Forgetting ourselves in flow: an active inference account of flow states and how we experience ourselves within them

The paper addresses two questions about flow: (i) which kind of self-awareness diminishes during flow and (ii) why flow modulates self-awareness in that way. The authors situate flow as an "almost automatic, effortless, yet highly focused" state that paradoxically involves reports of diminished self-awareness alongside clear skill and agency. Drawing on Metzinger’s self-model theory and related phenomenology, they distinguish between pre-reflective self-as-subject (minimal phenomenal selfhood) and reflective self-as-object (including a temporally extended self in planning and a conceptually represented, historicized self). They motivate the need for a computational account that explains how flow can inhibit reflective self-awareness while preserving pre-reflective, bodily subjectivity. They propose active inference as the formal framework to resolve these puzzles.

The authors survey classic and contemporary flow literature (Csikszentmihalyi and successors), highlighting varied descriptions of self-loss (loss of reflective self-consciousness, low self-referential thinking, action–awareness merger). They review phenomenological work on selfhood (Merleau-Ponty, Sartre, Zahavi) and Metzinger’s self-model theory distinguishing minimal phenomenal selfhood (MPS) and epistemic self models/agent models (ESM/EAM). They discuss predictive processing/active inference accounts of self, agency, and sensorimotor integration; the role of interoception and DMN in self-referential processing; and related altered states (meditation, psychedelics) that modulate temporal depth and self-awareness. Prior distinctions between habits vs goal-directed control, boredom vs exploration, and choking under pressure are integrated to frame flow’s phenomenology and boundary conditions.

This is a theoretical/computational analysis using the active inference framework rather than an empirical study. The authors formalize flow as precision-weighting and planning dynamics within a hierarchical generative model (discrete POMDP and continuous predictive coding notions). Key model components: - Generative model parameters: A (likelihood mapping from states to observations), B (state transition beliefs), C (prior preferences over outcomes), D (prior over initial states), E (priors over policies/habit strength). Second-order precision variables modulate these (e.g., likelihood precision ζ, transition precision ω, preference precision γ). - Contextual cueing: Re-entering a well-practiced, challenging context triggers habitual mental policies (E) that deploy high precision over (i) beliefs about action-dependent state transitions (B) and (ii) prior preferences over expected sensory outcomes (C). This biases action selection toward exploiting pragmatic affordances. - Attention as precision optimization: The challenging, volatile environment requires wholesale attentional focus, realized as high precision on the likelihood mapping (A), prioritizing present, task-relevant sensory evidence and suppressing mental time-travel (planning and retrospection). - Shallow planning horizon: Despite expertise, environmental volatility enforces temporally shallow policy selection. The system infers it will know what to do at the next moment, eschewing deep counterfactual planning. - Expected free energy (EFE) formalism (Equation 1): Policy evaluation combines epistemic value (information gain) and pragmatic value (preference satisfaction). In flow, precise B and C and low ambiguity reduce incentives for epistemic foraging; exploitation dominates. - Effort formalization (Equation 2; Q(u)=σ(E−G)): Effort relates to the KL divergence between habitual priors over policies (E) and context-sensitive policies minimizing EFE (G). In flow, alignment between E and G is high for the mental action of precision deployment, yielding the phenomenology of effortlessness. - Hierarchical Bayes graph (Figure 1): Contextual inference at a higher level initializes mental state and policy selection, which parameterizes precision at the perceptual/action level, producing moment-to-moment action sequences driven by successive contextual cues (n→n+1→n+2). This formal account is used to derive mechanistic explanations for flow’s features (loss of reflective self-as-object, effortlessness, autotelic affect, distinctions from habits, boundary conditions such as boredom or choking).

