In standard formulations of the causal Markov condition, each variable in a system is assumed to be probabilistically independent of its non-effects, given its direct causes. This principle underlies much of contemporary work in causal inference and bayesian networks. When we turn to cognitive systems, however, the underlying assumptions of locality, variable individuation, and temporal order that make the causal markov condition plausible in physical domains become far less straightforward. Cognitive processes unfold in a massively parallel architecture, are implemented in a high-dimensional and recurrent brain, and exhibit context sensitivity that is hard to encode in simple directed acyclic graphs. As a result, directly importing the standard condition into models of perception, thought, and action threatens to mischaracterize the structure of mental dependence and misidentify which variables are genuinely screening off others.
A first difficulty arises from the way cognitive states are individuated. In canonical applications, variables are relatively well-defined physical magnitudes or events, such as spins, temperatures, or switches. By contrast, mental states are functional and representational: beliefs, desires, intentions, and perceptual experiences are characterized by their roles in reasoning and by their semantic contents. When a belief is realized in distributed neural patterns, it may not correspond to a single node but to a complex pattern spanning many nodes and time slices. The causal Markov condition presupposes a set of variables where “parents” and “children” can be specified at the right grain of description. Cognitive systems undercut this assumption, as the same neural substrate participates in multiple overlapping computations and representations, blurring the distinction between direct and indirect causes.
Moreover, the brain is heavily recurrent, with loops connecting higher and lower levels of processing, as well as lateral connections among peer areas. In a simple feedforward network, causal arrows can be ordered neatly from early to late processing stages. In contrast, in a recurrent system, what looks like an effect at one temporal slice can also be a cause of its earlier inputs via feedback signals. Standard causal bayesian networks rely on directed acyclic graphs, which rule out cycles by design. Cognitive systems, however, appear to violate acyclicity at the implementational level: activity in frontal regions can modulate processing in sensory cortices, which in turn changes the very evidence that influences frontal regions. The causal Markov condition, originally framed for acyclic structures, thus requires reinterpretation or extension to accommodate the temporal unrolling and dynamical nature of these feedback loops.
Predictive processing and related frameworks sharpen this tension. On these approaches, the brain continuously generates top-down predictions about incoming sensory signals, compares them with the actual inputs, and propagates prediction errors up the hierarchy to revise its internal models. At any given moment, activity in a cortical area is governed jointly by bottom-up error signals and top-down predictions shaped by prior beliefs. If we attempt to model such a system using the standard causal Markov condition, we would expect that once we condition on the direct causal parents of a prediction error signal—say, local prediction units and incoming sensory data—other, more distal variables become probabilistically irrelevant. Yet in predictive coding architectures, distal prior states can exert a residual influence even after conditioning on local parents, because priors are shared across multiple processing pathways and timescales. This undermines the idea that a small, local set of parents cleanly screens off everything else.
Another pressure on the classical condition comes from the pervasive role of internal modulation and control in cognition. Attention, expectation, and task goals all influence how information flows, which variables are even instantiated, and which connections are effectively “open” or “closed” at a given time. In a causal graph that respects the causal Markov condition, the existence and strength of arrows are typically treated as fixed features of the system. By contrast, in a cognitive system, the effective connectivity between, say, visual areas and decision-making circuits depends on current goals and attentional focus. This makes the relevant causal structure context dependent at a fine temporal grain. Any fixed set of conditional independence relations that might be captured by a single causal Markov condition will be an approximation at best, applying only under a specific attentional and task configuration.
Representational content poses an additional challenge. Many mental states bear semantic relations to external states of affairs: a belief that it is raining is about the weather, not merely about neural configurations. When causal models are applied to physical systems, variables are typically individuated in non-semantic terms. However, cognitive models often use representational descriptions as variables precisely because they are explanatory in psychology and cognitive science. The difficulty is that semantic relations can introduce systematic dependencies that do not track local causal structure in the brain. Two beliefs about related propositions may covary because of shared inferential roles or shared background assumptions, rather than because one directly causes the other. A naive application of the causal Markov condition might misinterpret these semantic dependencies as evidence of direct causal connections, rather than as consequences of a higher-level inferential architecture that is not easily captured by local conditional independencies.
Temporality and memory further complicate the picture. Cognitive systems integrate information across multiple timescales: short-term perceptual traces, working memory states, long-term memories, and slowly changing traits all interact. In physical systems, adopting a Markovian description often involves bundling enough of the past into the present state so that future states depend only on that present. For minds, however, there may be no natural way to define a present cognitive state that screens off all relevant past states without collapsing the distinction between different forms of memory and representation. Memory traces themselves are often reactivated and reconsolidated, changing their content and strength in ways that cannot be captured simply by conditioning on current variables. This suggests that the naive Markovian assumption—that the present cognitive state renders the past and future independent—may fail systematically.
