Literature Review: Extended Mind and Distributed Cognition: Thinking Beyond the Brain

This literature review systematically examines extended mind and distributed cognition theories, which argue that cognitive processes extend beyond the biological brain into tools, artifacts, and social environments. Synthesizing 58 works, it explores criteria for genuine cognitive extension, such as functional integration, and highlights disability and prosthetics research as key empirical evidence. Studies show environmental scaffolds measurably restructure cognitive work, while learning sciences indicate cognitive offloading requires careful calibration to avoid impeding durable learning. Research in expert and collaborative settings reveals competence often emerges from person-artefact ensembles, with outcomes depending on structures that foster genuine engagement. The framework challenges traditional views of memory and identity, particularly in dementia care, and raises urgent questions regarding generative AI and cognitive justice. It concludes that cognitive extension is an empirically grounded concept with significant implications for education, technology design, and equity, demanding interdisciplinary responses.

1. Introduction

The question of where cognition ends and the world begins has occupied philosophers and cognitive scientists for decades, yet it has rarely carried the urgency it does today. As artificial intelligence systems grow increasingly embedded in everyday reasoning, as digital prosthetics blur the boundary between biological memory and external storage, and as collaborative work is mediated by ever more sophisticated technological environments, the theoretical frameworks designed to describe human thought are under pressure to expand. The doctrine that cognition is confined to processes occurring within the skull — a view that long held near-axiomatic status in cognitive science — has been subject to sustained and searching challenge since the mid-1990s. That challenge, organized loosely under the headings of extended mind theory and distributed cognition, has generated one of the most generative and contested debates in the philosophy of mind and cognitive science [1, 2]. This systematic literature review synthesizes fifty-eight works spanning the period from 1994 to 2025 to map the terrain of that debate, assess the empirical and theoretical evidence on both sides, and examine what the accumulated scholarship now demands of our understanding of memory, skill, identity, and the ethics of cognitive life.

The intellectual provocation that animates this field was sharpened in the late 1990s by the proposition that cognitive processes need not be bounded by skin and skull to count as genuinely mental [3]. If a notebook reliably performs functions that would otherwise be executed by working memory, the argument runs, then on functionalist grounds the notebook is part of the cognitive system [3, 4]. This coupling thesis — simple in formulation, radical in implication — opened a fault line in cognitive science that has not since closed. Defenders elaborated and refined the original position [5]; critics raised concerns about the principled criteria needed to distinguish genuine cognitive extension from the trivial observation that humans use tools [1, 6]. That foundational controversy remains alive, and this review begins by attending to it carefully.

Alongside theoretical dispute, a substantial empirical tradition has developed. Research on external representations, cognitive offloading, and interactivity has demonstrated that the organization of information in the environment — the layout of a cockpit, the spatial arrangement of a chess problem, the annotations in a margin — can constitute part of the cognitive work, not merely its context [7, 8, 9]. Distributed cognition, as a framework, has proved particularly productive in the analysis of expert practice and collaborative systems, revealing how competence and intelligence are frequently properties of coordinated person-artefact ensembles rather than of individuals considered alone [10]. Learning sciences and computer-supported collaborative learning (CSCL) research has independently demonstrated that when cognition is distributed across learners, tools, and collaborative structures in educational settings, qualitatively different forms of knowledge — sharing, construction, and creation — emerge depending on how the distribution is designed and scaffolded [11]. This empirical body of work forms the second major focus of the review.

The third cluster of themes takes a deeper mechanistic and evolutionary view. Predictive processing frameworks [12, 13, 14], embodied cognition research [15, 16], and evolutionary accounts of brain-body-world scaffolding [17] collectively suggest that the brain’s architecture may be constitutively oriented toward offloading and coupling, rather than merely tolerating it. Memory in this light is not simply a biological storehouse occasionally supplemented by the external, but a fundamentally extended and reconstructive process shaped by the affordances of the material and social environment.

It is, however, the emergence of generative artificial intelligence that lends a particular timeliness to this review. Large language models and related technologies now routinely mediate reasoning, writing, and problem-solving in ways that raise acute versions of questions the extended mind literature has long posed in more abstract registers [18, 19]. Whether these systems constitute genuine cognitive partners or sophisticated instruments [20, 21], and what their integration implies for intellectual autonomy and skill development, are no longer purely philosophical questions. They carry immediate practical and institutional stakes, and the existing theoretical apparatus is only beginning to be brought to bear on them.

The practical urgency of these theoretical questions is nowhere more apparent than in the domain of disability and assistive technology. Millions of people worldwide navigate daily life with acquired brain injuries, progressive neurodegenerative conditions, and congenital cognitive disabilities, relying on memory aids, brain-computer interfaces [22, 23], structured environments, and digital scaffolding to sustain reasoning, communication, and identity. For these populations, the question of whether a cognitive prosthetic merely assists cognition or partially constitutes it is not a philosopher’s thought experiment but a determination with immediate consequences for selfhood, legal personhood, and the ethics of care [24]. The growing body of research at the intersection of extended mind theory, disability studies, and assistive technology provides some of the most empirically consequential real-world cases the field has produced, and this review integrates these findings systematically.

Finally, the review attends to the normative and political dimensions of situated cognition — a domain that has historically received less systematic attention than the theoretical and empirical questions, but which can no longer be deferred. If cognition is genuinely distributed across social and material environments, then access to cognitive resources is a matter of justice, and the design of cognitive environments is a form of political power. Questions about epistemic dependence, cognitive liberty, and the unequal distribution of external scaffolding require the extended mind framework to account for structural conditions, not merely functional couplings. Critical disability studies scholarship has sharpened this demand [25, 24], demonstrating that differential access to cognitive scaffolding constitutes a form of cognitive stratification with consequences that are both measurable and morally urgent.

Together, these five thematic clusters constitute a field at an inflection point — theoretically mature enough to have produced sustained internal critique, empirically rich enough to have moved well beyond armchair speculation, and newly consequential enough that its central questions are now inseparable from the most pressing challenges of technological life.

2. Methodology

The literature informing this review was identified through a systematic search of the OpenAlex database, using five targeted queries designed to capture the principal conceptual and empirical dimensions of extended mind and distributed cognition research. The queries addressed, respectively, the criteria for cognitive extension and tool use, empirical evidence concerning cognitive processes and environmental artefacts, distributed cognition in expert practice and collaborative systems, the philosophical implications of extended mind theory for memory, skill, and identity, and recent empirical developments within the broader 4E cognition framework through 2025 and 2026. This multi-query approach was adopted to ensure that both the foundational theoretical literature and emerging empirical work received adequate coverage.

Search and Selection

The search and selection process began with an initial pool of 201 candidate records returned across five queries. Relevance scoring with a threshold of 0.6 yielded 30 papers for inclusion, filtering out tangentially related material while retaining sources with demonstrable conceptual alignment with the review’s focus. Citation network expansion was also attempted; although no additional papers entered the corpus through conventional forward or backward citation tracing, this expansion stage examined the network and contributed 14 newly identified relevant papers against 93 rejected, achieving a coverage delta of 0.13 before the collection target was reached. The process terminated upon reaching the pre-specified collection ceiling of 90 candidates needed to support a final corpus of 30 papers.

A supplementary search was subsequently conducted targeting the intersection of extended mind theory [3] with disability studies, cognitive prosthetics, and assistive technology. This search was motivated by the recognition that disability contexts — including dementia care, brain injury rehabilitation, and brain-computer interface use [26, 27] — provide the most empirically consequential real-world instantiations of the extended mind debate’s core claims [6], yet were underrepresented in the initial corpus. The philosophical stakes of these contexts are underscored by work examining how disabled body-minds interact with sociotechnical environments [24], and by neurorehabilitation research demonstrating that BCI-based interventions can produce measurable functional recovery following stroke [28] — findings that bear directly on questions of where cognitive and agentive boundaries lie. The supplementary search yielded 12 additional papers meeting the same relevance and quality criteria, bringing the total corpus to 49 works.

A further supplementary search was conducted targeting the intersection of extended mind and distributed cognition [10] with learning sciences, educational psychology, and computer-supported collaborative learning (CSCL). This search was motivated by the recognition that educational settings — where cognition is deliberately distributed across learners, tools, and collaborative structures — provide a critical empirical domain for evaluating distributed cognition claims [29]. Additionally, learning sciences research offers the most detailed evidence concerning when cognitive offloading [8] and external scaffolding [30] enhance versus impede durable learning. The growing literature on 4E cognition — embodied, embedded, extended, and enactive approaches [15, 2] — also informed this supplementary strand, given its increasing empirical footprint in both educational and clinical domains. This search yielded 9 additional papers meeting the same relevance and quality criteria, bringing the final corpus to 58 works.

Quality filters were applied consistently across the candidate pool. Papers published outside a two-year recency window were required to meet a minimum citation threshold of five, ensuring that older work retained in the corpus had demonstrated some disciplinary uptake. To prevent temporal skew in the other direction, a recency quota of 35% was maintained, guaranteeing that contemporary scholarship — particularly the growing empirical literature on 4E cognition, assistive technology, and CSCL — was meaningfully represented alongside the canonical theoretical contributions.

Retrieval and Processing

All papers selected for inclusion underwent full-text analysis; there were no cases of metadata-only processing or irrecoverable retrieval failure in the final corpus. However, during the acquisition phase, 10 papers initially identified as suitable could not be retrieved and were replaced with alternative papers drawn from the remaining candidate pool. This substitution preserved the target corpus size without compromising the relevance criteria already established.

The resulting corpus spans publication years from 1994 to 2025, a range that reflects the intellectual arc of the field from its foundational formulations—most notably Clark and Chalmers’ original articulation of the extended mind hypothesis [3]—through to recent empirical and applied scholarship [18]. The 31-year span is analytically appropriate given that extended mind theory and distributed cognition emerged as coherent research programmes in the mid-1990s [5, 10], and any review that omitted this foundational literature would misrepresent the theoretical context within which contemporary empirical work is situated.

Thematic Organisation

Following selection and processing, the corpus was organised into five thematic clusters. This structure enabled the review to distinguish between different types of research, including theoretical debates over the criteria for genuine cognitive extension [3, 6, 14] — which encompass longstanding disputes about what constitutes the “mark of the cognitive” and the coupling-constitution distinction [1] — empirical studies of artefact-mediated cognition [7, 8], and analyses of distributed cognition in professional and collaborative settings [10, 31]. The clustering process grouped papers according to shared conceptual concerns rather than relying solely on methodology or publication period. This approach produced a thematic architecture that reflects the multi-layered nature of the field itself.

