by Marlon Barrios Solano
Abstract. Active inference – Karl Friston’s framework for perception and action – posits that self-organizing agents act to minimize their expected free energy. A key component of this process is the epistemic or informational value of actions, often characterized in terms of “epistemic affordances”: opportunities in the environment that an agent can exploit to reduce uncertainty. Here we explore Friston’s notion of epistemic affordance in depth, tracing its theoretical roots (from Gibsonian affordances to Bayesian active inference) and illustrating how agents actively seek out uncertainty-reducing information. We then examine how this perspective bears on central questions of sentience and consciousness. We argue that sensitivity to epistemic affordances unifies biological drives for curiosity, exploration, and self-modelling, and suggest that aspects of subjective experience (e.g. attention, “feeling aware,” flow states) naturally arise within an active inference account. We also address philosophical critiques (concerning representationalism, the “hard problem,” and testability) and clarify what (and what not) the concept of epistemic affordance contributes to debates on mind and consciousness. We conclude that while Friston’s epistemic affordance concept provides a powerful formalism for studying how agents interact with their world, careful articulation is needed to avoid overextension; nonetheless, it has rich implications for cognitive science and theories of subjective experience.
Active inference is a leading theoretical framework in cognitive neuroscience, proposing that living systems behave as if they minimize a quantity called variational free energy (Friston). Roughly speaking, agents carry generative models of their environments and constantly select actions that make the world more predictable with respect to those models. An important insight of active inference is that agents do not merely pursue externally rewarded goals, but also seek to reduce uncertainty about their world – a drive sometimes characterized as intrinsic or epistemic motivation. In practice, this means agents will choose some actions for exploration (to gain information) as well as for exploitation (to secure preferred outcomes). The notion of epistemic affordances arises in this context: roughly, those aspects of the environment or potential interactions that promise information gain. In the ecological psychology tradition, affordances (Gibson) are opportunities for action offered by the world; here, epistemic affordances are opportunities for learning or reducing uncertainty.
Friston and colleagues have explicitly connected epistemic affordances to the formalism of active inference. Mathematically, the expected free energy of a policy (plan of actions) decomposes into an extrinsic or goal-directed term and an epistemic term (sometimes called salience or intrinsic value). The epistemic term is essentially the expected information gain (mutual information) about hidden states or model parameters that would result from following that policy. In other words, the agent selects actions that it expects will reveal the true state of the world, thereby reducing surprise in the future. Concretely, Friston et al. explain that the intrinsic value of a policy is its epistemic value or affordance, “the expected information gain afforded by a particular policy” about hidden states or parameters. This formalization shows that an epistemic affordance – say, an eye movement that would clarify an ambiguous visual cue – is precisely the option that maximizes expected information gain.
This essay elaborates Friston’s concept of epistemic affordance from multiple angles. First we review its theoretical foundations: how active inference, Bayesian decision theory, and the embodied cognition literature converge on the idea that cognition involves active exploration of an uncertain world. We clarify how epistemic affordances generalize Gibson’s classical affordances by focusing on information rather than mere action possibility. Next we illustrate epistemic affordances with examples and models (e.g. rodents exploring mazes) that show agents behaving “as if” seeking uncertainty-reducing information. We then extend the discussion to sentience and consciousness. We argue that sensitivity to epistemic affordances underwrites key aspects of agency, self-modeling, and attentional awareness. For instance, we consider how the hierarchical generative model of a conscious agent might encode longer-term epistemic affordances (as noted by Bruineberg & Rietveld) and how this relates to phenomenology of attention and flow states. Finally, we address criticisms: some worry that the affordance concept is stretched too far, or that active inference cannot by itself explain qualitative experience. We discuss these concerns in the context of cognitive science and philosophy of mind, noting both the strengths and limits of the active inference view.
By weaving together formal theory, empirical examples, and conceptual analysis, we aim to present a comprehensive account of epistemic affordances in active inference. We show that Friston’s idea – agents implicitly guided by information-seeking affordances – has far-reaching implications for understanding curiosity, intentionality, and even the contours of conscious experience (with references to relevant literature in each case). The discussion is framed within an academic style, with references to Friston’s work and related sources in cognitive neuroscience and philosophy of mind.