These structural and representational complications are amplified by the fact that cognitive systems are embedded in rich environmental and social niches. External artifacts, linguistic practices, and social interactions shape and partly constitute cognitive processes. If many of the causes of a cognitive state lie outside the individual organism—distributed in notebooks, devices, conversational partners, and institutions—then a model that confines itself to internal neural states will violate the spirit of the causal Markov condition. Hidden common causes in the environment will induce dependencies among internal states that cannot be eliminated by conditioning on any subset of purely internal parents. In such a case, the condition, as usually interpreted, encourages us either to treat environment variables as parents or to accept systematic violations of the expected screening-off behavior.
In addition, the normative dimension of cognition introduces complications that have no clear analogue in basic physical systems. In reasoning and decision-making, cognitive states are evaluated in terms of rationality, coherence with evidence, and responsiveness to reasons. These normative constraints structure dependencies among beliefs and intentions in ways that are not easily reducible to causal connections. For instance, if an agent revises a belief in light of new evidence, the resulting conditional dependencies among beliefs are guided by logical and probabilistic relations, not only by physical interactions in the brain. Applying the causal Markov condition in this setting presumes a clean separation between normative relations and causal structure, yet in practice, many models of reasoning intertwine the two. This raises the possibility that causal models of cognition must explicitly accommodate inferential and normative dependences, revising what counts as a legitimate screening-off relation.
Taken together, these considerations indicate that a straightforward transplantation of the standard causal Markov condition from physical systems to minds is unlikely to succeed. Cognitive systems exhibit recurrent connectivity, context-dependent effective structure, distributed implementation of representational states, and deep entanglement with environmental and social scaffolding. Each of these features disrupts the tidy mapping from causal parents to conditional independencies that underwrites the traditional condition. Any adequate revision of the principle for cognitive systems will have to relax, re-interpret, or supplement the original assumptions in ways that respect both the dynamical complexity of the brain and the representational and normative dimensions of mentality.
Challenges for applying causal bayesian networks to minds
Attempting to apply causal bayesian networks to minds immediately runs into the issue of how to carve the system into variables without distorting the phenomena under study. Classical uses of bayesian networks assume a relatively clean mapping from physical or experimental manipulations to nodes: a switch is on or off, a particle decays or not, a treatment is administered or withheld. Mental states, by contrast, resist such neat discretization. A single belief may be graded in strength, contextually activated or inhibited, and partially implicit rather than fully explicit. Emotions and perceptual experiences can blend, fade, and reconfigure in ways that make it unclear whether we should treat them as one variable, many variables, or as parameterized families of states. When the grain of individuation is wrong, the causal markov condition can be made to fail or hold trivially, not because of any deep fact about cognition, but because the modeling choices are too coarse, too fine, or misaligned with the system’s functional architecture.
The difficulty of variable selection is magnified by the multi-level nature of the brain. A single psychological construct may be implemented by large-scale networks, local microcircuits, and synaptic changes, all interacting across time. If we define variables at the neural level, the network may need millions of nodes to capture the relevant dynamics; if we define variables at the psychological level, we risk erasing the very causal mechanisms that underpin mental transitions. Causal inference using bayesian networks presumes that the chosen set of variables is reasonably closed under the causal processes at issue, such that adding or removing a few nodes does not radically alter the overall pattern of independencies. In cognitive science, however, moving between implementational, algorithmic, and computational descriptions can dramatically change which conditional independencies appear and which are masked, making it difficult to find a level of description at which the causal markov condition is both non-trivial and empirically adequate.
Recurrent and dynamical features of cognitive systems pose additional obstacles. Standard bayesian networks rely on directed acyclic graphs to represent causal structure, but the brain is suffused with feedback loops: higher-level areas project back to lower-level ones, lateral interactions abound, and neuromodulatory systems broadcast signals that influence vast regions in parallel. A common workaround is to “unroll” time and treat the system as a dynamic bayesian network with separate nodes for each variable at each time step. Yet this move depends on being able to specify a time discretization that captures all relevant causal relations without smearing them across temporal slices. Neural processes often evolve on multiple overlapping timescales, with fast oscillations nested within slower integrative processes, and with synaptic plasticity unfolding over minutes, hours, and days. Any single temporal grid risks misrepresenting which events are genuinely contemporaneous and which are earlier or later, undermining the basic assumptions that license the causal markov condition in dynamic settings.
Another pervasive challenge comes from latent structure and unobserved common causes. In classical causal inference, violations of the expected independencies often signal hidden variables that should be added to the model. For mental systems, however, many of the most important influences are precisely those that are hard to measure or even to characterize: background knowledge, long-term character traits, entrenched habits, implicit biases, and stable environmental affordances. These factors simultaneously shape multiple mental states, inducing dependencies that cannot be removed by conditioning on any set of readily accessible variables. Because we rarely have direct access to the full complement of priors and long-term dispositions that structure an agent’s cognition, a bayesian network constructed from observable behavior or self-report will systematically violate the predicted screening-off relations, not due to the failure of causal markov at a fundamental level, but because the relevant hidden causes are pervasively missing from the graph.