3. Foundations and Debates in Extended Mind Theory

The question of where the mind ends and the world begins has proven to be one of the most generative and contentious disputes in the philosophy of cognitive science. What began as a speculative thought experiment in the late 1990s has expanded into a sprawling research programme touching on neuroscience, robotics, phenomenology, and the social sciences, while simultaneously attracting a sustained body of internalist critique that has never been fully answered. Tracing this debate chronologically reveals not merely an accumulation of positions but a genuine dialectical evolution in which each intervention has sharpened, complicated, or partially undermined what came before.

Active Externalism and the Parity Principle

The modern debate originates with the landmark proposal by [3], which introduced active externalism: the thesis that cognitive processes can extend beyond the brain and body when external elements participate in real-time, reliable coupling with an agent’s mental processing. The paper’s rhetorical centrepiece contrasts Inga, who retrieves a memory location from biological memory, with Otto, who performs the functionally equivalent operation by consulting a notebook. Clark and Chalmers argued that if the two processes are functionally isomorphic, principled grounds are lacking for treating one as genuinely cognitive and the other as mere external scaffolding [3]. This Parity Principle—which states that if a process would count as cognitive were it inside the skull, it counts as cognitive wherever it occurs—became the organizing heuristic of extended mind theory. Crucially, the Principle applies only where the external element satisfies stringent coupling conditions: it must be reliably available, regularly engaged, and its outputs automatically endorsed by the agent, thereby distinguishing genuine cognitive extension from contingent tool use [6].

Clark subsequently developed and defended these commitments across two closely related monographic treatments [32, 5]. In these works, the argument broadened considerably: human cognition is characterized not by neat skull-bound representations but by the habitual recruitment of bodily actions, linguistic tools, and environmental structures as genuine components of cognitive machinery [32]. The Parity Principle was reaffirmed as a criterion that cuts through substrate chauvinism—the assumption that biological neural tissue enjoys a privileged cognitive status merely by virtue of its physical location [6, 1]. Empirical material from robotics and cognitive neuroscience was marshalled to show that biological nervous systems have always been “designed” to exploit environmental regularities [5, 16]. At this stage, the extended mind hypothesis stood not as philosophical speculation but, Clark argued, as a thesis with substantial empirical backing.

The Functionalist Foundation of the Parity Principle

The force of the Parity Principle — and indeed, the intelligibility of the entire extended mind debate — depends on a philosophical commitment that is often assumed rather than examined: functionalism about the mind. Functionalism, in its broadest formulation, holds that mental states are individuated not by their physical substrate but by their causal-functional roles — the patterns of inputs, outputs, and relations to other states that define what a mental state does rather than what it is made of [33, 34]. The Parity Principle is, at bottom, a functionalist argument: if an external process plays the same functional role as an internal process that we unhesitatingly call cognitive, then the external process merits the same label, because functional role is what makes something cognitive in the first place [3].

Making this functionalist commitment explicit is essential for understanding why subsequent debates take their specific form. However, there are importantly different varieties of functionalism, and they yield different verdicts on how liberal the Parity Principle can be. Classical machine functionalism, which individuates mental states by their position in a computational state table — abstracting entirely away from the physical medium that implements those states — permits extension wherever the right abstract relations obtain, regardless of physical medium. Causal-role functionalism, which defines mental states by their typical causal relations to sensory inputs, behavioural outputs, and other mental states within a broader theory, is more permissive about material substrate but ties functional identity more closely to the causal profile of the whole system. Psychofunctionalism, which appeals to the functional roles specified by mature cognitive science rather than by common sense, introduces empirical constraints that may restrict which external processes genuinely qualify [33, 1]. The extended mind hypothesis does not require allegiance to any single variant, but the choice matters: a strict psychofunctionalist might demand that external processes exhibit the same fine-grained computational profile as their internal counterparts [1], while a more liberal causal-role functionalist need only require broad equivalence of input-output relations.

Wheeler [34] has argued explicitly that extended cognition and functionalism are not merely compatible but interdependent — that an extended functionalism treating neural, bodily, and environmental resources together as elements of a single functional organisation provides the most coherent philosophical basis for the extended mind thesis. Clark himself, in addressing tensions between embodied cognition and extended cognition, endorsed this position: the cognitive significance of any component — whether neural tissue, bodily action, or environmental artefact — is determined by the functional role it plays within the overall balance of resources recruited for a cognitive task [16, 5]. On this account, the body’s contribution to cognition is important not because flesh has special metaphysical standing, but because it occupies particular functional positions that happen to be difficult for other substrates to fill — a point that preserves the multiple realisability central to functionalist philosophy [33] while acknowledging the empirical importance of embodiment [15]. Indeed, debates over where precisely to draw the boundary of cognitive systems — what Piredda [6] calls the problem of identifying the “mark of the cognitive” — are themselves downstream of this foundational question about which variety of functionalism one adopts.

The functionalist grounding of the Parity Principle is what gives the internalist critiques their particular bite: the most powerful objections to extended cognition are, at root, objections to the kind of functionalism that underwrites it. Recognising this makes the logical structure of the ensuing debate considerably more transparent.

Internalist Critiques and the Coupling-Constitution Fallacy

The internalist response to extended cognition was swift and technically precise. Early work by [35] challenged whether extended cognition claims had genuine empirical support, arguing that cognitive processes remained, as a matter of scientific fact, bounded by brain and body even if boundary-crossing was conceptually possible. A parallel challenge came from [1], who argued that the embedded cognition hypothesis — which posits that internal cognitive systems are heavily shaped by and oriented toward environmental structures — offers a more parsimonious and empirically tractable alternative to full-blown constitutional extension.

The critique deepened considerably in the subsequent decade. [36] identified what they termed the coupling-constitution fallacy: the logical error of inferring that because an external object is causally coupled to a cognitive agent, it therefore constitutes part of the cognitive system. Adams and Aizawa argued that being connected to something cognitive does not make that thing cognitive; rather, the intrinsic properties of internal processes — including the involvement of non-derived, or original, intentional content — are what individuate genuinely cognitive states [36]. On this view, Otto’s notebook is merely a tool that influences cognition, not a part of it, and Clark’s Parity Principle trades on an equivocation between causal contribution and constitutive membership.

Crucially, the original intentionality criterion that Adams and Aizawa advance represents an explicitly anti-functionalist move: it reinstates a substrate-specific condition on cognition by insisting that only states possessing non-derived content — content whose meaning is intrinsic rather than conventionally assigned — qualify as genuinely cognitive [36, 33]. If this criterion is accepted, then no external artefact whose representational content is derived from human convention (such as notebooks, diagrams, or digital records) can be constitutively cognitive, regardless of its functional role. The debate thus bottoms out in a disagreement about the adequacy of functionalism itself: Clark’s camp holds that functional role is sufficient for cognitive status, while Adams and Aizawa insist that it is not — maintaining that something about the intrinsic character of internal states escapes functional characterisation [5].

However, as [6] has argued, this criterion is less decisive than it first appears. Since every extended cognitive process necessarily includes neural components that do possess non-derived content, the original intentionality requirement can be satisfied by the extended system as a whole without demanding that each constituent element independently exhibit it. A complementary challenge comes from predictive processing frameworks, which propose their own system-level “mark of the cognitive” that sidesteps the derived/non-derived content distinction altogether [14]. The force of the objection thus depends on whether the “mark of the cognitive” must be a property of every component or of the system that those components jointly constitute — a question that remains open but whose stakes become considerably clearer once the functionalist premises are made explicit.

This critique has remained influential precisely because it targets the logical structure of the extended mind argument rather than any particular empirical claim. Clark’s response, developed at length in Supersizing the Mind [5], has been to question whether the internalists’ criterion — original intentionality — is itself theoretically well-grounded, pointing out that no principled non-question-begging account of what makes content “non-derived” has been supplied. No consensus resolution has emerged. The debate thus reveals a deeper disagreement about what makes something cognitive in the first place, a question that has never received a fully satisfying answer from either camp.

Systems-Based Demarcation and Organismic Boundaries

Rupert [1] presented a structurally distinct challenge to the extended view, rejecting it on what he framed as methodological rather than purely logical grounds. He argued that the most empirically productive criterion for demarcating cognitive systems is organism-centred: the integrated, evolved, and developmentally stable organization of biological organisms provides the natural unit for cognitive scientific inquiry. According to this account, extended cognitive systems are neither empirically tractable nor theoretically necessary; the Hypothesis of Extended Cognition (HEC) offers no explanatory advantages over an Hypothesis of Embedded Cognition (HEMC) that keeps cognition inside the organism while acknowledging its profound environmental sensitivity [1]. Crucially, Rupert’s empirical case rests on the observation that organism-internal cognitive systems exhibit a clustered, mutually supporting profile of properties — including cross-task integration, sensitivity to learning, and characteristic error patterns — that external artefacts do not reliably share [1]. This systems-based demarcation placed Rupert in direct opposition to Clark [5]: where Clark saw organism-environment coalitions as genuine cognitive wholes, Rupert saw boundary inflation that sacrifices scientific precision for philosophical permissiveness.

Rupert’s position is philosophically instructive because it does not reject functionalism outright but constrains it. According to his account, the functional organisation relevant to cognition is specifically the organisation of the organism, shaped by evolutionary and developmental history in ways that external artefacts, however useful, do not share [1]. This represents a restricted functionalism in which the causal-role relations that individuate mental states must be anchored in a biologically integrated system. The disagreement with Clark is therefore not over whether functional role matters — both agree that it does — but over the spatial and material scope of the systems within which functional roles are defined [5, 16]. The coupling-constitution distinction is directly implicated here: critics such as Piredda [6] have argued that Clark’s Parity Principle commits a fallacy by inferring constitution from mere causal coupling, a charge that sharpens the demarcation problem Rupert identifies. Slors [4] has further sharpened this point by distinguishing between implementation extension, where artefacts are incorporated into a functionalist cognitive architecture in the manner Clark envisions, and impact extension, where social practices and institutions shape cognition through normative and coordinative mechanisms that resist straightforward functionalist analysis. This distinction suggests that the functionalism underwriting the Parity Principle may be adequate for some forms of cognitive extension (artefact-based) while requiring supplementation or replacement for others (socially constituted).