Active inference stems from the free energy principle, which proposes that any system persisting in a changing environment acts so as to minimize surprise (variational free energy) about its sensory states. For cognitive agents, this minimization is achieved by maintaining a generative model of the world and updating its internal states (beliefs) and actions to maximize model evidence (or equivalently minimize free energy). In practice, one computes the expected free energy of possible policies (sequences of actions) and chooses the policy with lowest expected free energy. Crucially, the expected free energy G(π) of a policy π contains two terms: an extrinsic (pragmatic) term that drives the agent toward preferred outcomes, and an epistemic term that drives the agent toward uncertainty resolution. The extrinsic term is the negative expected utility (e.g. achieving goals), while the epistemic term is the expected information gain about hidden states or model parameters. In Friston’s words, “the intrinsic value of a policy is its epistemic value or affordance… the expected information gain afforded by a particular policy.”
This decomposition mirrors the exploration–exploitation tradeoff known in reinforcement learning: agents sometimes explore to learn about the world even at the cost of immediate reward. Within active inference, however, this tradeoff arises naturally from Bayesian principles. When an agent is uncertain about its environment, the epistemic term dominates: actions that gather information (reduce uncertainty) have high expected value. As confidence increases, the epistemic value of those same actions declines, and the agent shifts to exploiting known rewards. For example, in a classic T-maze task, a rat may initially search all branches (to find where food is), driven by epistemic affordances of cues; once the food location is learned, the epistemic benefit disappears and the rat reliably exploits the reward path. Parr et al. note that “systems that attain general steady-states will look as if they are responding to epistemic affordances” – i.e. behaving bayes-optimally to resolve uncertainty.
Importantly, Friston’s scheme bridges information theory and control theory. The expected information gain (epistemic value) can be formalized as the mutual information between future states and observations. Equivalent formulations call it Bayesian surprise or expected reduction in posterior entropy. Such formal identity with information-theoretic measures has led to connections with other fields: Schmidhuber’s theory of artificial curiosity (maximizing model improvement), optimal experiment design (Lindley), and perceptual salience in neuroscience. In essence, an action is epistemically valuable if it “offers the agent the opportunity to learn ‘what would happen if I did that?’”
Thus active inference gives us a precise foundation for epistemic affordances: they are not vague possibilities for action, but quantifiable opportunities for reducing uncertainty as encoded in a generative model. This view of action selection – balancing goal achievement with curiosity-driven exploration – aligns with many ideas in cognitive science about intrinsic motivation and active perception. The next section explores how this formalism connects to the broader notion of affordances from ecological psychology.
The term affordance originates with James Gibson’s ecological psychology. Gibson (1979) famously defined affordances as the actionable possibilities in the environment, relative to an organism (e.g. a chair affords sitting). Gibson’s affordances emphasize the direct perception of opportunities without needing internal computation. In subsequent decades, cognitive scientists have extended the idea: besides physical affordances (e.g. graspability of objects), there are social affordances, attentional affordances, and (relevant here) epistemic affordances.
An epistemic affordance can be loosely described as an environmental situation that affords knowledge gain. For instance, a visual pattern that is ambiguous at first but can become clear if the observer moves closer would be an epistemic affordance: it invites the agent to explore for information. Schwarzfischer (2021) outlines this idea in design terms, but it also permeates ecological theories. Bruineberg and Rietveld (2014) discuss how affordances can be long-term goals for exploratory behavior; Linson et al. (2018) note that agents show “sensitivity to long-term epistemic affordances.” Linson et al. emphasize that epistemic affordances relate to salience and novelty: salient affordances reduce uncertainty about the world’s state (immediate sensory variables), while novel affordances reduce uncertainty about the model’s parameters or context.
From an active inference standpoint, these ecological concepts are naturalized. The environment presents many affordances for the agent, but the generative model (and its priors) bias which ones are relevant. Active inference suggests that the agent will preferentially notice and sample those affordances that are expected to be informative. For example, attending to a moving object confers high salience if it resolves ambiguity in visual input; conversely, a motionless but unfamiliar object offers novelty (in uncertainty about hidden causes). Pezzulo et al. describe this as epistemic foraging: agents plan and execute actions that maximize expected information gain about hidden states (the world) or contextual variables.
This formal view dovetails with earlier ideas of epistemic actions (Kirsh & Maglio 1994), where agents simplify cognitive tasks by actively changing the world to gain information (e.g. physically rotating a puzzle instead of mentally simulating it). It also links to the idea of Bayesian experimental design: an agent as scientist choosing tests to maximize learning. In robotic and developmental contexts, algorithms inspired by intrinsic motivation and curiosity (Schmidhuber 2006; Oudeyer & Kaplan 2007; Barto et al. 2013) effectively implement epistemic affordances by driving exploration. Linson et al. note that when an agent pursues the “free-energy principle,” uncertainty-reducing epistemic policies naturally ensure it quickly finds out “what would happen if I did that.”