Interventions, which play a central role in the theory of causal bayesian networks, are also problematic in the mental domain. Ideal interventions are supposed to change the value of a variable while leaving the rest of the system’s causal structure intact. In the laboratory, we can approximate such manipulations in physical systems with clamps, switches, and targeted stimuli. Mental interventions, by contrast, tend to be diffuse and meaning-laden. Presenting a sentence, an image, or a choice option rarely changes just one mental variable; it reconfigures a network of beliefs, expectations, and affective states in ways that are partly determined by the subject’s history and context. Pharmacological manipulations and brain stimulation may appear more localized, but they also modify global states such as arousal, attention, and mood. As a result, the mapping between experimental manipulations and clean surgical interventions on mental nodes is tenuous, reducing our ability to use interventionist tests to validate or refute proposed bayesian network structures in cognitive science.
At the methodological level, data limitations significantly constrain causal modeling of minds. Many of the canonical results linking bayesian networks to the causal markov condition rely on large samples and stable probability distributions. Cognitive phenomena, however, often change with learning, fatigue, and strategic adaptation. The distribution of responses in a decision task today may differ systematically from the distribution tomorrow, even for the same individual, because exposure to the task modifies underlying priors, heuristics, and attentional habits. Treating all observations as draws from a single stationary distribution glosses over these shifts, but accommodating time-varying parameters within a bayesian network quickly leads to models of intractable complexity. The tension between the need for rich, dynamic models and the constraints of limited, noisy data makes it difficult to infer fine-grained causal structure without overfitting or imposing strong, and potentially distorting, prior assumptions.
Furthermore, the normative and representational dimensions of cognition blur the boundary between causal relations and logical or evidential constraints. In classical applications of bayesian networks, dependencies reflect physical interactions or shared causes. For mental states, patterns of dependence often arise from constraints of rationality and consistency: if an agent believes that all mammals are animals and that all dogs are mammals, then rationality demands that believing “this is a dog” should cohere with a high probability assigned to “this is an animal.” These relations are not merely statistical regularities; they track inferential norms and semantic structures. When such patterns are encoded as probabilistic dependencies in a network, the resulting graph risks conflating rational connections with causal ones, thereby making the meaning of “parent,” “child,” and “screening off” ambiguous. The causal markov condition, interpreted naively, then misinterprets inferential dependence as evidence of underlying causal wiring, obscuring the distinct kinds of explanation at play.
The context sensitivity of cognition also conflicts with standard modeling strategies. Causal bayesian networks typically assume that the conditional probability distributions attached to each node are context-invariant: once the parents of a variable are fixed, its distribution does not depend on what other variables are being observed or manipulated. Yet cognitive processing is heavily modulated by task demands, social cues, emotional state, and environmental structure. The same sensory input may elicit different perceptual categorizations depending on instructions, incentives, or subtle framing effects. Likewise, the transition from one belief state to another can vary with conversational context and background assumptions. Capturing these phenomena would require an explosion of context-specific conditional probability tables or the introduction of higher-order context variables, which again threatens tractability and raises the question of whether a single, stable causal bayesian network can represent a mind across its diverse engagements.
Another difficulty arises from the way in which cognitive systems routinely exploit external structures as part of their processing. If notebooks, diagrams, digital devices, and social collaborators function as extensions of memory and reasoning, then the relevant causal processes cut across the boundaries of the individual organism. Classical bayesian networks are often constructed under the assumption that the system of interest can be reasonably demarcated from its environment, with exogenous variables representing external influences. For many cognitive tasks, however, the distinction between internal state and external scaffold is fluid: rearranging objects on a desk to solve a problem is as much a part of cognition as silent mental simulation. Modeling only internal neural or psychological variables yields patterns of dependence that reflect omitted external links and shared environmental causes, leading to persistent violations of the expected independencies. To restore conformity with the causal markov condition, one would have to broaden the network to include environmental artifacts and social partners, yet doing so dramatically increases complexity and raises contentious questions about where to stop.
There is an epistemic challenge specific to introspective and self-report data. Many causal models of mental states rely on subjects’ reports of their own beliefs, desires, and experiences, but introspective access is partial, theory-laden, and influenced by conversational norms. Self-reports tend to smooth over hesitations, ambivalences, and rapid fluctuations in state that may be causally important. When such reports are used as nodes in a bayesian network, the resulting variables may systematically misrepresent the underlying processes, collapsing multiple distinct internal states into a single coarse-grained category. The derived independence and dependence relations then reflect artifacts of linguistic practice and self-interpretation rather than genuine constraints among cognitive mechanisms. In this way, data collection practices themselves become an additional source of hidden structure that complicates any straightforward application of causal markov-based modeling to minds.
Mental causation and the structure of causal independence
Mental causation is often framed as the claim that mental states such as beliefs, desires, and intentions are not merely epiphenomenal correlates of neural activity, but play a genuine role in bringing about behavior and subsequent mental states. From the perspective of causal inference, however, we must ask what it means for one mental state to causally influence another, and how such influence is reflected in patterns of conditional independence. The standard picture inherited from the causal Markov tradition suggests that if a belief causally contributes to an intention, then conditioning on that belief should screen off the intention from more distal antecedents. Yet the internal organization of cognition makes this simple screening-off story fragile. Because mental states are embedded in webs of background assumptions, affective dispositions, and standing goals, there are often multiple pathways connecting the same pairs of states, some of which are semantic or normative rather than strictly mechanistic. The upshot is that the structure of causal independence among mental states does not map straightforwardly onto the causal Markov patterns familiar from physical systems.