A related interdisciplinary analysis offered by [37] approached the boundary question from a different angle, arguing that the individual itself — as a unit of analysis — requires reconstruction across philosophy, psychology, and the history of the human sciences. Rather than simply adjudicating between extended and internalist positions, this work underscored that the boundaries of mind are themselves theoretical constructs shaped by disciplinary commitments. This point carries methodological significance for how the empirical debates should be framed.

The Self-Evidencing Brain and Predictive Processing

The predictive processing framework introduced a new dimension to the debate about cognitive boundaries. According to Hohwy [38], the brain functions as a self-evidencing organ when understood as a prediction error minimization engine. This means the brain’s entire operation consists of generating models that account for its own sensory input, which requires a principled evidentiary boundary between the inferring system and the environment it models. Central to Hohwy’s argument is what he calls the “evidentiary-explanatory circle” (EE-circle), whereby a hypothesis and its supporting evidence form a mutually reinforcing loop during Bayesian inference. Within this framework, only brains and sensory organ states properly reside inside the evidentiary boundary—external objects like notebooks are excluded from the prediction error minimization (PEM) machinery by definition [38]. Hohwy contends that the predictive brain cannot literally extend into the world without losing the epistemic structure that makes it function as an inference engine at all. On this view, the boundary is not a philosophical prejudice but a functional necessity [38]. Hohwy’s framework thus provided new internalist ammunition precisely when predictive processing was being enthusiastically recruited by extended mind theorists. Clark, who had himself drawn on predictive processing to support extension [12, 13], was thereby placed in the position of arguing that the same computational framework could support opposite conclusions about cognitive boundaries—a tension that remains incompletely resolved.

At a deeper level, the Clark–Hohwy dispute represents a disagreement about what kind of functionalism predictive processing entails. Hohwy’s position implies that the prediction error minimization architecture defines a functional boundary that is constitutive of the cognitive system. For Hohwy, the evidentiary boundary is itself a functional property, and its integrity is what makes the system cognitive [38]. Clark’s extended reading, by contrast, treats prediction error minimization as a functional process that can recruit external states through precision-weighting mechanisms. This approach allows the functional system to expand beyond the evidentiary boundary without dissolving it [13, 12]. Notably, critics of Hohwy such as Kersten [14] have argued that PEM can instead serve as a positive “mark of the cognitive” for extended systems. According to this view, if external components genuinely participate in algorithmic-level PEM—as illustrated by cases of prosthetic devices that shape proprioceptive inputs and constrain action selection—then no principled internalist boundary follows from the PEM framework itself. Both positions are functionalist in character, but they disagree about whether the functional architecture of prediction inherently demarcates the cognitive system or merely describes computational operations that can be distributed across system boundaries. This framing reveals that the internalism–externalism debate is not a conflict between functionalist and non-functionalist philosophies of mind so much as a dispute about the grain and scope of the functional descriptions that matter.

4E Cognitive Science: Coherence or Cluster?

By the mid-2020s, the proliferation of extended, embodied, enactive, and embedded research programmes prompted a meta-level question: does 4E cognitive science constitute a unified research programme at all? The analysis offered by [2] provides a sobering assessment: the field resists straightforward definition because it lacks any single unifying essence, and existing explanatory strategies—whether definition-based, contrastive, or essentialist—either misrepresent the diversity of actual research or introduce internal incoherence.

This analysis implies that 4E cognitive science may be better understood as a family resemblance cluster in the Wittgensteinian sense: a set of overlapping commitments united by shared opposition to classical computationalism [15, 16] rather than by positive theoretical consensus [2]. This finding complicates the entire extended mind debate, since the arguments of Clark [5], Rupert [1], and Adams and Aizawa—each advancing distinct criteria for what counts as genuinely cognitive—may be directed at targets that are not fully commensurable.

A related pressure point concerns what has been termed the “mark of the cognitive”: the absence of agreed necessary and sufficient conditions for cognitive processes means that boundary disputes over extension, embedding, or enaction inherit a shared foundational instability [14, 6]. This foundational instability is further deepened by the relationship between cognition and behavior explored in [39]: if we lack agreed criteria for what constitutes a cognitive process, debates about the boundaries of such processes inevitably inherit that same instability.

Cognitive Disability and the Real-World Stakes of the Parity Debate

The theoretical debate over the Parity Principle has largely been conducted through stylised thought experiments whose very neatness sometimes obscures whether real-world cases vindicate the theoretical intuitions at stake. It is in the domain of cognitive disability and assistive technology that the Parity Principle confronts its most consequential empirical tests. The populations who depend most heavily on external cognitive scaffolding are also those for whom the distinction between genuine extension and mere tool use carries immediate implications for selfhood, legal personhood, and the ethics of care.

Three bodies of evidence are particularly instructive here. First, participatory design research by [40] demonstrates that physical artefacts used by people living with dementia — familiar objects, postcards, spatially arranged environments — do not merely prompt recall but function as genuine mediators of agency. These artefacts enable intentional social action that other interfaces cannot reliably elicit. Crucially, they satisfy the Parity Principle’s core criteria as originally formulated by [3]: they are reliably accessed, habitually trusted, and functionally integrated into the cognitive routines that sustain the person’s engagement with the world.

Evidence on augmentation-related brain plasticity reinforces this interpretation from a different direction. Prolonged use of prosthetic and assistive devices produces measurable cortical remapping in sensory, motor, and frontoparietal networks. For example, intraneural implants in amputees reverse aberrant reorganisation and restore representational maps of the missing limb [41]. When the organism’s own functional architecture restructures itself to incorporate external artefacts, Rupert’s organism-centred demarcation [1] becomes empirically harder to sustain.

These considerations find their most systematic legal and philosophical elaboration in [21], whose concept of empersonification identifies the threshold beyond which a durably integrated device ceases to be an external tool and becomes a cognitive constituent of the person. This transformation carries concrete consequences for how damage to such devices should be classified in law. Whether any given technology actually reaches that threshold, and what ethical implications follow if it does, are questions evaluated in detail by [42].

Section 5 develops these cases in full mechanistic and ethical detail, examining the evidence on dementia artefacts, prosthetic plasticity, and empersonification within the broader framework of cognitive disability and moral status. The purpose here is simply to establish that disability contexts are not peripheral illustrations of the Parity debate but its most empirically consequential arena — the domain where abstract theoretical distinctions between coupling and constitution [6] acquire urgent practical stakes.

Gaps and Outstanding Questions

Across this literature, several limitations stand out. First, despite decades of argument, no agreed-upon empirical criteria have emerged for distinguishing genuine cognitive extension from mere causal coupling—the very question Adams and Aizawa posed in 2001 remains open [35, 36], though proposals grounded in prediction error minimization [14], neo-mechanical constitutive relevance [43], and graduated integration thresholds drawn from disability and assistive technology research [42, 21] represent promising avenues. Second, as [2] makes clear, the definitional instability of the broader 4E programme undermines the clarity of the debate’s central terms. Third, the relationship between phenomenal consciousness and extended cognition—briefly gestured at in Clark and Chalmers’ original treatment of extended belief [3]—has not been adequately theorized: it remains unclear whether extension claims that are defensible for representational states carry over to conscious experience. Fourth, both internalist and externalist positions have developed almost exclusively within Western analytic philosophy, with little engagement with non-Western philosophical traditions that have long theorized porous or relational boundaries between mind and world. Fifth, and perhaps most fundamentally, the functionalist commitment that underwrites the Parity Principle is rarely subjected to direct scrutiny within the extended mind debate itself: participants typically argue about the consequences of functionalism for cognitive boundaries without stepping back to ask whether functionalism — in any of its variants — provides the right framework for individuating the mental in the first place [33, 34]. These gaps suggest that the foundational debate, far from being exhausted, continues to open onto uncharted territory.

4. Distributed Cognition, External Representations, Cognitive Offloading, and Interactivity

Cognition, as studied through the lens of distributed systems, extends well beyond the boundaries of individual minds into the social, material, and technological environments that humans inhabit and transform. This section traces the theoretical and empirical development of distributed cognition research, from its foundational programmatic statements through experimental studies of cognitive offloading, gesture, and interactivity, to the most recent philosophical challenges posed by extended cognition to classical action theory. Together, these contributions share a commitment to expanding the unit of cognitive analysis and examining the mechanisms by which external resources are recruited into cognitive work.

Theoretical Foundations: Sociotechnical Systems and the Distribution of Mind

Early work in this tradition established the conceptual architecture that subsequent research would build upon and contest. Hollan, Hutchins, and Kirsh argued that cognitive processes are not confined within individual skulls but are distributed across people, artifacts, and environments in sociotechnical systems [10]. Their framework treated social organization itself as a form of cognitive architecture — a structure that shapes how information flows across groups and how errors propagate or are caught. Cultural artifacts, on this account, are not merely convenient tools but accumulated cognitive resources that encode solutions to recurring problems and permit forms of reasoning that would be impossible without them [10, 7]. Kirsh’s subsequent work on external representations reinforced this point, demonstrating empirically that physical and symbolic structures in the environment actively transform the nature of cognitive tasks rather than simply supplementing internal processes [7]. This was a genuinely radical claim: it implied that the appropriate unit of cognitive analysis is not the individual reasoner but the entire system within which reasoning takes place — a position that finds further elaboration in Clark’s account of the mind as constitutively extended into its material and social surroundings [5]. Slors has more recently argued that such cognitive extension scales beyond individual artifact-use to encompass institutions and notebooks as sites of genuinely symbiotic cognition [4].

Malone and Crowston’s coordination theory offered a complementary and more formally tractable perspective on distributed cognitive work [44]. Defining coordination as “managing dependencies between activities,” they articulated a framework applicable across human organizations, computational systems, and biological systems alike. Critically, they identified how information technology reduces coordination costs and thereby transforms organizational structure — increasing overall coordination activity while substituting automated processes for human coordinators [44]. Although originating in organizational science rather than cognitive psychology, coordination theory anticipates central questions about how cognitive labor is partitioned and integrated across hybrid human-computational systems.

