free states induced by Reduced Environmental Stimulation Therapy (Feinstein et al., 2018), “deconstructed” consciousness induced by 5-MeO-DMT (Timmermann et al., 2025), or “empty” meditative states (Metzinger, 2024). In this way, we may better elucidate the (maybe diffuse) border between sentient, adaptive physiological regulation, and blind organismic life.