Consciousness in active inference: Deep self-models, other minds, and the challenge of psychedelic-induced ego-dissolution

Predictive processing and its process-theoretic formulation, active inference, have become influential frameworks for explaining many aspects of perception, action and cognition, and they hold promise for illuminating the computational underpinnings of phenomenological states. Critics, however, argue that predictive processing as usually formulated is insufficiently constrained to yield a theory of consciousness, in particular because it lacks a clear computational account that would allow us to determine which non-human biological or artificial systems are conscious (the 'other systems' argument). Another challenge comes from reports of selfless experiences—most notably psychedelic-induced ego-dissolution—which appear to show vivid conscious states in the absence of ordinary self-consciousness and therefore threaten theories that treat self-modelling as constitutive of consciousness. Deane sets out to show that active inference has the resources to deliver a theory of consciousness by developing a self‑modelling account of subjectivity. The central proposal is that phenomenal consciousness is underpinned by a hierarchically deep inference about precision on the action model—what the author calls 'subjective valuation'—that evaluates how well a system is realising self‑evidencing (allostatic) outcomes across multiple timescales. The paper develops this account by explicating the mechanics of active inference, mapping components of conscious contents to the generative model, considering disruptions of ordinary selfhood (depersonalisation, meditation), examining implications for attributing consciousness to other systems, and responding to the 'selflessness challenge' via an analysis of psychedelic ego-dissolution using models such as REBUS (Relaxed Beliefs Under pSychedelics).

This is a theoretical and conceptual paper rather than an empirical study. Deane conducts a structured conceptual synthesis: first reviewing and explicating the formal mechanics of the active inference framework (generative models, variational Bayes, expected free energy, precision-weighting, allostatic control), then developing a self‑modelling theory of subjectivity built on those mechanics. The author sequentially maps elements of conscious content to elements of the hierarchical generative model (perception, precision, active inference, affective inference) to show how a deep control model could underpin subjectivity. To support and illustrate the account, the paper draws on a range of empirical and phenomenological sources already present in the literature: examples from sensory cue integration and motor control, clinical phenomena (depersonalisation, schizophrenia), comparative neuroethology (corollary discharge in simple and complex animals), meditation studies, and the growing computational literature on psychedelics (notably the REBUS model). Phenomenological reports (including anecdotal reports of psychedelic experiences) are used as illustrative material. The extracted text does not report any systematic search strategy, new data collection, experimental methods, or quantitative meta-analytic procedures; the approach is argumentative and synthetic rather than empirical.