External Representations: Transformation, Not Substitution

Where early distributed cognition theory established the importance of artifacts in principle, Kirsh’s theoretical analysis sharpened the mechanism [7]. The crucial move was distinguishing between external representations as mere memory aids and external representations as genuinely transformative cognitive tools. Externalizing mental structures, Kirsh argued, converts them into shareable, stable, and manipulable objects of thought that enable people to reach cognitive states that would otherwise remain inaccessible [7]. Physical interaction with an external representation does not simply substitute for internal processing; it reshapes the cognitive terrain itself. A diagram does not just record a solution; it affords discovery of structure that the unaided mind would not perceive. Steffensen et al.’s fine-grained case analyses illustrate this vividly: participants solving insight problems achieved breakthroughs not through pre-planned internal reasoning, but through serendipitous material interactivity — physical manipulation of external objects that generated affordances invisible to purely mental effort [9]. This distinction between offloading and transformation remains insufficiently integrated into experimental paradigms, most of which operationalize external representation use as a performance-enhancing memory strategy [8, 45] rather than as a cognitive enabler of qualitatively different thought [7].

Cognitive Offloading: Strategy, Habit, and Metacognitive Control

The experimental study of cognitive offloading brought distributed cognition hypotheses into the laboratory, but with an important narrowing of scope. Risko and Gilbert defined cognitive offloading as a strategy whereby individuals externalize cognitive processes to reduce mental effort, reviewing evidence for how people use external resources and environmental structures to support cognition [46]. This framing positions offloading as a metacognitive act — a deliberate choice made on the basis of cost-benefit reasoning about internal versus external processing. Gilbert’s related work developed this metacognitive account further, emphasizing that strategy selection is itself a cognitive achievement that depends on accurate self-monitoring and environmental sensitivity [8]. More recent philosophical analysis has extended this picture, examining how cognitive offloading relates to the causal structure of intentional human action and raising questions about the degree to which externalized processes remain genuinely agentive [45].

Empirical findings from Risko and Dunn’s memory span study complicated this picture, however [8]. Spontaneous writing increased predictably with memory load and did provide performance benefits at larger set sizes. Yet nearly half of participants chose to write even when remembering only two letters — a set size where writing provided no measurable performance benefit. This pattern suggests that offloading behavior is not always strategically optimal; it may become habitual, deployed even when the cognitive cost-benefit analysis does not favor it [8]. A tension therefore exists between the metacognitive strategy account and the possibility that offloading behavior is over-extended, a kind of cognitive default that persists beyond its domain of utility.

Learning sciences research has brought this tension into sharper focus by identifying a distinctive risk of cognitive offloading in educational contexts: the elimination of desirable difficulties — the productive cognitive struggle through which durable learning is constructed. Systematic research on writing-to-learn interventions demonstrates that structured journaling supported by metacognitive prompts produces learning effects approximately double those found in broader meta-analyses of unprompted writing interventions (Hedges’s g = 0.78), with the crucial finding that the sequencing of prompts — metacognitive monitoring strategies before cognitive elaboration strategies — significantly moderates outcomes [47]. Moreover, the same programme of research establishes an expertise reversal effect: scaffolding that initially benefits novice learners becomes counterproductive as expertise develops, requiring adaptive fading of support rather than permanent provision [47, 29]. For novices, structured scaffolding reduces unproductive load and directs attention toward relevant features of the task; for more expert learners, the same scaffolding imposes redundant processing that competes with, rather than supports, the more efficient cognitive schemas they have already constructed [29]. The implication for cognitive offloading theory is significant: the cost-benefit framework that treats offloading as a rational response to cognitive demand must be supplemented by a developmental account in which the consequences of offloading depend not only on immediate task performance but on the long-term construction of the knowledge structures that enable expertise [20]. Offloading that enhances performance while bypassing effortful encoding may, over time, prevent the very skill development it appears to facilitate — a dynamic that will bear directly on the analysis of AI-mediated learning tools taken up in Section 6.

Philosophical Complications: Davidson, Action Theory, and Extended Cognition

The most recent contribution to this cluster, Britten-Neish’s 2025 philosophical analysis, reveals a fundamental tension between the theory of action that implicitly underpins much cognitive offloading research and the extended cognition hypothesis that motivates the distributed cognition program more broadly [45]. Davidson’s “standard story” of action, with its Strict Bodily Constraint — the requirement that intentional actions consist only in movements of an agent’s body — is directly incompatible with the Hypothesis of Extended Cognition (HEC), which holds that some cognition inherently involves bodily and environmental resources in a constitutive rather than merely causal relation [45, 3]. Originally articulated by Clark and Chalmers through the contrast between Inga, who relies on biological memory, and Otto, who uses a notebook as a cognitive prosthetic, HEC treats external resources as genuine constituents of cognitive systems when they are reliably coupled with an agent’s mental activity [3, 5]. When participants in offloading studies write information down or arrange objects to support memory, the standard story treats this as a bodily action that causally supports cognition without being part of it. Extended cognition proponents would resist this framing, arguing that causal coupling with environmental resources legitimately grounds constitutive cognitive relations once methodological individualism is relinquished [6]. Whether Davidson’s action theory can be revised to accommodate extended cognitive processes, or whether it must be replaced, represents an unresolved and consequential philosophical question for the field [45]. This is not merely an academic dispute: it determines whether cognitive offloading constitutes genuine cognitive extension or only causal scaffolding — and therefore what explanatory work distributed cognition theories are doing [1, 6].

Gesture and Interactivity: Embodied Cognition in Action

Alongside these theoretical debates, empirical work on gesture and problem-solving interactivity has provided rich evidence for how external, bodily, and material resources are recruited under cognitive pressure. Pouw and colleagues synthesized evidence demonstrating that gesture arises most frequently during high cognitive demand, particularly in visuospatial processing, novel tasks, and when working memory is taxed [48]. Crucially, individuals with lower spatial working memory capacity and weaker spatial transformation ability produce gestures more frequently and are more negatively affected when gesturing is constrained [48]. This establishes gesture not merely as communicative behavior but as an embedded and potentially extended cognitive resource — one whose function is most salient precisely when internal resources are most limited. Complementing this, cognitive offloading research has similarly shown that agents strategically recruit the environment to reduce internal memory and processing demands, a pattern that holds across both action-based and representational tasks [45, 8].

Vallée-Tourangeau and colleagues contributed a parallel finding from a different methodological tradition [9]. Using Cognitive Event Analysis of participants solving the 17 Animals problem, they demonstrated that successful problem-solving transitions — from arithmetic to set-theoretical frameworks — occurred through serendipitous material interactions rather than pre-conceived mental plans. Aesthetic actions, behaviors that imposed visual order on materials without explicit task purpose, functioned as cognitive motors that inadvertently revealed task-relevant affordances [9]. The implication is that physical interaction with the environment does not merely implement decisions already made internally; it actively generates the cognitive events from which insight emerges. This resonates strongly with Kirsh’s account of how external representations transform and simplify cognitive problems by restructuring the informational environment [7], and with Hutchins’s distributed cognition framework, but extends them with fine-grained qualitative evidence about the moment-by-moment dynamics of cognitive discovery [10]. Taken together, these findings suggest that the boundaries of productive cognitive work are drawn not at the skull, but at the edges of the agent–environment system as a whole [5].

Creative Collaboration as Distributed Cognitive System

The analysis of Elizaveta Svilova’s film editing collaboration with Dziga Vertov applies this distributed cognition framework to a creative domain, examining how expertise, embodiment, and material inscription practices constitute a cognitive system larger than any individual contributor [31]. Drawing on Clark and Chalmers’ extended mind framework [3, 5], Pearlman, Mackay, and Sutton argue that editing requires kinesthetic imagination and embodied expertise rather than purely intellectual decision-making, and that creativity in this case emerges from the integrated functioning of brain, body, and world [31]. The role of material inscription — the physical manipulation of film strips, annotations, and editing tables — functions analogously to the representational media studied in distributed cognition accounts of expert professional practice [10, 7]. This case study demonstrates distributed cognition’s applicability to expert creative work, though it also illustrates a persistent gap: the theoretical richness of distributed cognition accounts of real-world practice tends to outrun their experimental operationalization [1].

Distributed Cognition in Educational Settings: Evidence from CSCL

The distributed cognition framework receives some of its most systematic empirical support from computer-supported collaborative learning (CSCL) research, where the distribution of cognition across learners, artefacts, and digital platforms is not merely observed but deliberately designed. This body of work provides ecologically valid evidence for the theoretical claims advanced by Hollan, Hutchins, and Kirsh [10] while simultaneously revealing important constraints on when and how distributed cognitive systems produce genuinely productive outcomes.

A foundational empirical contribution is the tripartite distinction between knowledge-sharing, knowledge-construction, and knowledge-creation discourse [11]. Knowledge-sharing involves the exchange of pre-formed beliefs; knowledge-construction entails collaborative refinement of ideas; knowledge-creation, the most cognitively demanding mode, involves sustained, community-driven advancement of ideas as epistemic objects. The roots of this framework lie in Scardamalia and Bereiter’s foundational work on knowledge-building communities, which first articulated the vision of classrooms as communities whose collective goal is the advancement of shared knowledge rather than the acquisition of individual understanding [49]. In a design experiment with secondary students using a collaborative knowledge-building platform, the highest-performing group consistently outperformed peers on measures of knowledge quality, with discourse analysis revealing that idea-centred and genuine inquiry moves — rather than volume of participation — distinguished knowledge-creation from lesser forms of collaborative discourse [11]. Earlier empirical work with ninth-grade students had already demonstrated that classes using collaborative platforms alongside explicit knowledge-building principles and portfolio assessment outperformed peers on measures of inquiry quality and conceptual understanding, with portfolio scores measuring collective inquiry contributing significant variance in understanding beyond individual-level predictors [50]. Together, these studies establish that knowledge-creation discourse is neither automatic nor common, and that its cultivation depends on intentional alignment between tools, principles, and assessment practices.

The conditions under which productive collaborative cognition emerges have been further specified through research on collaboration scripts. Early empirical demonstrations showed that structured scripts can meaningfully elevate the quality of argumentative discourse during peer interaction, scaffolding participation in forms of reasoning that learners do not naturally perform well without guidance [51]. Meta-analytic synthesis of 22 studies comprising 34 comparisons subsequently established that collaboration scripts in CSCL environments produce a small but meaningful effect on domain-specific knowledge acquisition (d = 0.20) and a large effect on collaboration skills (d = 0.95), with a theoretically consequential moderator: only scripts that prompted transactive activities — contributions in which learners explicitly build upon, extend, or challenge one another’s reasoning — produced reliable benefits for domain learning [30]. Scripts lacking transactive prompts showed no advantage over unstructured collaboration, reframing the design question from whether to scaffold collaborative cognition to how to scaffold it in ways that generate genuine epistemic engagement rather than procedural compliance. More recent theoretical work has further argued that such epistemic scripts must be understood not merely as procedural scaffolds but as structuring mechanisms that shape the very character of collaborative inquiry — organising which knowledge moves are available, valued, and sequenced within a given collaborative system [52].