One dimension of complexity arises from the fact that mental causation typically operates through structured representations rather than through featureless nodes. A belief that there is a glass of water on the table can contribute to a desire’s being satisfied, to the formation of an intention to reach for the glass, and to an expectation about how the scene will unfold. These transitions are sensitive to the logical and semantic relations among propositions: the belief that there is a glass of water supports an expectation that something drinkable is nearby because the agent also accepts generalizations about glasses, liquids, and drinking. When we attempt to encode such relations in a graphical model, dependencies may appear not because one mental state exerts an independent causal influence on another, but because they are both constrained by shared inferential norms or conceptual structures. In such cases, the absence of conditional independence is not diagnostic of a direct causal link in the sense presupposed by causal bayesian networks; it may instead reflect common membership in an inferential network that is itself realized across many neural and environmental substrates.
This becomes particularly clear when we consider chains of reasoning. Suppose an agent infers Q from P and the rule “if P then Q,” and then infers R from Q. If we model the mental states corresponding to P, Q, and R as nodes, we will observe strong statistical dependencies: P is correlated with Q, Q with R, and P with R. Conditioning on Q may fail to render P and R independent, because the same background rules and priors that support the inference from P to Q can also support alternative pathways from P to R. Here, the causal Markov template—according to which Q should screen off P from R—does not neatly apply, because what sustains the transition from one state to another is not a localized causal channel but a shared normative framework governing what counts as a good inference. To the extent that mental causation piggybacks on such normative patterns, the resulting structure of (in)dependence resists reduction to the kind of sparse, modular graph required by standard causal models.
Another source of complication is that the brain implements mental causation via distributed, recurrent dynamics. A single mental state is rarely realized by a compact neural configuration that serves uniformly as a parent node. Instead, it emerges from interactive activity across multiple regions, many of which are themselves engaged in representing other concurrent states. When a desire contributes to the formation of an intention, for example, the underlying neural processes will typically involve loops linking prefrontal decision areas, limbic valuation systems, and sensorimotor circuits. Within such loops, signals that correspond to the “same” mental state at different times continually shape and reshape each other. This cyclic structure makes it hard to isolate clean units that satisfy simple Markovian independence properties. Mental causation, in practice, unfolds as the evolution of a high-dimensional dynamical system, whereas the theory of causal Markov and bayesian networks is rooted in acyclic graphs whose nodes can, at least in idealization, be manipulated independently.
The structure of causal independence among mental states is further blurred by context-sensitive gating and modulation. Whether a given belief will causally influence an intention often depends on the agent’s current goals, attentional focus, and affective state. A standing belief that exercise is healthy may typically remain causally dormant until a particular cue—such as feeling lethargic or seeing a gym—activates it as a practical consideration. If we attempt to read causal structure off observed dependencies, the influence of this belief on behavior may be visible only in certain contexts, and conditioning on other variables, such as mood or situational cues, may either create or dissolve apparent dependencies. What looks like a violation of screening-off may simply be an effect of conditionalizing on variables that gate whether a causal pathway is open. The independence structure of mental causation is thus not fixed across contexts but dynamically reconfigured by higher-order control mechanisms, a feature that sits uneasily with the static, global character of the usual causal Markov condition.
Normativity also changes how we should interpret causal independence in mental life. In many domains, rational constraints demand that certain belief updates proceed in lockstep: learning that a source is generally unreliable should simultaneously lower one’s credence in many propositions derived from that source. From the standpoint of probabilistic modeling, this looks like a single event—the revision of reliability expectations—causing correlated updates across a wide network of beliefs. However, at the personal level, we are inclined to say that each belief is revised for its own reasons: the agent now judges that the evidence supporting each proposition is weaker than previously assumed. When we encode such changes as transitions among mental-state variables, we may find that no small subset of “parent” nodes fully screens off the rest: the same abstract consideration about evidential weight pervades all transitions. In effect, the structure of causal independence is shaped by reason-giving relations that operate at a global level, so that attempts to localize causation to a few discrete nodes risk misrepresenting the diffuse, pattern-based character of rational influence.
These pressures invite a distinction between at least two layers of causal organization: an implementational layer at which neural events interact according to biophysical laws, and a personal or cognitive layer at which mental states interact according to patterns of reasoning, goal-directed planning, and meaning. At the neural level, something like a causal Markov property may still hold once we identify appropriate state variables and time scales. But at the personal level, where we speak of beliefs causing intentions and experiences causing judgments, the relevant variables are individuated by content and rational role. Here, the absence or presence of conditional independencies often reflects the structure of conceptual and practical relations, rather than sparse physical connectivity. As a result, the structure of causal independence among personal-level mental states may systematically deviate from what we would expect if those states were merely coarse-grained summaries of an underlying Markovian process.