- Flow’s core mechanism: high precision weighting over (i) expected sensory consequences of action (B) and (ii) prior preferences over outcomes (C), combined with high likelihood precision (A) due to task-demanded attention. This yields exploitation of pragmatic affordances and shallow, present-focused control. - Self-awareness modulation: Flow inhibits reflective self-as-object (both the temporally-extended self from deep planning and the conceptually represented, historicized self from meta-cognition), while preserving and foregrounding pre-reflective, bodily self-as-subject that is both transparent (world given in a bodily mode) and performative (felt embodied agency and know-how). - Effortlessness: The mental policy of precision deployment in flow closely matches habitual priors (E), minimizing the KL(E‖G) term and thus subjective effort. - Autotelic positive valence: Flow often exceeds pessimistic expectations due to forgetting hyper-priors between sessions, producing precision (gamma) increases and positive affect; also consistent with accounts tying valence to faster-than-expected error reduction (error dynamics). - Distinction from habits: Habits are state–action mappings that bypass state-value inference and counterfactual depth; flow retains (shallow) state-based active inference and goal-directed control while being contextually cued by habitual mental actions (precision deployment). - Boundary conditions: Flow is maintained despite small, rapidly resolved prediction errors; breaks when large uncertainty arises (e.g., salient mistake) leading to planning/self-objectification or when tasks are too easy (boredom) or too hard (choking). Boredom shifts behavior to epistemic exploration; choking reflects precision capture by higher-level narrative preferences. - Neural correlates: Proposed alignment with reduced DMN activity (implicated in counterfactual planning and self-referential thought) during flow, meditation, psychedelics, consistent with contracted temporal depth and diminished reflective self-awareness. - Learning: Explicit, propositional learning likely occurs post-flow; implicit sensorimotor learning may occur within flow. Expertise is prerequisite; training builds the contextual cue → precision deployment linkage and action–outcome associations.

The framework resolves the flow paradox by showing how precise, context-triggered beliefs and shallow planning suppress reflective self-modelling while preserving and enhancing pre-reflective embodied subjectivity and skilled control. The model links multiple flow features—effortlessness, autotelic affect, temporal distortion, action–awareness merger—to a single computational substrate (precision allocation and EFE minimization). It differentiates flow from habits, explains why boredom and choking disrupt flow, and predicts variability across domains, including altered states that also contract temporal depth. This advances theoretical integration between phenomenology of expertise, predictive processing/active inference, and self-model theories, offering testable predictions (e.g., precision/attention dynamics, DMN modulation, micro-phenomenological signatures).

The authors propose that flow arises from high precision weighting over second-order beliefs about expected consequences of action and preferred outcomes, together with high likelihood precision and a contracted planning horizon. This drives exploitation of pragmatic affordances, inhibits reflective self-as-object (temporal and conceptual) while maintaining pre-reflective bodily self-as-subject, and yields effortlessness and positive valence. The account unifies diverse flow phenomena within active inference and suggests empirical avenues (e.g., neuroimaging of precision/DMN, micro-phenomenology) and extensions to other selfless experiences. The paper highlights implications for expertise and wellbeing and calls for investigating purely mental flow and the learning dynamics within and around flow.

- The work is theoretical; no empirical data are collected. Proposed mechanisms (precision deployment, shallow planning horizon, DMN modulation, effort as KL(E‖G)) require empirical testing. - Flow learning dynamics remain open: the extent of implicit learning within flow vs post-flow explicit consolidation is not resolved. - Conceptual mapping between MPS, ESM/EAM, and pre-reflective vs reflective self-awareness involves theoretical tensions; alternative phenomenological accounts (e.g., interoceptive vs sensorimotor emphasis) are acknowledged but not adjudicated. - Neural claims (DMN attenuation, network dynamics) are tentative and based on existing correlational findings. - Scope focused on bodily flow; it remains to be explored whether purely mental flow exhibits the same self-awareness profile and precision dynamics. - Model hybridizes discrete (POMDP) and continuous predictive coding descriptions; some dependencies are simplified, and temporal depth is only partially represented in Figure 1.