The cognitive architecture underlying these collaborative effects has been formalised through Collaborative Cognitive Load Theory (CCLT), which extends classical cognitive load theory to group settings [29]. CCLT proposes that collaboration’s benefit is conditional on a favourable balance between two competing forces: a collective working memory effect, in which the combined cognitive resources of group members exceed what any individual could bring to bear on a complex task, and transactive costs — the real cognitive expenditure required for coordination, communication, and mutual monitoring [29]. When transactive costs are high relative to task complexity, collaboration can impair rather than enhance learning — a finding that explains the longstanding empirical inconsistency of collaboration’s benefits and that connects directly to the metacognitive offloading literature: the challenge in both individual and collaborative settings is to calibrate external support so that it reduces unproductive load without eliminating the effortful processing that consolidates learning. CCLT further invokes the expertise reversal effect — the same principle identified in individual scaffolding research [47] — to argue that collaborative structures appropriate for novice groups may impose unnecessary overhead on more experienced ones. For novices, structured collaboration scripts reduce the search space of possible epistemic moves; for experienced learners, those same structures may fragment the more fluid and integrated cognitive strategies they have developed, imposing coordination costs without compensating benefits. This prediction has direct implications for adaptive design of collaborative learning environments, and — as will be examined in Section 6 — it bears with particular force on AI-mediated learning contexts, where the question of how automated support should be calibrated to learner expertise is both theoretically central and practically urgent [29, 30].

Empirical work has further demonstrated that task design independently shapes the regulatory and knowledge-construction dynamics of collaborative groups. A mixed-methods study of secondary students collaborating across two differently structured science tasks found that hands-on construction tasks prompted more frequent socioemotional interaction and emotion-motivation regulation, while conceptually oriented tasks fostered sustained cognitive engagement and deeper meaning-making [53]. Emotion-motivation regulation emerged as the most frequently interconnected regulatory mode across both conditions, suggesting that affective processes are not peripheral to distributed cognition but structurally central to how collaborative cognitive systems function [53]. This finding resonates with broader neuroscientific arguments that emotion and cognition are functionally inseparable — that affect is not a disruptive force on rational processing but a constitutive component of the learning process itself [54]. Taken together, these results extend the distributed cognition framework by demonstrating that the distribution of cognitive labour is shaped not only by artefacts and representations but by the emotional and motivational dynamics that task environments elicit — a dimension that Kirsh’s transformative representation account [7] and the foundational distributed cognition framework [10] did not explicitly theorise.

Visualisation tools provide a further empirical illustration of how external representations function differently depending on their specificity. In controlled experiments with educational psychology students working in dyads, content-specific visualisation tools — as opposed to generic or no-tool conditions — led groups to reference more appropriate domain concepts, engage in more genuine intellectual conflict rather than superficial agreement, and produce collaborative solutions of significantly higher quality [55]. The content-specificity of the tool was consequential: generic visualisation did not produce equivalent benefits, confirming Kirsh’s theoretical claim that the representational affordances of a tool must align with the epistemic demands of a domain to elevate discourse quality.

Taken together, this educational evidence strengthens the distributed cognition programme in two respects. First, it provides ecologically valid, classroom-based demonstrations that cognitive processes are genuinely restructured — not merely supplemented — when distributed across collaborative systems, confirming in educational settings what Kirsh’s theoretical analysis predicted for external representations more broadly [7]. Second, it reveals that productive cognitive distribution is a design achievement, not an automatic consequence of providing collaborative tools or social configurations — a finding with significant implications for the broader extended mind debate, where the conditions under which coupling becomes constitutive remain contested.

Gaps and Outstanding Questions

Several important gaps emerge across this body of work. The experimental cognitive offloading literature predominantly employs simple memory tasks, leaving open how offloading dynamics operate in the complex, multi-step, ecologically valid cognitive work that distributed cognition theory was originally developed to explain [46, 8]. Beyond working memory capacity, individual differences in offloading propensity remain poorly characterized; traits such as impulsivity and related dispositional factors have begun to attract attention in adjacent literatures [56], but systematic mapping onto offloading behaviour is absent. The integration between coordination theory and cognitive offloading paradigms — which share obvious conceptual territory in questions about task decomposition and dependency management — has not been systematically pursued [44, 46]. Cross-cultural variation in external representation use and offloading strategies has received almost no empirical attention, despite the emphasis in foundational theory on cultural artifacts as accumulated cognitive resources [10, 7].

The learning sciences evidence, while providing crucial ecological validity, introduces its own gaps. Longitudinal studies tracking whether collaboratively distributed cognitive gains transfer to individual competence remain scarce, with existing work focusing primarily on in-situ group performance rather than post-collaborative individual outcomes [57, 50]. Furthermore, the interaction between script design, task structure, and learner expertise has not been systematically modelled despite clear theoretical predictions from Collaborative Cognitive Load Theory that these variables should interact [29, 30]. Finally, the philosophical question posed most sharply by recent work — whether offloading constitutes genuine cognitive extension or merely causal scaffolding — awaits an answer that will require both conceptual precision and empirical ingenuity. Current debate over the coupling-constitution fallacy [6, 1] and proposals for intermediate positions such as symbiotic cognition [4] indicate that the field has not yet converged on the right theoretical vocabulary, let alone the right evidence [45].

5. Memory, Embodiment, Predictive Processing, and Evolutionary Perspectives on Brain-Body-World Architecture

The architecture through which brains, bodies, and environments couple to produce cognition has become one of the most contested and productive areas of cognitive science. Rather than treating cognition as a property of isolated neural circuits, a growing body of theoretical and empirical work situates mental processes within dynamic loops that recruit bodily systems, artifacts, social partners, and evolutionary history. The papers examined here share a concern with how such coupling is enabled, constrained, or transformed — whether approached through computational neuroscience, memory taxonomy, cognitive archaeology, or social cognition. Critically, these accounts do not converge on a single picture: they generate productive tensions that are reshaping the field’s understanding of where cognition begins and ends.

Predictive Processing and the Clark–Hohwy Dispute in the Memory Context

As detailed in Section 3’s treatment of the self-evidencing brain, Clark and Hohwy reach diametrically opposed conclusions from the same predictive processing framework: Clark’s precision-weighting account holds that external resources can be dynamically recruited as quasi-internal components of the predictive hierarchy [13], while Hohwy argues that the brain’s need to maintain a stable evidentiary boundary between its generative models and their sensory evidence constrains cognition to the skull [38]. This internalist constraint has been contested on formal grounds, with Kersten [14] arguing that predictive processing does not, by itself, supply a principled “mark of the cognitive” that would exclude external vehicles from constituting genuine cognitive states. The foundational computational architecture motivating both positions — hierarchical bidirectional prediction and error propagation established by [12] — is not rehearsed here. What Section 3’s treatment did not develop, however, is the specific implication of this dispute for memory: if precision-weighting allows external memory systems to be incorporated as functional components of the predictive hierarchy, as Clark argues [13], then the boundary between neural and technological memory may be permeable not only philosophically but computationally. Empirical work on cognitive offloading lends indirect support to this permeability thesis, demonstrating that agents strategically redistribute memory load onto environmental structures in ways that alter the causal profile of subsequent recall [8, 45]. This memory-specific corollary becomes consequential when examined alongside the expanded memory taxonomy considered below [58]. The Clark–Hohwy debate at the level of memory systems remains empirically unresolved [12, 13, 38, 14] and, as will be argued, its resolution may require precisely the kind of longitudinal evidence that current memory research has yet to produce.

Expanded Memory Taxonomies and the Memory Symbiosis Framework

One domain where the question of genuine extension versus mere information storage becomes empirically tractable is memory research. Early cognitive psychology worked with a taxonomy largely confined to individual, brain-based systems — episodic, semantic, procedural, and working memory — a framework codified across decades of research into biological memory architecture, including influential work on the strict capacity constraints of working memory [59]. More recent theoretical work has argued that this taxonomy is fundamentally incomplete. [58] proposes a four-stage expansion: external memory (both social and technological), biological memory systems distributed throughout the body (including immune, genetic, and dermal memory), collective memory at the group level, and natural external memory embedded in environmental structures. This taxonomy is not merely additive; it is grounded in the claim that these systems are functionally continuous with neural memory in their role of storing, retrieving, and transmitting information that shapes adaptive behavior.

Central to [58]‘s account is the Memory Symbiosis Framework, which characterises internal and technological external memory as complementary and interdependent rather than redundant. Each system has distinct strengths — internal memory is flexible, context-sensitive, and associatively rich, while technological memory is stable, high-capacity, and shareable — and the two systems co-evolve in use. Empirical work on cognitive offloading lends support to this picture: individuals strategically redistribute memory load onto environmental structures when doing so is advantageous, suggesting that the boundary between internal and external storage is actively and flexibly managed rather than fixed [8]. External representations more broadly have been shown to restructure the cognitive tasks agents perform, rather than merely duplicating information already held internally [7]. This framing has direct implications for the Clark–Hohwy debate introduced above: if technological memory and neural memory form a genuinely symbiotic system, then the evidentiary boundary Hohwy posits may be permeable in practice even if maintained in principle — a tension Clark himself identifies in arguing that neural and non-neural resources can form deeply integrated cognitive systems [5, 3]. However, [58] acknowledges that the Memory Symbiosis Framework currently lacks longitudinal empirical support; how internal memory systems change as technological memory becomes more heavily relied upon remains an open question.

Dementia, Prosthetic Memory, and the Moral Significance of Extended Cognitive Systems

The Memory Symbiosis Framework acquires its most ethically urgent application in the domain of dementia and cognitive decline, where the symbiosis between internal and external memory is not a theoretical abstraction but a daily necessity. When biological memory deteriorates progressively, external scaffolding does not merely supplement what remains — it increasingly constitutes the functional memory system through which the person maintains orientation, exercises preference, and sustains social participation.