Failures of screening-off can also arise from the way mental states systematically co-occur as parts of larger configurations or “profiles.” Consider an agent’s political outlook, which might include clusters of beliefs about economic policy, social norms, and international relations. Specific beliefs within such a cluster often exhibit strong mutual dependencies even after conditioning on plausible mediators, because they are shaped by shared identity commitments, affective orientations, and source-trust patterns that are not easily decomposed into independent parents. When a new piece of information leads to a shift in one belief, the change tends to propagate across the cluster in ways that defy clean graphical separation. For modeling purposes, we can introduce higher-level latent variables representing identity or worldview, but these are themselves complex, evolving mental constructs rather than simple exogenous causes. Mental causation, in these cases, involves changes in a structured profile, and the independence relations observable among its parts are emergent products of that profile’s internal coherence conditions, not of separable, modular pathways.
Memory processes illustrate similar tensions. Recollection is not a passive retrieval of static traces but an active, reconstructive process shaped by current goals and expectations. When a cue triggers a memory, the content that comes to mind is influenced by prior experiences, background narratives, and the agent’s current interpretive stance. From a causal standpoint, a present cue and a stored trace together produce the recollection; yet other concurrent states, such as current mood or self-concept, can systematically bias the reconstruction. Conditioning on the cue and the trace may therefore fail to render the recollection independent of these broader factors. Moreover, repeated recollection alters the trace itself, so that past acts of remembering causally shape future memories in a way that entangles states across time. The resulting dependence structure is more like a web of mutually reinforcing interpretations than a sequence of temporally ordered events that satisfy a simple Markovian screening-off property.
Attention to these complexities does not entail abandoning the project of modeling mental causation, nor does it require rejecting causal inference tools altogether. Instead, it suggests that the structure of causal independence in mental systems is layered, context-sensitive, and partially governed by representational and normative relations. While neural-level descriptions may admit representations in terms of directed acyclic graphs with an associated causal Markov property, personal-level descriptions must reckon with the fact that beliefs and desires are connected by networks of meaning, reasons, and shared background commitments. These networks can generate pervasive dependencies that do not track direct causal influence in the narrow sense. Understanding mental causation, therefore, requires models that can differentiate between dependencies arising from mechanistic influence and those arising from rational or semantic organization, and that can represent how such forms of organization jointly shape the patterns of independence and dependence we observe in cognitive life.
Proposed modifications to the causal markov condition for mental states
Any attempt to adapt the causal Markov condition to mental states must begin by loosening the requirement that a single, fixed set of variables and edges capture all relevant aspects of cognition. Instead of a global principle that applies uniformly to every level and context, a more promising approach is to formulate a stratified, context-indexed version. On this view, distinct layers of description—neural, subpersonal computational, and personal—each admit their own approximate Markov properties, relative to appropriately chosen variables and timescales. At the neural level, we may retain a relatively orthodox condition, provided we work with state variables that coarse-grain over fast micro-dynamics and capture the main channels of information flow. At the subpersonal computational level, we define variables in terms of algorithmic components (e.g., prediction-error units, value-estimation modules) and require only that each such variable be conditionally independent of non-effects, given its direct algorithmic inputs, under a specified task context. At the personal level, where variables are content-laden beliefs, desires, and intentions, the Markov condition becomes explicitly contextual and partial: it applies within restricted domains of reasoning and practical deliberation, rather than across the entirety of an agent’s mental life.
This stratified picture motivates what might be called a layered causal Markov condition. Roughly, a variable at a given level L should be independent of non-effects at level L, conditional on its parents at L together with the relevant “interface variables” from adjacent levels (e.g., neural encodings feeding into a computational node, or summary outputs from a computational process feeding into a personal-level state). Instead of insisting that personal-level beliefs are screened off from one another solely by other beliefs, we allow that their independence relations may depend on variables representing subpersonal computations or even environmental scaffolds. Thus, the Markov condition is not abandoned but relativized: it states that, once we condition on a variable’s direct causes at the appropriate mix of levels, there should be no additional residual dependencies that track further causal pathways. Violations of this layered condition then serve as guides for where our description is missing key cross-level links rather than as refutations of causal Markov tout court.
A second modification concerns the static and acyclic nature of standard formulations. Cognitive systems are inherently dynamical and recurrent, so a more adequate principle should be framed in terms of trajectories rather than instantaneous states. One way to achieve this is to shift from node-based to path-based screening-off. Instead of requiring that the current value of a variable be independent of non-effects given its parents in a directed acyclic graph, we require that, given the full history of its causal ancestors along allowed dynamical paths over a specified time window, the variable be independent of non-effects not reachable by such paths. Technically, this amounts to a Markov property on paths in a dynamic system rather than on nodes in a static graph. For cognitive modeling, this opens the door to using tools from dynamical causal inference and state-space representations: we treat the brain as implementing a high-dimensional Markov process in continuous time, but allow mental variables to correspond to functionals of these trajectories, such as integrated prediction error over a period or the stable attractor towards which the system is moving.