We illustrate with a simplified example. Consider a mobile robot (or virtual rat) in a T-maze with cues at branch points. Early on, the robot does not know which branch leads to the goal, so all cues have high epistemic affordance. Active inference drives the robot to sample them (epistemic action) to reduce uncertainty. As learning proceeds and one branch is known to be rewarding, the epistemic affordance of the other branches vanishes, and the robot shifts to straightforward exploitation. This switch from exploration to exploitation has been repeatedly simulated (e.g. an active inference model of vicarious trial-and-error) and matches animal behavior. Importantly, this behavior “comes free” from the formalism: the single objective of minimizing free energy entails both exploring unknown affordances and avoiding surprises once the world is learned.
In sum, epistemic affordances in active inference are precisely those action opportunities that promise high expected information gain. They can be understood as facets of Gibsonian affordances that pertain to knowledge acquisition. As such, they provide a bridge between ecological, embodied accounts of perception and formal Bayesian models of cognition. Equipped with this concept, we can now ask: what does it mean for a cognitive system to be guided by epistemic affordances? In particular, how does this relate to an agent’s inner experience, sentience, and consciousness? We address these questions in the next sections.
What does it mean, phenomenologically or biologically, to seek epistemic affordances? In practical terms, any animal or agent that attends to ambiguous stimuli or explores novel situations is, in effect, acting on epistemic affordances. A mouse poking its whiskers around a dark box is sampling uncertainties about tactile cues; a person reading multiple sources of news is sampling ambiguous information about reality. Active inference suggests that such behaviors are not incidental but functionally essential: organisms must reduce uncertainty to survive and self-organize.
Sentience, broadly, refers to the capacity to have subjective sensations or feelings – essentially, what it is like to experience being a living agent. By itself, the active inference framework does not directly explain qualia (the raw feel of experience). However, it does imply a form of minimal sentience: any agent that models its environment and tries to fulfill its needs must possess some internal state representing sensory evidence. In Friston’s words, self-organizing systems “actively confirming that ‘I exist’” through self-evidencing behaviors. When an agent pursues epistemic affordances, it is essentially tuning into the world in a purposeful way, presumably accompanied by changes in neural activity and possibly feeling. For example, reducing uncertainty about immediate threats or opportunities would plausibly be accompanied by affective valence (relief, confidence) and attentional engagement – markers of subjective salience.
Active inference accounts tie closely to interoception and affect: agents not only sample external affordances but also monitor internal bodily states as part of their model. Hence, epistemic affordances may include those that clarify interoceptive ambiguities (e.g. movements that resolve whether one is hungry or just thirsty). In this view, sentience arises from the integrated loop of perception and action: by acting to test its model of itself and world, the agent maintains a minimally coherent sense of “there is something that it is to be me” (a self-model). Friston et al. have argued that having any self-model and pursuing evidence for it (self-evidencing) is tantamount to a primitive conscious “I”: the model’s own existence is continually inferred and affirmed.
By extension, an agent sensitive to epistemic affordances will have a dynamic interplay of expectancies and surprises. This interplay underwrites aspects of sentience like the feeling of uncertainty or surprise. When an outcome fails to match the prediction, a salient (high epistemic value) event occurs, commanding attention. Conversely, smoothly predicted or learned affordances may fade from awareness (the information is no longer new). In other words, epistemic affordances may correspond phenomenologically to the “interest” or “curiosity” one feels when encountering novelty, and the cognitive closure or satisfaction that follows learning.
A practical example: flow states in skilled activity. The neuroscience of flow (e.g. during dancing or sports) suggests a balance between challenge and skill, with focused attention and reduced self-consciousness. Recent work (Limanowski & Blankenburg 2013; Limanowski & Friston 2018) using active inference frames flow as a state where only goal-relevant (pragmatic) affordances dominate, and epistemic affordances are suppressed. In flow, the system has encoded high confidence in the task (“no posterior uncertainty”) so that exploration (epistemic action) is unnecessary. This results in the phenomenology of effortless action and loss of self-as-object (the “flow” experience). In contrast, mundane or uncertain tasks invoke sensitivity to many epistemic affordances, leading to more reflective thought and awareness of self as one solves problems. Thus active inference suggests that the feel of engagement depends on the balance of epistemic vs. pragmatic affordances being actualized.