Participatory design research with people living with dementia demonstrates this dynamic with unusual empirical clarity. [40] found that physical artefacts — postcards, scrapbooks, spatially arranged objects — functioned not as passive memory cues but as active mediators of agency: residents in a care facility used these materials to express sharing intentions, indicate choices, and engage in social interaction that screen-based digital interfaces could not reliably support. Crucially, Lazar, Edasis, and Piper theorised agency in this context not as an intrinsic individual property but as a sociomaterial achievement distributed across persons, objects, and environments — a finding that extends the Memory Symbiosis Framework into the domain of identity and selfhood [40].

This reframing carries profound implications for how we understand the relationship between memory and personal identity. On traditional internalist accounts, memory loss in dementia is a subtraction from a pre-given self — an erosion of what was once stored within. Such accounts, rooted in psychological continuity theories of personal identity, treat the self as constituted by chains of internal memory states, such that their disruption straightforwardly diminishes the person [58]. On the extended account supported by this evidence, however, the self was always constituted in part through its material and social scaffolding [3], and dementia renders visible the extent of that constitutive dependence. Critically, this is not merely a philosophical dispute about the boundaries of cognition: the coupling-constitution distinction bears directly on the normative stakes [6]. The implication is that disrupting a person with dementia’s access to habitual cognitive artefacts — removing a familiar memory book, relocating a person from a structured environment they have learned to navigate, or replacing trusted physical objects with unfamiliar digital interfaces — constitutes a meaningful harm to the person, not merely an inconvenience to their routine. Heinrichs [42] makes the broader point that situated cognition theses carry direct ethical entailments: if external resources genuinely constitute rather than merely cause cognitive states, then interference with those resources is interference with the cognitive subject, not merely with their tools.

The concept of empersonification, introduced by [21] and discussed in Section 3 in the context of brain-computer interfaces and disability, acquires a further dimension here that Section 3’s treatment did not develop: its application to memory and identity continuity in dementia care. Where Section 3 examined empersonification primarily as a framework for device-person integration in contexts of motor or sensory augmentation, the dementia context foregrounds a different and starker set of stakes — not the extension of existing capacities but the prosthetic maintenance of selfhood itself. If a cognitive device becomes sufficiently integrated with a person’s mental processes to cross the threshold of empersonification, then its removal or disruption is not a property matter but an assault on the person’s identity and continuity [40, 21]. In dementia care, this framework suggests that memory aids, environmental scaffolds, and digital companions durably incorporated into a person’s cognitive routines may warrant protections extending beyond those accorded to ordinary personal property — a conclusion with significant implications for care ethics, institutional policy, and legal personhood.

Neuroscientific evidence on augmentation-related brain plasticity, reviewed in Section 3 in the context of sensorimotor prosthetics, provides convergent support for the constitutive reading developed here. The brain’s demonstrated capacity to reorganise its representational maps in response to sustained prosthetic use — including cortical remapping in amputees fitted with intraneural multielectrode implants, where representations of the missing hand were restored and normalised [41] — implies that the boundary between internal and external memory is not merely philosophically porous but neurobiologically permeable. The Memory Symbiosis Framework proposed by [58] and the plasticity evidence documented by [41] converge on the same conclusion: internal and external memory systems are not independent modules that happen to cooperate but deeply interdependent components of a single functional architecture whose boundaries shift with experience, learning, and — in the case of dementia — progressive neurodegeneration. Slors [4] extends this point from the individual to the institutional level, arguing that symbiotic cognition is not merely a dyadic relation between a person and an artefact but is embedded in and enabled by broader social and institutional arrangements — an observation particularly salient in the care home context, where institutional policies govern access to the very objects that constitute a resident’s functional memory.

Emerging experimental work extends these concerns into the anticipatory register, asking how people who do not yet have dementia imagine and prepare for the possibility of cognitive decline. [60] deployed chatbots as instruments for studying intertemporal self-reflection about ageing, manipulating whether the chatbot was framed as the participant’s future self or as a stranger, and whether it simulated dementia symptoms or presented as neurotypical. Framing the chatbot as a future self with dementia symptoms produced stronger emotional responses and deeper personal reflection, though it did not shift generalised attitudes toward dementia [60]. This dissociation between emotional engagement and attitudinal change is theoretically instructive: it suggests that the psychological mechanisms linking empathic identification with a future self to broader normative commitments are more complex than simple perspective-taking models predict, and that designing technologies to foster dementia awareness requires attending to the specific pathways through which self-relevant emotion translates — or fails to translate — into sustained moral concern.

Taken together, these findings transform the relationship between memory research and identity theory. Memory, on the account emerging from this literature, is not a fixed biological storehouse that external aids merely supplement; it is a dynamic, hybrid system whose boundaries are defined functionally rather than anatomically [3, 5], and whose disruption — whether by neurodegeneration or by the removal of external scaffolding — constitutes an assault on the conditions of selfhood itself.

Cognitive Archaeology: Evolutionary Architecture of Tool Integration

A complementary evolutionary perspective is provided by cognitive archaeology, which asks how the brain-body-world coupling visible in contemporary cognition was assembled over deep time. [17] argues that human cognition emerged from dynamic interactions between brain, body, and environment rather than from neural computation alone, with the expanded human parietal cortex playing a central functional role. Parietal expansion in the human lineage is linked to what Bruner terms “prosthetic capacity” — the ability to integrate tools into body schema and cognitive processes, effectively extending the boundaries of the bodily self. This account provides an evolutionary mechanism for the kind of extension that Clark’s precision-weighting framework describes computationally [12, 13]: the parietal cortex may be the neural substrate through which external artifacts are recruited as quasi-bodily resources [17]. This hypothesis is lent direct causal support by electrical brain stimulation (EBS) research, which has established that the parietal lobe — and the inferior parietal lobule in particular — is the cortical region most comprehensively linked to bodily self disturbances, capable of evoking alterations across all five phenomenal components of the bodily self including self-location, agency, body ownership, and perceptual body image [61]. The augmentation-related plasticity evidence reviewed by [41] — including the brain’s capacity to maintain coherent ownership representations for supernumerary limbs and to incorporate entirely novel sensory modalities through cross-modal reorganisation — provides contemporary empirical confirmation that the prosthetic capacity Bruner identifies in the evolutionary record remains dynamically operative in living brains.

This interpretation, however, sits in tension with accounts that emphasize external scaffolding as the primary driver of cognitive change rather than neural specialization. [17] is careful to frame parietal expansion and external tool use as co-evolving rather than causally ordered, but the archaeological evidence is largely correlational. Causal mechanisms linking tool use to neural reorganization — and specifically the direction of that causation — remain poorly specified, a limitation that [17] explicitly acknowledges as a priority for future research.

Social Metacognition and Abstract Concept Formation

A further dimension of brain-body-world architecture concerns the social routes through which abstract concepts are formed and stabilized. [62] argues that abstract concepts differ fundamentally from concrete ones in their dependence on social interaction, emotional resonance, interoceptive grounding, and linguistic scaffolding. Particularly important is the role of social metacognition — the capacity to monitor one’s own conceptual states and to align them interactively with those of social partners — in acquiring concepts whose meanings are vague and socially negotiable. This account implies that many abstract concepts are genuinely distributed across individuals and interactions rather than stored in any single mind [62].

The relationship between this account and the predictive processing framework is suggestive but unexplored. If precision-weightings govern the recruitment of external resources, then the social partners and interactions that [62] identifies as constitutive of abstract concept formation might themselves function as precision-weighted inputs to the predictive hierarchy. How social metacognition interacts with predictive processing during concept acquisition remains a significant gap in both literatures. Similarly, the expanded memory taxonomy proposed by [58] could accommodate social and collective memory as nodes in the symbiotic system, but integration between these frameworks has not yet been attempted.

Synthesis and Outstanding Tensions

Taken together, these accounts converge on a view of cognition as fundamentally distributed across brain, body, artifacts, evolutionary history, and social interaction [10, 15], while disagreeing sharply about the mechanisms, boundaries, and evolutionary origins of this distribution [1, 16]. The predictive processing framework provides the most developed computational account but remains internally divided over whether its logic supports or forecloses genuine extension [12, 13, 38]. Memory taxonomy work extends the relevant systems well beyond the brain but needs longitudinal empirical support [58]. Cognitive archaeology offers evolutionary grounding but struggles with correlation-causation problems [17]. Social cognition research identifies underappreciated routes to concept formation that neither predictive processing nor memory science has yet incorporated [62] — though recent work attempting to embed social actors within active inference frameworks represents a promising step toward closing this gap [63]. And dementia research reveals that the stakes of these theoretical disputes are not merely academic: the moral significance of extended cognitive systems becomes acutely visible when biological memory fails and the self depends on scaffolding for its very continuity [40, 21]. The integration of these perspectives into a unified account of brain-body-world architecture — one that can accommodate computational, evolutionary, social, and clinical dimensions simultaneously — remains the central unfinished task of this research cluster [2, 4].

6. AI, Generative Technologies, and the Extended Mind

The relationship between artificial intelligence and human cognition has become one of the most contested domains in contemporary philosophy of mind, cognitive science, and AI ethics. What began as a relatively settled debate about whether external tools could constitute genuine cognitive extensions has been dramatically reopened by the advent of large language models and generative AI systems — technologies whose interactive, generative, and adaptive qualities mark a qualitative departure from the notebooks, calculators, and search engines that earlier theorists had in mind. The papers gathered under this theme represent a convergence of philosophical analysis, conceptual framework development, and computational architecture design, united by the question of whether AI extends, diminishes, or fundamentally transforms human cognitive capacities. Far from resolving into consensus, this body of work reveals deep disagreement about the basic character of human-AI cognitive integration, with significant implications for education, ethics, and the design of AI systems themselves.

Generative AI as Cognitive Extension: The Continuity Argument

The foundational theoretical position in this cluster is articulated by Andy Clark, whose 2025 philosophical analysis extends his longstanding extended mind thesis explicitly to generative AI [18]. The extended mind thesis — first advanced by Clark and Chalmers [3] and elaborated across subsequent work [5] — holds that human minds have never been purely biological: from gesture and spoken language to writing, print, and digital search, cognitive work has always been distributed across biological and non-biological substrates. On this view, generative AI represents not a rupture but the latest chapter in a long history of hybrid cognitive systems, and techno-pessimism about AI-induced cognitive diminishment fundamentally misidentifies the nature of human cognition by treating the biological brain as the privileged locus of genuine thought [18]. Drawing on predictive processing frameworks [12, 13], Clark argues that the brain is not a passive receiver of external information but an active anticipatory system that continuously integrates internal and external resources — a characterisation that, if correct, makes the boundary between mind and AI tool philosophically unstable. Predictive processing accounts — in which the brain generates top-down probabilistic predictions and updates them against incoming sensory signals — naturalise the use of external scaffolds as part of the cognitive process itself [14].