This dynamical reinterpretation also suggests that some apparent violations of causal Markov at the personal level are artifacts of projecting a non-Markovian trajectory onto a coarse set of instantaneous mental variables. To address this, a revised condition could explicitly license “thick” state variables that summarize a window of underlying dynamics. For instance, instead of a belief being identified with a momentary neural pattern, it might correspond to a quasi-stable region in phase space, characterized by its robustness across small perturbations and its pattern of couplings to sensory and motor channels. The modified condition would then say: conditional on the thick state that realizes a belief and its directly connected thick states (e.g., relevant desires, currently active percepts), further past states of the system are causally irrelevant for predicting near-future behavior, modulo explicit memory variables. This thicker Markov assumption acknowledges that memory, mood, and context must sometimes be treated as separate, slowly varying state variables that prevent any overly simple present-screens-off-past idealization.
A third family of modifications concerns the treatment of context, attention, and control. Standard causal Markov, as used in causal bayesian networks, typically presupposes that once we fix the parents of a node, its probabilistic behavior is stable across contexts. Cognitive systems emphatically violate this assumption. A reasonable adjustment is to treat attentional and task variables as higher-order modulators that enter the Markov condition as “conditionalizers on the condition.” Concretely, for each mental variable X, we specify a set of parents Pa(X) and a set of context variables C(X) that gate which causal links are operative. The revised requirement is then: for any fixed value of C(X), X is independent of its non-effects, conditional on Pa(X) and C(X). Here, context variables might include task demands, current goals, global arousal, or social framing cues. By explicitly building such modulators into the causal structure, we avoid misinterpreting context-dependent changes in dependence patterns as violations of Markov; they become expected consequences of conditioning on or marginalizing over C(X).
Nonetheless, adding context variables risks unmanageable complexity if every subtle environmental or internal factor must be explicitly represented. To avoid this, the revised framework can adopt a minimal-sufficiency principle: we introduce context variables only when observed violations of conditional independence remain robust across repeated measurements and manipulations that control for noise. In effect, the causal markov condition becomes a tool for detecting which contextual factors must be elevated to explicit causal status. When observed data systematically contradict the predicted screening-off relations for a given set of Pa(X), we search for candidate modulators M such that, once M is added to C(X), the modified condition is restored. This pragmatically oriented, iterative use of Markov violations as prompts for model enrichment fits the methodological realities of cognitive science, where models are often refined in light of surprising residual dependencies in behavioral, neural, or psychophysiological data.
A further revision must accommodate the fact that many dependencies among mental states arise from shared representational content and rational norms, not just from mechanistic influence. One way to handle this is to distinguish two kinds of arrows in our graphs: mechanistic-causal arrows and inferential-structural arrows. The former are meant to represent channels along which physical influence passes—for example, neural activity propagating from predictive coding units to motor plans. The latter denote constraints imposed by semantic and rational relations—such as the necessity, given one’s prior commitments, of adopting a certain belief when presented with particular evidence. We can then impose a restricted causal Markov condition only over the mechanistic-causal subgraph, while allowing the inferential-structural subgraph to generate additional dependencies that are not constrained by screening-off. In practice, this means that when we perform causal inference from data, we treat violations of Markov that can be accounted for by explicit inferential relations (e.g., known logical entailments or normative updating rules) as non-diagnostic of missing causal variables.
Formally, suppose that for each mental variable X we have two sets of parents: causal parents PaC(X) and inferential parents PaI(X). The modified condition states that X is probabilistically independent of its non-effects with respect to the causal graph, given PaC(X) and any relevant context variables, conditionalizing also on any logically or evidentially mandated relations that run through PaI(X). Violations of conditional independence that persist even after factoring in known inferential constraints then point to unmodeled mechanistic causes. This dual-graph approach allows us to capture the important role of priors, evidential relationships, and reasoning norms in shaping patterns of dependence without forcing them into a strictly causal template. It also aligns with the observation that some features of mental life—such as coherence requirements among beliefs—are best understood as constraints on rational updates rather than as causal transmissions in the sense presupposed by standard causal inference frameworks.
Central examples for this dual-graph perspective come from predictive processing theories of the brain. In these models, top-down predictions encode prior beliefs about the causes of sensory input, while bottom-up prediction errors convey information about mismatches between expectations and observations. In naive causal bayesian networks, we might represent prediction errors simply as children of priors and sensory inputs, and expect that, conditional on these parents, errors are independent of more distal aspects of cognition. Empirically, however, prediction errors often remain correlated with remote factors such as long-term motivational states or high-level interpretive frames, even after conditioning on local priors and current stimuli. On the revised view, some of these residual dependencies may be explained by inferential-structural connections—e.g., global assumptions about self or world that systematically shape the range of priors entertained across contexts—while others indicate omitted mechanistic modulators, such as neuromodulatory systems or slow-varying affective variables. The modified causal Markov condition thus operates as a sieve: dependencies that can be traced to explicit inferential structures are filtered off as non-causal, while unexplained residues prompt the inclusion of new causal parents or context nodes.
