Flow states, described as highly focused yet effortless states of consciousness, are experienced across various domains. A puzzling aspect of flow is the reported loss of self-awareness despite apparent agency and skill, and the feeling of effortlessness despite task complexity. Existing flow literature offers diverse conceptualizations of this self-awareness loss. Some suggest a loss of self-awareness in general, while others highlight a loss of self as a social actor or reflective self-consciousness. This paper aims to address two key questions: (i) what type of self-awareness is attenuated during flow, and (ii) why does flow modulate self-awareness? The paper utilizes the self-model theory of subjectivity (SMT) to distinguish between various levels of self-experience, from minimal phenomenal selfhood (MPS) to high-level reflective self-representations in an epistemic self-model (ESM). The authors propose that flow primarily affects the ESM, specifically inhibiting higher-order self-conceptualization and the temporally extended self-as-object, but not pre-reflective, bodily self-awareness. This framework offers insights into flow's other characteristics and proposes a computational explanation rooted in active inference, a process theory explaining how complex entities persist in dynamic environments by minimizing variational free energy (VFE). Prior research has not thoroughly examined flow states through the lens of active inference; this computational approach addresses the qualitative nature of much of the existing flow literature and offers potential applications to understanding conditions involving dysfunctional selflessness, such as depersonalization, and well-being.

The literature review examines diverse definitions of flow and self-awareness within flow states. It highlights the inconsistencies in defining the nature of self-awareness loss during flow, ranging from a complete loss of self-awareness to a more nuanced loss of reflective self-consciousness or self-referential thinking. Existing research emphasizes the importance of a balance between challenge and skill in inducing flow. Qualitative studies provide rich phenomenological descriptions of the flow experience, including a distortion of temporal experience, intrinsic reward, and effortless control. These observations, often expressed through quotes from individuals experiencing flow in diverse activities, highlight the key aspects that the authors will attempt to explain using active inference. This body of research reveals both similarities and contradictions across flow experiences, making the need for a computational account that can unify these findings particularly salient. Previous attempts to utilize active inference in the study of flow states have been limited, creating a gap that this research seeks to fill.

The paper uses the active inference framework as its primary methodology. Active inference posits that perception and action are forms of active Bayesian inference aimed at self-organization, or minimizing variational free energy (VFE). The authors build their model by considering the dual aspects of the flow context: (i) the agent's learned expectations about the situation based on extensive practice and (ii) the challenging nature of the task. The engagement in flow elicits a contextual inference that triggers high-precision weighting over beliefs about the impact of actions on latent state transitions (B tensor in POMDP models) and expectations about sensory outcomes (C tensor). This precision weighting drives action selection towards maximizing pragmatic value, reducing the need for information gain or novelty-seeking. The EFE (expected free energy) equation is used to formally show that high precision weight drives the system towards pragmatic affordances, and reduces attentional resources required for uncertainty resolution. The challenging nature of flow-inducing activities introduces two additional constraints: (i) a requirement for focused attention on incoming sensory data, leading to high precision weighting on the likelihood mapping (A tensor), inhibiting mental time travel, and (ii) a restricted, shallow temporal horizon for action planning, further limiting mental time-travel and contributing to a reduction in reflective self-awareness. This contrasts with everyday flow states involving less complex actions, which don't necessitate the same level of total attention and thus don't inhibit self-reflection to the same degree. The paper then explores the relationship between flow states and habits, emphasizing that, while both involve action selection without deliberation, flow involves a higher level of goal-directed behavior compared to habit-driven actions. The difference between habitual mental actions (contextual cue triggers precision weight adjustment) and physical actions (precision weighting increases influence of EFE) is discussed in the context of the model's hierarchical structure. The authors present a hierarchical Bayes graph illustrating this inferential architecture, distinguishing between contextual cues for mental actions and the precision weighting that guides physical actions. This distinction is central in explaining how flow experiences can be effortless, connecting it to previous research on the phenomenology of effort and the KL divergence between context-sensitive and context-insensitive action beliefs. This framework helps to understand how flow states can be maintained or broken, explaining the dynamics of maintaining flow in the face of minor prediction errors and the loss of flow when the task is too simple or too complex. Finally, the authors look into the interaction between flow and learning. While learning may implicitly occur through sensorimotor adjustments in flow, explicit learning usually happens after the flow state concludes. This analysis extends to the role of expertise in flow, highlighting that the process of acquiring the skill necessary for flow relies on explicit exploration before the transition to implicit pragmatic action. A diachronous decline in capacities and confidence (forgetting), compensated by the positive surprise of successful performance, is linked to the autotelic nature of flow and its connection to positive valence. The authors also discuss how different accounts of valence align with their active inference model, concluding that the positive valence in flow stems from unexpectedly efficient free energy minimization.