This continuity thesis sets the stage for the System 0 framework developed by Chiriatti, Bergamaschi Ganapini, and Panai [19], which offers a more granular mechanistic account of how AI integrates into human reasoning. Building on Kahneman’s dual-process theory [64], which distinguishes fast, automatic System 1 from slow, deliberative System 2 reasoning, they propose that AI now functions as a prior stage — System 0 — a cognitive preprocessor that shapes the informational landscape before conscious thought is engaged [19]. This is a significant theoretical innovation: rather than being simply a tool that System 2 consults, AI is positioned as structuring the very inputs to human reasoning, potentially bypassing deliberative engagement altogether. The empirical implication is striking — their framework predicts, and finds support for the claim, that human-AI collaboration can enhance creativity precisely because System 0 enriches associative inputs, but may simultaneously degrade analytical decision-making by reducing the cognitive demands that sharpen deliberative capacities [19].

The Hollowed Mind: Cognitive Sovereignty and Expertise Duality

The continuity and System 0 frameworks are directly and forcefully challenged by Klein and Klein, whose synthesis of dual-process theory, cognitive load theory [29], and neurobiology produces a starkly different diagnosis [20]. Their central concept, the “hollowed mind,” names a condition in which habitual reliance on AI-generated answers progressively erodes what they term Cognitive Sovereignty — the capacity to engage in autonomous, foundationally grounded reasoning. Where Clark sees extended cognition [18], Klein and Klein see cognitive displacement: the offloading of cognitive work to AI does not extend the mind but empties it [45], particularly when foundational knowledge — the structured schemas that enable expert pattern recognition and analogical reasoning — is never built in the first place [20]. This concern is not merely theoretical; research on cognitive offloading demonstrates that habitual externalization of memory and reasoning can alter metacognitive monitoring and reduce the depth of internal knowledge consolidation [8].

Crucially, Klein and Klein introduce the Expertise Duality, a distinction that adds important empirical texture to the debate. AI use is not uniformly beneficial or harmful: for novices on routine tasks, AI levels performance by compensating for absent schemas, while for genuine experts on complex tasks, AI acts as an amplifier, extending existing competence [20]. This asymmetry has profound implications for educational design and workforce preparation [65]. It suggests that the cognitive risks of AI dependency are most acute precisely at the developmental stages — student years, early career — when foundational knowledge is being constructed. The worry is not that AI harms experts but that it may prevent novices from ever becoming experts, creating populations dependent on AI scaffolding that, if removed, would reveal an underlying cognitive deficit [20, 66].

Frameworks for Technologically Induced Cognitive Diminishment

Fasoli, Cassinadri, and Ienca contribute a formal theoretical framework that systematises the concerns raised by Klein and Klein into a broader account of technologically induced cognitive diminishment [67]. Rather than treating cognitive diminishment as an incidental side effect, their framework positions it as a potential structural outcome of technologies designed for enhancement — a kind of cognitive rebound effect in which the very mechanisms that make AI useful also erode the capacities that render that utility meaningful. This framing finds empirical grounding in Klein and Klein’s analysis of large language models, which argues that LLMs are qualitatively distinct from earlier cognitive tools such as calculators or search engines precisely because they can perform integrative reasoning — synthesis, argumentation, and evaluation — thereby enabling users to bypass the very mental effort that builds durable intellectual capacity [20]. Neuropsychological evidence cited in support of this position includes EEG studies showing that LLM use reduces frontal theta power, indicating diminished working memory engagement, and AI-assisted writing research reporting up to a 55% reduction in cortical activity alongside impaired memory integration [20]. This framework directly opposes Clark’s continuity thesis: where Clark argues that dependency on external cognitive resources is constitutive of human cognition and therefore not a pathology [5], Fasoli and colleagues argue that not all dependencies are equivalent, and that some forms of AI reliance cross a threshold into genuine cognitive harm [67, 18].

The unresolved question these frameworks collectively expose is whether there exist principled criteria for distinguishing genuine cognitive extension — meeting Clark and Chalmers’ original parity and trust conditions [3] — from mere offloading that leaves underlying capacities underdeveloped or atrophied. Clark’s position implies that such criteria may be philosophically misguided, since hybrid cognition is the norm rather than the exception [18]. The Fasoli et al. framework, by contrast, implies that such criteria are both possible and urgently necessary, particularly as AI becomes embedded in educational and professional contexts where, as Klein and Klein document, the protracted development of the prefrontal cortex leaves younger learners especially vulnerable to AI-induced cognitive atrophy during a sensitive period of prefrontal consolidation [67, 20]. The concept of empersonification proposed by [21] — the threshold at which an integrated AI system ceases to be a tool and becomes a constituent of the person — offers one promising approach to operationalising this distinction, though it has been developed primarily in the context of brain-computer interfaces rather than the large language models at the centre of current debate.

Human-AI Shared Regulation and Posthumanist Framings in Education

Educational research has begun to develop frameworks that engage more directly with the specific mechanisms through which generative AI integrates into learners’ cognitive processes, moving beyond the general philosophical question of whether extension occurs to the pedagogically consequential question of what kinds of human-AI cognitive relationships produce genuine learning. The Human-AI Shared Regulation of Learning (HASRL) model proposes that the defining characteristic of effective human-AI partnership in education is not the capability of either party in isolation but the quality of their joint regulation of cognitive, emotional, and motivational processes [65]. This represents a meaningful departure from self-regulated learning frameworks that locate regulatory agency exclusively within the individual learner [47, 68]: the HASRL model distributes regulatory responsibility between human and AI, formalising how this distribution unfolds temporally and across different learning phases [65]. Complementary work on smart learning environments has similarly identified self-regulated learning as a process that can be strategically leveraged through AI-mediated scaffolding [69], reinforcing the view that regulatory capacity is not a fixed individual trait but a relational achievement shaped by how learning tools are designed. If shared regulation is the core mechanism of hybrid intelligence, then educational AI systems must be evaluated not primarily for their content delivery or adaptive accuracy but for their capacity to sustain meaningful regulatory dialogue with learners — a design criterion that the extended mind framework would recognise as a condition for genuine cognitive integration rather than mere tool use.

A complementary critique from posthumanist epistemology targets the conceptual frameworks that educators and researchers currently employ to understand AI in learning. [66] argue that widely adopted models such as TPACK and SAMR fundamentally misrepresent generative AI by treating it as an inert technological object rather than as a relationally active participant in knowledge construction. In place of these frameworks, they propose the concept of symbi(AI)tic ecology — a vocabulary drawn from ecological metaphor and posthumanist theory to describe the spectrum of relationships between human learners and GenAI systems, ranging from mutualistic configurations in which both parties develop through interaction, to parasitic configurations in which one party’s development comes at the cost of the other’s [66]. The parasitic pole of this spectrum provides a conceptual vocabulary for the cognitive dependency risks that Klein and Klein’s hollowed mind thesis identifies [20]: AI interaction may reduce epistemic struggle, curtail exploratory thinking, and attenuate the development of independent knowledge-making capacities. Research on cognitive offloading supports this concern, demonstrating that habitual externalisation of cognitive processes to environmental supports — including digital tools — can alter metacognitive monitoring and reduce the development of internal representational capacity [8, 45]. Critically, neither the HASRL model nor the symbi(AI)tic ecology framework treats these risks as inevitable; both position them as design and pedagogical challenges contingent on how human-AI relationships are structured and reflexively managed.

These educational frameworks sharpen the theoretical stakes of the broader AI-extended mind debate in a specific and practically important way. In educational contexts, not all forms of cognitive extension are developmentally equivalent. As the expertise reversal effect established in Section 4 predicts [29], integration that scaffolds the learner’s own processing while preserving productive struggle may constitute genuine cognitive extension, whereas integration that substitutes AI output for effortful engagement may produce collective working memory gains without the corresponding development of individual cognitive capacity — precisely the condition that the desirable difficulties evidence [47] and research on knowledge-creation discourse [11] collectively warn against. The System 0 account of AI as a cognitive extension further underscores this tension, noting that the degree to which AI augmentation translates into durable individual competence depends critically on whether the human partner remains an active, effortful participant in the reasoning process rather than a passive consumer of AI-generated outputs [19]. The distinction between mutualistic and parasitic modes of integration — which the HASRL model’s emphasis on shared regulation and the symbi(AI)tic ecology framework both attempt to capture — may therefore prove more consequential for educational policy than the philosophical question of whether AI is genuinely “part of the mind.”

Symbolic Epistemic Agents and the Architecture of Cognitive Extension

A different register of engagement with these questions appears in the EDL framework proposed by Figurelli, which shifts from philosophical and psychological analysis to computational architecture [70]. EDL — a declarative domain-specific language for configuring symbolic cognitive agents — is presented as an attempt to implement epistemic agency computationally, enabling AI systems to maintain interpretability, coherence under symbolic degradation, and traceable reasoning through mechanisms including interpretive binding, mutation-based adaptation, and graceful semantic failure [70]. The EDL project is significant in this context because it represents an attempt to design AI systems that instantiate, rather than merely support, epistemic virtues — systems that could, in principle, satisfy stronger versions of the trust and integration conditions that extended mind theorists have proposed as criteria for genuine cognitive extension [3, 5, 70]. Clark and Chalmers’ original formulation required that an external resource be reliably available, automatically endorsed, and easily accessible for it to qualify as part of the extended cognitive system [3]; EDL’s emphasis on symbolic coherence and traceable inference can be read as a computational attempt to satisfy precisely these desiderata.

However, the EDL framework currently lacks empirical validation beyond symbolic physics domains, and its applicability to the open-ended, contextually sensitive reasoning that characterises human cognition remains to be demonstrated [70]. The framework’s emphasis on symbolic coherence and interpretability [71, 72] also sits in tension with the statistical, subsymbolic nature of the large language models at the centre of the broader debate [73], raising questions about whether insights from one architectural paradigm transfer to the other.