A more radical but complementary adjustment is to relax the requirement that every mental variable admit a clean Markov blanket within the mental domain alone. In standard bayesian networks, a variable’s Markov blanket—the set of its parents, children, and co-parents of its children—suffices to shield it from the rest of the network. For mental states, however, an adequate Markov blanket often includes environmental and social elements: notebooks, digital devices, conversational partners, and institutional roles that together determine how a state is maintained, updated, and expressed. The proposed revision is therefore to treat mental Markov blankets as extended, potentially crossing the skull and even the individual’s body. In extended or distributed cognition scenarios, a belief’s immediate mechanistic parents may comprise both neural encodings and stable environmental structures that are regularly consulted. The causal Markov condition is then evaluated relative to this broadened boundary: once we condition on the extended Markov blanket, the belief should be independent of more distal causes.
This extended-blanket move has two main payoffs. First, it renders tractable many cases where purely internal models exhibit stubborn apparent violations of causal Markov because external scaffolds are omitted. Second, it provides a principled way to decide when an external factor should be treated as part of a mental variable’s causal neighborhood rather than as a distant influence: if including the factor in the blanket systematically restores predicted conditional independencies across diverse tasks, we have good reason to treat it as functionally integrated with the mental state in question. This operationalizes some of the claims of extended mind theories in terms congenial to causal inference, while avoiding the need to draw a sharp metaphysical boundary between internal and external components in advance.
Finally, the revised framework should be indexed to timescales and learning regimes. Many cognitive processes are non-stationary: experience, practice, and aging reshape priors, heuristics, and processing strategies. It is implausible that a single, time-independent set of conditional independence relations could capture these evolving structures. A more realistic approach is to adopt a family of time-sliced or regime-specific causal models, each with its own approximate causal Markov property. For a given period in which an agent’s cognitive architecture is relatively stable (e.g., a particular phase of skill acquisition), we posit a model that satisfies the layered and context-sensitive conditions described above. As the system transitions to a new regime—say, after extensive training or significant environmental change—the model itself is updated, and with it, the set of Markov constraints. Rather than demanding that one static graph endure across all learning histories, we allow the graph to evolve, while using persistent patterns of (in)dependence across regimes as clues to deep structural features of the mind.
Within this dynamic-regime framework, the role of the causal Markov condition is less that of a universal law and more that of a guiding idealization: in each regime, we aim to identify a representation in which mental variables are “as Markovian as possible,” subject to the constraints imposed by recurrent dynamics, inferential structures, and environmental embedding. Deviations from ideal Markov behavior then have a clear interpretive function: they point either to missing levels, missing modulators, or to domains where personal-level descriptions are so thoroughly governed by normative and semantic relations that mechanistic independence notions must be supplemented by other explanatory tools. In this way, proposed modifications to the causal Markov condition for mental states do not simply weaken the principle; they reconfigure it into a multi-layered, context-sensitive, and methodologically explicit framework better suited to the complexity of minds.
Implications for philosophy of mind and cognitive science
Revising the causal Markov condition for mental states reshapes several core debates in philosophy of mind, beginning with the status of mental causation. If the usual screening-off pattern is not reliably observed among personal-level states, this loosens the grip of arguments that treat failures of independence as evidence against mental-level causal efficacy. Rather than inferring that beliefs and desires are epiphenomenal whenever their probabilistic profile does not conform to a simple graph, we can regard these irregularities as reflections of layered, context-sensitive organization. On a stratified view, mental events can be genuinely causal even when they participate in dense webs of inferential and normative relations that violate the sparse conditional independencies presupposed by standard causal bayesian networks. This allows defenders of non-reductive physicalism to say that mental properties are realized by neural states with something like a neural-level causal markov property, while permitting the personal level to exhibit a richer, more holistic dependency structure without forfeiting causal relevance.
This restructuring of dependence patterns affects how we think about reduction and multiple realizability. Traditional worries hold that if personal-level states fail to respect clean, Markovian boundaries, then they cannot figure in respectable causal explanations and must be replaced by lower-level neuroscience. The layered framework weakens that inference. Personal-level descriptions may lack strictly local Markov blankets within the mental domain, yet still hook into well-behaved mechanistic structures at the neural and computational levels. Philosophically, this supports a picture on which explanatory autonomy is compatible with a background of strict physical determination. Cognitive-level models remain indispensable because they track patterns of rational organization, semantic content, and task-specific context that are intractable to express purely in terms of neural trajectories, even though the brain obeys its own implementation-level causal markov constraints.
The proposed revisions also influence discussions of mental content and externalism. If Markov blankets for many cognitive states extend beyond the skull, then environmental and social scaffolds are not merely background causes but partial realizers of mental processes. This gives a new, operational gloss to externalist slogans: whether an external item counts as part of a cognitive system depends on whether including it in the extended Markov blanket restores predicted conditional independencies in our causal inference. For philosophy of mind, this offers a way to adjudicate disputes over the boundaries of the mind that is sensitive to empirical patterns of dependence, rather than relying solely on intuitive judgments about what “really” belongs to cognition. Extended beliefs, socially distributed memory, and tool-mediated reasoning can be treated as legitimate targets of causal modeling when they function as stable members of these broadened causal neighborhoods.