In sum, epistemic affordances play a crucial role in shaping sentient behavior. Any creature that actively gathers information – through saccades, foraging movements, social queries – is effectively tuning its attention and affect according to the expected value of knowledge. In this sense, being sentient could be seen as having the capacity to detect and exploit epistemic affordances in one’s environment and body. Note that this view does not require consciousness as in access to language or explicit reasoning; even low-level organisms (microbes, plants) can exploit chemical gradients or sunlight (epistemic affordances about nutrient distribution) without being “aware” in a human sense. It simply posits that living adaptive systems, by being driven to reduce surprise, will display behavior that looks curious and exploratory.
Moving from sentience to consciousness usually implies higher-order features: self-reflection, subjective awareness of experience, and reportability. How might epistemic affordances inform theories of consciousness? Active inference and predictive processing proponents (Seth, Hohwy, Friston, etc.) have suggested that conscious perception arises when predictions about sensory causes include a model of the self (a deep self-model) and when certain precision-weighting of signals occur. In this context, epistemic affordances are tied to what Hohwy (2013) calls “precision-weighted prediction errors” that reach conscious access. An action choice among epistemic affordances can be seen as a form of attention: an agent “chooses” to sample a specific sensory input because it expects high information.
One way to frame this is via the global workspace or integration of information. When an epistemic affordance is realized (an exploratory action yields new data), the resulting prediction error may become globally broadcast to update the model. The ease or difficulty of integrating this new information could relate to the vividness or clarity of conscious percepts. Conversely, if an affordance is pursued mostly subcortically or robotically (as in over-learned habits), the information may remain unconscious. In active inference terms, only those uncertainties that surpass the precision threshold (salience) will pull processing resources. Thus an epistemic affordance is not conscious per se, but its outcome – the reduction of uncertainty – can register in the agent’s phenomenal awareness as a resolved prediction.
Some theorists explicitly link epistemic dynamics to consciousness. For example, Deane (2021) argues that sensitivity to long-term epistemic affordances can explain complex self-modeling and even ego-dissolution in psychedelics. The idea is that a deep hierarchical model anticipates future epistemic opportunities (e.g. to learn about one’s long-term projects or social environment); consciousness may emerge when one reflects on such extended affordances. Active inference could also illuminate social consciousness: attending to other agents’ behaviors involves epistemic affordances about their hidden states (theory of mind). In this way, consciousness might be seen as an extended inference process over complex affordance landscapes.
Philosophically, the epistemic affordance view resonates with enactivist and predictive processing accounts of consciousness. Enactivists (Noë, Thompson, Hutto et al.) emphasize that perception and cognition are active and world-involving. Predictive processing advocates claim that consciously perceived world is the brain’s best guess at affordances that fit its model. Here we can state: Friston’s formalism gives a precise meaning to such speculation. An agent’s conscious beliefs about the world will align with the habitual exploitation of epistemic affordances it has learned. There is no separate “Cartesian theater”; rather, awareness reflects the dynamic tension of reducing prediction error about the self and world.
However, it is important to be cautious. Active inference itself is a functional account, not a solution to the “hard problem” of why experience is felt at all. Sensitivity to epistemic affordances might correlate with conscious attention, but does not automatically explain the qualitative feel (qualia). For example, two agents might both fixate on a novel object (responding to an epistemic affordance), but one might be fully awake while the other is anesthetized – the formal process of information gain is similar, but the subjective experience differs dramatically. Thus, we can say that epistemic affordances shape which information enters the system and thus what is potentially experienced, but the embodiment of that experience (the “what-it-is-likeness”) likely depends on additional neural or phenomenological factors.
In cognitive science, connections have been drawn between active inference and existing consciousness theories. For instance, Metzinger’s self-model theory (2010) suggests that consciousness arises from a transparent self-model; active inference can supply a framework for how a multilayer model of self (interoceptive, attentional, social) is maintained and updated. Seth (2015) has proposed that affective experience (“interoceptive prediction errors”) plays a key role in consciousness. Under active inference, interoceptive signals themselves present epistemic affordances (e.g. moving or ingesting to resolve hunger), linking homeostatic drives to conscious feelings of hunger or satiety.