The paper's key findings revolve around the active inference model of flow states. The model posits that flow is characterized by high-precision weighting over second-order beliefs about the consequences of actions and the resulting sensory observations. This leads to the selection of actions that maximize pragmatic value, thereby minimizing expected free energy. The challenging nature of flow-inducing activities restricts the temporal depth of the generative model. This reduces the individual's capacity for mental time travel (planning and retrospection). This explains the reported attenuation of reflective self-awareness, specifically the temporally extended self-as-object and the conceptually represented self-as-object, within the framework of the self-model theory of subjectivity (SMT). The model highlights that pre-reflective, bodily self-awareness is retained during flow; this is supported by observations from various qualitative studies. The paper clarifies that flow states are distinct from habits; while both involve action without deliberation, flow is more goal-directed and involves state-based inference. The model also elucidates why flow is effortless by suggesting that the effort in performing flow-inducing actions stems from the divergence between context-sensitive and context-insensitive action beliefs. This difference is small in flow, resulting in the experience of effortlessness. Furthermore, the model explains how flow states can be maintained or broken. Minor prediction errors are seamlessly integrated through skillful motor behavior, while larger errors or overly simple tasks disrupt the state. The authors find that while implicit learning might occur during flow, explicit learning and insights typically emerge after the flow state has ceased. The model proposes that the autotelic nature of flow results from the positive surprise experienced when the individual's performance exceeds their prior, pessimistic expectations, which also explains the positive valence associated with the experience. These insights are connected to existing theories of valence which emphasize the role of successful error minimization.

The active inference model proposed in this paper provides a comprehensive explanation for the seemingly paradoxical aspects of flow states. The model successfully integrates various phenomenological observations from qualitative studies, formalizing them within a computationally grounded framework. By emphasizing the role of precision weighting, the model sheds light on how the balance between challenge and skill contributes to the experience of both effortless control and reduced self-awareness. The proposed model is not merely descriptive; it offers testable predictions regarding the neural correlates of flow and how different levels of precision weighting might influence the experience of self. The contrast drawn between everyday flow states and canonical flow states clarifies the role of attentional resources in mediating the experience of self. By addressing the debate regarding the role of learning in flow, this work contributes to a more nuanced understanding of the cognitive processes involved and opens avenues for future research. The discussion of the autotelic nature of flow, linking it to positive surprise and successful free energy minimization, expands the understanding of intrinsic reward and motivational aspects of flow. The computational framework suggests potential applications beyond flow, offering new tools for investigating experiences of selflessness, embodied expertise, and well-being, even extending to understanding dysfunctional states like depersonalization.

This paper offers a novel active inference-based model of flow states. It explains the seemingly contradictory features of flow—effortless control alongside reduced self-awareness—through precision weight dynamics and a shallow temporal planning horizon. The model clarifies that flow does not eliminate self-awareness entirely but rather alters its phenomenal form, maintaining pre-reflective bodily self-awareness while inhibiting reflective aspects associated with the epistemic self-model. Future research could investigate the specific neural correlates of precision weighting and temporal depth in flow, exploring how individual differences in skills and preferences impact the flow experience. Further qualitative studies using micro-phenomenological interviews could deepen our understanding of the subtle phenomenological aspects of the “transparent” and “performative” body during flow. Exploring whether purely mental flow exists and its potential phenomenology would also be a valuable extension of this work.

While the model provides a comprehensive account of flow, it relies heavily on the active inference framework. The applicability of this framework to all aspects of flow experiences remains to be fully tested empirically. The model primarily focuses on canonical flow states; the extent to which it generalizes to less intense or more contextually specific flow experiences requires further investigation. The integration of qualitative and quantitative data relies on a degree of interpretation, and future studies might benefit from more rigorous methodological approaches that bridge these two approaches. The model emphasizes the computational aspects of flow, leaving room for future work to explore the neurobiological underpinnings and the exact neural mechanisms responsible for the precision weighting and temporal horizon adjustments.