One immediate payoff for cognitive science is a more realistic conception of modeling practice. Instead of seeking a single, stationary network that captures an agent’s entire mental life, researchers are encouraged to build regime-specific and task-specific models whose conditional independencies hold only relative to particular contexts, attentional settings, and stages of learning. This is methodologically liberating: the goal is no longer to force all data into one master graph, but to identify families of models in which variables are “as Markovian as possible” for the limited domain at hand. In experimental design, this suggests shifting attention toward careful manipulation of context variables—goals, instructions, framing, social cues—and explicitly treating them as modulators in the graphical structure. Doing so reframes puzzling failures of screening-off not as defects of causal inference but as signals that crucial contextual parents have been omitted.
Similarly, the dual-graph distinction between mechanistic-causal and inferential-structural arrows clarifies how cognitive science can integrate normative theories of reasoning with mechanistic accounts of the brain. Models of belief updating, decision-making, and perception often conflate evidential and causal dependencies, leading to confusion about what sorts of interventions are appropriate and how to interpret departures from rational norms. Under the revised picture, inferential links encode rational or semantic constraints that shape probability distributions without necessarily corresponding to independent channels of physical influence. Causal graphs then track the flow of neural and computational signals, while inferential graphs encode how priors, likelihoods, and rules of reasoning jointly constrain possible belief states. This separation allows, for example, predictive processing models to attribute systematic misbeliefs either to distorted mechanistic pathways (e.g., aberrant precision weighting in the brain) or to atypical inferential structures (e.g., idiosyncratic background assumptions), rather than collapsing both into a single undifferentiated dependency network.
For debates over rationality, bounded reasoning, and idealization, the revised use of causal Markov has further consequences. Many normative theories tacitly assume that ideal reasoners can be modeled as if their beliefs and intentions were nodes in a well-structured graph satisfying the usual screening-off principles. Recognizing that real cognitive systems deviate from this template in systematic, interpretable ways suggests a richer taxonomy of reasoning failures. Some deviations will reflect noise or resource limits in the mechanistic implementation; others will stem from the fact that personal-level dependencies are partly governed by global coherence pressures, identity commitments, and long-term narratives that operate above the level of local Markov blankets. Rather than treating all such dependencies as irrational, we can classify them according to whether they arise from principled, content-based organization or from distortions in the underlying causal machinery. This reframing opens space for models of “structured bounded rationality” that respect both mechanistic constraints and normative ideals.
The shift to dynamic, regime-indexed models also reframes learning and development. If each learning phase corresponds to a partially distinct causal graph, then questions about concept acquisition, skill formation, and developmental change become questions about how graphs evolve—how new edges appear, how Markov blankets reconfigure, and how context variables become more finely differentiated. Cognitive science can treat observed transitions in independence structure as indicators of structural learning, not just parameter updating. For philosophy of mind, this supports pictures on which cognitive capacities are not static modules but evolving webs of causal and inferential relations. The familiar tension between nativism and empiricism can then be recast in terms of constraints on possible graph topologies and the kinds of environmental input needed to drive transitions between them.
Within cognitive neuroscience, the layered Markov view offers guidance on how to relate large-scale brain dynamics to task-level cognition. Techniques such as Granger causality, directed coherence, and dynamic causal modeling already probe directed dependencies among neural regions, but their interpretation is often contested. Embedding these tools within a multi-level causal markov framework clarifies what is being inferred: approximate mechanistic edges at the neural level that may or may not map cleanly onto personal-level states. The framework encourages explicit bridging constructs—algorithmic variables linking brain areas to psychological functions—and treats persistent mismatches between neural and personal-level independence structures as data about where existing cognitive theories mislocate or mischaracterize functional components. This tightens the connection between neural evidence and high-level theorizing without demanding a naïve one-to-one mapping from regions to folk-psychological categories.
The treatment of extended Markov blankets has parallel implications for social and cultural cognition. Many cognitive phenomena of interest to philosophers and social scientists—shared intentions, group beliefs, institutional facts—depend on coordinated patterns of interaction that no single individual’s brain fully contains. The extended-blanket approach suggests operational criteria for when such group-level constructs merit causal modeling: when conditioning on appropriate social and institutional variables restores predicted independencies among individual attitudes and behaviors. This provides a bridge between individualistic and collectivist explanations in psychology and the social sciences. Group-level states can be acknowledged as causally efficacious, not by stipulation, but when they demonstrably function as stable nodes in broader networks of dependence that include but transcend individual minds.
Reinterpreting causal Markov for mental systems reframes what counts as a successful explanation in cognitive science. Explanations that appeal to algorithms, representations, and reasons have sometimes been criticized for lacking solid grounding in causal structure. Under the revised framework, such explanations are vindicated when they identify variables and relations that sit at “sweet spots” of partial Markovianity—levels and contexts where conditional independencies are good enough to support reliable causal inference and prediction, even if not perfectly obeyed. This encourages pluralistic methodologies: mechanistic experiments, behavioral studies, and normative analyses can all contribute to refining the multi-level graphs that jointly characterize a cognitive system. Rather than aiming for a monolithic causal image of the mind, philosophy of mind and cognitive science can pursue a coordinated set of models, each tuned to particular questions and scales, and each evaluated by how well its causal and inferential structures capture the rich patterns of dependence that characterize thinking, perceiving, and acting.