To summarize this section: epistemic affordances frame consciousness as oriented toward information. A conscious agent is one that not only seeks sensory goals but also reflects on the uncertainty of its own model. Consciousness involves tracking how well one’s predictions are doing (metacognition), which naturally arises when an agent monitors its free energy over time. In flow or deep focus, epistemic exploration is minimized and consciousness “narrows” (loss of self-awareness). In learning or social contexts, epistemic exploration is high and consciousness “broadens” to incorporate new information. These align with phenomenological reports, suggesting the active inference account offers a unified view: attention and awareness are essentially determined by epistemic affordances at multiple levels of the cognitive hierarchy.
While appealing, Friston’s epistemic affordance framework invites several criticisms. Critics often target active inference broadly, and these critiques also affect the notion of epistemic affordances in particular.
Falsifiability Critique. The free energy principle is so general that it seems tautological (any adaptive behavior can be explained post hoc as free energy minimization). If every action can be called either epistemic or pragmatic by adjusting the model, then “epistemic affordance” might lose explanatory power. In reply, proponents point to specific models and experiments (e.g. agents simulated to perform epistemic foraging) as concrete tests of the theory’s predictions. The challenge remains to design critical tests that distinguish active inference from simpler heuristic accounts (e.g. ε-greedy exploration).
Over-Representation. Ecological psychologists (Gibson, Chemero, Rietveld, Bruineberg) might argue that internal Bayesian models are unnecessary overhead: organisms directly perceive affordances without constructing detailed probabilistic representations. From this view, talking about “expected information gain” in the organism’s brain could be seen as anthropomorphizing or mischaracterizing animal behavior. Friston and colleagues counter that the formalism is a process theory, not necessarily a literal description of neural code, and that it can be interpreted in embodied terms.
Empirical Evidence. While the math is elegant, direct neural evidence for the free-energy computation and epistemic value signals is still sparse. Studies have found brain correlates of uncertainty and novelty signals, and dopamine has been linked to salience (epistemic value), but the link to Friston’s precise expected free energy remains indirect. Skeptics (e.g. Clark 2016) emphasize that many models of decision-making produce similar exploration behavior without invoking variational free energy.
The Hard Problem. Critics say active inference captures cognitive function but does not address the “hard problem”: why those neural processes have subjective quality. They caution against hand-waving explanations where epistemic affordances become “phenomenological features.” We acknowledge this: epistemic affordance models the function (information-seeking) but not the feeling of curiosity itself.
Representational Debate. Enactivists deny the need for internal world-models. The counterargument is that Friston’s scheme can be seen as enactivist-friendly, because it anchors representations in action and embodiment, and because the “model” is really the physical organism itself (viewed through a statistical lens).
In short, epistemic affordances within active inference spark rich debate. Some see them as a breakthrough unifying concept; others see risks of metaphorical overstretch or lack of parsimony. Our view is that while caution is warranted, the idea has explanatory fruitfulness: it connects information theory, behavior, and phenomenology in novel ways.
Karl Friston’s notion of epistemic affordances in active inference offers a powerful lens for understanding how agents seek information. The framework shows that curiosity and exploration are not mysterious add-ons but mathematically necessary components of any adaptive system minimizing free energy. We have seen that agents acting under this principle naturally behave “as if” the world is inviting them to resolve uncertainty: eye movements, foraging, attention, and even social inquiry can all be interpreted as exploiting epistemic affordances. Moreover, this perspective sheds light on sentience and consciousness by linking subjective attention, motivation, and affect to the same underlying drive for knowledge. In particular, modes of experience (flow, curiosity, doubt) correspond to different balances of epistemic and pragmatic affordances.
However, linking epistemic affordances to consciousness must be done carefully. Active inference describes what cognitive processes do, not why they feel a certain way. It remains an open question how far this functional story can carry us toward a theory of consciousness. Future work might clarify which neural mechanisms implement epistemic value computation, and how this relates to phenomenological reports of curiosity or insight. Empirical tests (e.g. manipulating information salience vs. reward) will help validate or constrain the theory’s claims.
In academic terms, this essay has aimed to deepen the concept of epistemic affordances by integrating cognitive neuroscience, cognitive psychology, and philosophy of mind. The active inference framework provides an elegant quantitative formalism, but it gains richness when connected to ideas of affordances, enactive perception, and conscious experience. We conclude that epistemic affordances are a fruitful research concept: they emphasize that cognition is not just about achieving goals, but also about actively knowing. As Gibson noted for physical affordances, the most complete theory of mind may be one that understands not just what the world is, but what the world offers – especially in terms of information.