The main theoretical result is a coherent account that identifies subjective valuation—an inference about the precision of the action model and the expected free energy of policies across hierarchical levels—as the constitutive computational core of phenomenal consciousness. Deane argues that consciousness arises when an organism evaluates how well its generative model will realise phenotype-congruent, self‑evidencing outcomes over varying timescales, and that this hierarchically deep inference shapes perception, agency and affect. Several explanatory components are unpacked. Perception is cast as posterior state estimation (model inversion) under a generative model. Precision is the inferred reliability (inverse variance) of signals and priors and functions like attentional gain; precision‑weighting determines how top‑down priors and bottom‑up prediction errors combine. Active inference casts action selection as planning-as-inference, where policies are chosen to minimise expected free energy (a quantity combining pragmatic value—fulfilling prior preferences—and epistemic value—resolving uncertainty). Agentive control—the inference that one can endogenously control sensory outcomes—is tied to sensory attenuation mechanisms. Two kinds are distinguished: physiological sensory attenuation (used in movement initiation by transiently down-weighting proprioceptive evidence) and perceptual sensory attenuation (top‑down filtering of afferent inputs that supports a sense of agency and a self/other distinction). Affective inference is framed as inference about precision on prior preferences and epistemic affordances; valence tracks whether the current policy is achieving the expected rate of prediction error reduction and thus serves as a domain‑general controller that tunes policy precision. Deane uses this formalism to characterise clinical and phenomenological states: depersonalisation and some forms of depression are modelled as a domain‑general inferred loss of allostatic control (global loss of precision on policies), whereas focused attention meditation can be understood as training endogenous precision control, producing a perceived gain in control. Turning to other systems, the author sketches how mechanistic markers associated with the allostatic control model—particularly complex, higher‑order corollary discharge and sophisticated sensory attenuation supporting counterfactually rich, temporally deep models—can inform attributions of consciousness across species. Simple nervous systems (e.g. C. elegans) that exhibit only lower‑order corollary discharge and lack anticipatory, goal‑directed behaviour are argued to be unlikely candidates for phenomenal consciousness on this account; animals that show predictive, temporally extended sensorimotor control (bats, songbirds, primates) are more promising. On the selflessness challenge, Deane integrates the REBUS model to explain psychedelic‑induced ego‑dissolution. He explains that psychedelics relax high‑level priors, increasing the precision or impact of ascending prediction errors so that corollary discharge and sensory attenuation fail; consequently the system ceases to posit itself reliably as an endogenous controller, producing loss of ordinary self‑boundaries. Crucially, affective/inferential valuation remains: the psychedelic state often entails intense epistemic value and vivid affect, and Deane distinguishes 'complete' ego‑dissolution (often blissful, linked to relaxation of constraining priors that remove sources of suffering) from 'incomplete' ego‑dissolution (characterised by fear and distress). From this perspective, ego‑dissolution does not falsify the notion that subjectivity depends on self‑modelling because subjective valuation—the inference about how well one is self‑evidencing—remains present even when sensorimotor structures of selfhood collapse. The paper therefore identifies affective valence or evaluative inference as the deepest constitutive facet of experience.

Deane interprets the proposed account as showing that active inference can indeed ground a theory of consciousness: hierarchical self‑models that perform subjective valuation provide the computational mechanisms that make sensory states 'matter' for a subject. He positions this view relative to earlier work by emphasising that the account goes beyond explaining contents of consciousness to identify a constitutive role for hierarchically deep inference about policy precision and allostatic control. The author argues that this move helps answer the 'other systems' objection by pointing to neurocomputational markers—complex corollary discharge and context‑sensitive sensory attenuation—that can be sought empirically when assessing consciousness in non‑human systems. In addressing the selflessness challenge, the paper contends that psychedelic‑induced ego‑dissolution is better understood as a disruption of sensory attenuation and high‑level priors than as proof that consciousness can exist in the total absence of self‑modelling. Deane emphasises that affective valuation survives in ego‑dissolution and that this evaluative inference is the most basic phenomenal ingredient: even when sensorimotor or cognitive self‑representations fail, felt valence persists. The author acknowledges terminological and conceptual complexity around definitions of 'self‑consciousness' and notes that disputes about the necessity claim versus the typicality claim may hinge on definitional choices; on a stringent definition of self‑consciousness as representation of oneself 'as oneself', ego‑dissolution might count as truly selfless, but this possibility does not undermine the paper's claim that subjective valuation underwrites conscious experience. As for implications, Deane suggests that the active inference framework, combined with empirical work on psychedelics and related phenomena, can inform consciousness science, computational psychiatry and computational phenomenology by clarifying mechanisms by which experience is shaped and by identifying physiologically testable markers of deep self‑models. The extracted text does not supply a dedicated limitations subsection with empirical caveats, but the author does note conceptual uncertainties (e.g. polysemy of key terms) and concedes that sensorimotor and evaluative aspects of consciousness could vary independently in different organisms or states.

Deane concludes that phenomenal consciousness is plausibly understood within predictive processing as arising from hierarchically deep self‑models: a 'deep control model' that tracks endogenous control of self‑evidencing outcomes across timescales. The paper argues that this perspective addresses two major objections—insufficient specification for attributing consciousness to other systems and the challenge posed by psychedelic ego‑dissolution—by identifying mechanistic markers of deep self‑modelling and showing that ego‑dissolution preserves the core evaluative inference constitutive of subjective experience. The author therefore presents active inference as a promising and testable framework for future consciousness science.