Synthesis and Outstanding Questions

Taken together, these works reveal a field in productive tension, with several significant disputes remaining unresolved. The most fundamental is the absence of agreed-upon criteria for distinguishing genuine cognitive extension from offloading that undermines cognitive development — a gap that philosophical and empirical literatures must jointly address. The expertise duality identified by Klein and Klein [20] and formalised by Fasoli and colleagues [67] indicates that the stakes of this question are unevenly distributed: cognitive risks are most acute at precisely the developmental stages when foundational knowledge is being built, making educational contexts a site of particular urgency. Educational frameworks for shared regulation [65] and symbi(AI)tic ecology [66] offer conceptual tools for distinguishing mutualistic from parasitic forms of integration, but these remain largely theoretical rather than empirically validated. The EDL project points toward architectural responses [70], but at an early stage.

Significant empirical gaps persist. Longitudinal studies tracking how sustained AI use reshapes cognitive capacities over years rather than sessions are conspicuously absent — a lacuna made more pressing by existing evidence that even relatively circumscribed offloading practices, such as reliance on external memory stores, can alter metacognitive monitoring over time [8]. This concern is sharpened by Clark’s [18] analysis of generative AI as a qualitatively new kind of cognitive partner, one whose interactive and linguistically productive character distinguishes it from prior tools studied in offloading research, yet for which longitudinal cognitive outcome data remain essentially nonexistent. Cross-cultural and cross-demographic evidence on differential vulnerability to cognitive dependency remains thin, with current research concentrated in Western, educated populations — a sampling bias that neuroanthropological perspectives highlight as a structural limitation of the broader cognitive science literature [74]. And the empersonification threshold proposed by [21] — promising as a criterion for distinguishing tool from cognitive constituent — has yet to be operationalised for large language model contexts, where the fluid, generative, and linguistically sophisticated nature of AI outputs complicates the application of criteria developed for simpler artifact–cognition couplings [19]; Piredda’s [6] defence of the extended mind hypothesis likewise underscores how existing mark-of-the-cognitive criteria strain under the demands posed by such dynamic, open-ended systems. Whether responsible AI integration is achievable depends substantially on closing these gaps, and the field has not yet produced the sustained interdisciplinary collaboration that doing so would require.

7. Social, Ethical, and Political Dimensions of Situated Cognition

The extended mind thesis, initially developed as an epistemological claim about the boundaries of cognitive systems, carries profound implications that reach far beyond philosophy of mind into questions of justice, identity, and political organization. If cognition is genuinely distributed across bodies, artefacts, and social relationships, then the design of material environments, the structure of social interaction, and the distribution of cognitive technologies become matters of urgent normative concern. This section traces the evolution of this normative turn in extended mind research across four thematic movements: first, the analysis of how physical artefacts and built environments serve as vehicles of oppression by actively constituting social hierarchies rather than merely reflecting them; second, the examination of how abstract concepts central to moral and political life are coconstituted through social interaction, rendering the conditions of that interaction a matter of epistemic justice; third, the concrete implications for disability and assistive technology, where access to cognitive scaffolding determines the material conditions for full cognitive functioning; and fourth, the emerging framework of cognitive diminishment and cognitive sovereignty, which reveals how technologies designed to augment some cognitive capacities may simultaneously attenuate others in ways distributed unevenly across populations. Together, these four threads are substantially accelerated by critical disability studies scholarship that has placed structural access and cognitive justice at the centre of the debate.

Oppressive Artefacts and Material Cognitive Systems

Early foundational work in extended cognition largely bracketed political questions, treating the mind-world boundary as a primarily metaphysical or epistemological puzzle. The landmark paper by Clark and Chalmers [3] established the canonical case for cognitive extension through their parity principle — if an external process functionally mirrors what would count as cognition if done inside the head, it counts as cognition — yet said little about the normative or political valence of such extensions. Clark later elaborated this framework into a broader account of the mind as constitutively entangled with tools, artefacts, and bodily action [5], and has more recently extended these commitments into the predictive processing paradigm itself [13]. It was not until the turn of the 2020s that researchers began systematically examining how the material scaffolding of cognition could itself be a vehicle of oppression. The pivotal contribution here is [75], whose framework of “oppressive things” argued that physical artefacts and spatial environments do not merely reflect pre-existing social hierarchies but actively constitute and reproduce them. Liao and Huebner proposed that oppressive systems operate through three mutually reinforcing components — psychological, social, and physical — and that the material dimension has been systematically undertheorized. On their account, a narrow doorway, a neighborhood designed around automobile access, or a workplace organized around a normative body type does not merely inconvenience marginalized users but reconfigures cognitive and social possibilities in ways that entrench racism, sexism, classism, and ableism. The framework operates bidirectionally: material environments shape psychological dispositions and social norms, while those dispositions and norms in turn influence how environments are designed and maintained [75].

This analysis invites comparison with earlier work on the individual in distributed cognitive systems. [37] had already challenged the assumption that the mind is bounded by skull and skin, arguing through an interdisciplinary synthesis of philosophy, cognitive science, and human sciences that mental representation and consciousness must be understood as extending into the social and material world. Embodied and enactive approaches have further developed this insight, emphasizing that perception, action, and cognition are shaped by the structural properties of the environments in which agents are embedded [15]. Crucially, Gallagher’s enactive account underscores that social institutions and built structures are not merely context for cognition but co-constitutive of it — a point that directly amplifies the normative stakes of oppressive design [15]. While Wilson’s primary concern was metaphysical rather than political, his insistence on the permeability of the cognitive boundary established the theoretical scaffolding upon which later normative analyses would build. Once it is accepted that the built environment partially constitutes rather than merely influences cognition, it becomes impossible to treat oppressive architecture as politically neutral.

A significant tension runs through this literature, however. Critics might argue that oppressive effects of material environments are better analyzed through structural sociology or critical geography than through the apparatus of cognitive science — that invoking extended mind theory adds theoretical complexity without commensurate explanatory gain. Against this, [75] contend that the cognitive framing is precisely what reveals the mechanism: oppressive things work not only by blocking access but by shaping habit, attention, and expectation in ways that naturalize hierarchy. Recent work in neuroanthropology and cultural cognitive science similarly documents how materially structured environments produce lasting differences in attentional patterns, habitual inference, and affective disposition across socially positioned groups [74]. What remains insufficiently theorized, however, is the relationship between these oppressive artefacts and predictive processing frameworks. Clark’s influential account of the brain as a “prediction machine” — a hierarchical system engaged in continuous prediction error minimization (PEM) against a generative model of its sensory environment [12] — has been shown to lend principled support to extended cognition: if external structures genuinely participate in PEM at an algorithmic level, they qualify as constituents of a cognitive system rather than merely its causes [14]. Cheadle and colleagues further argue that active inference frameworks must incorporate social and environmental embeddedness as constitutive rather than merely modulatory factors, providing a formal basis for understanding how structured social niches differentially shape agents’ generative models [63]. Applied to the politics of the built environment, this entails that environments systematically designed to disadvantage particular bodies would differentially shape prediction error profiles and habitual inference patterns — a connection the existing literature has yet to fully elaborate.

Abstract Concepts, Social Metacognition, and Epistemic Power

A complementary dimension of the normative turn concerns not the physical scaffolding of cognition but its social-interactional dimension, particularly in the domain of abstract thought. More recent theoretical work [62] has made a compelling case that abstract concepts — including concepts like justice, democracy, freedom, and race — differ fundamentally from concrete concepts in their dependence on social interaction, linguistic scaffolding, and emotionally embedded negotiation. Borghi, Mazzuca, and Tummolini distinguish between the mere “socialness” of a concept and the active processes of social metacognition through which meanings are monitored, aligned, and contested across interlocutors. Abstract concepts, on their account, are not simply retrieved from internal lexicons but are coconstituted through iterative social interaction, rendering their meaning inherently vague and negotiable [62].

This has significant political implications. If abstract concepts central to moral and political life are formed and stabilized through social interaction, then the conditions of that interaction — who gets to speak, whose framings are treated as authoritative, whose emotional and interoceptive contributions to meaning-making are recognized — become matters of epistemic justice. The framework developed by [62] thus intersects with ongoing debates about epistemic injustice and hermeneutical marginalization: those excluded from dominant social interactions are not merely denied political voice but are potentially excluded from the very cognitive processes through which shared normative concepts achieve determinate meaning. This represents an important tension within the extended cognition tradition: the same social distribution of cognitive labor that enables abstract thought also creates conditions for epistemic domination.

The tension between individual cognitive autonomy and the necessarily social character of abstract concept formation remains unresolved. One might worry that acknowledging the deep social constitution of abstract concepts undermines the conditions for independent critical thought — if my concept of justice is always already shaped by the social interactions I have been permitted to have, what resources remain for critique? [62] do not fully address this worry, and it represents one of the more significant open theoretical questions in this domain.

Disability, Assistive Technology, and the Structural Conditions of Cognitive Extension

The normative implications of extended mind theory are nowhere more concrete than in the domain of cognitive disability, where the question of access to cognitive scaffolding is not an abstract matter of epistemic justice but a daily material reality with measurable consequences for agency, identity, and social participation. If the extended mind thesis is correct — if cognitive processes genuinely extend into external artefacts and environments — then people who are denied access to appropriate cognitive scaffolding are not merely inconvenienced; they are denied the material conditions for full cognitive functioning. Disability studies scholarship has brought this implication into sharp focus, and in doing so has revealed that the Parity Principle, as originally formulated, requires supplementation with a structural justice framework [5]. As [42] demonstrates through careful analysis of the Ethical Parity Principle, no straightforward inference from cognitive parity to moral or political parity can be made without substantive additional reasoning about autonomy, identity, and the specific functional roles of cognitive tools in particular lives — reasoning that necessarily engages questions of access and structural inequality.

Critical disability studies has undergone a significant theoretical transformation over the past decade, moving from the foundational distinction between impairment and disability toward a more ambitious engagement with posthumanist theory. [25] argue that posthuman disability studies dissolves the distinctions between human and non-human, material and immaterial, biological and technological that earlier models had left intact, demanding instead attention to the relational assemblages through which technologies, bodies, and normative expectations co-produce one another. This framework proves indispensable for evaluating assistive technologies, which are never merely neutral prosthetics but carry embedded assumptions about what constitutes a normal body and a normal mind. Empirical research on assistive technology acceptance confirms this theoretical concern: user uptake is shaped not only by functional performance but by how devices align with or challenge users’ self-concept, social roles, and the normative expectations of their environments — factors that purely technical design criteria systematically underweight [76].

The ethical stakes of these assumptions are demonstrated with particular clarity in research on wearable robotic devices. [77], drawing on expert consultations across multiple European workshops, develop an extensive taxonomy of ethical, legal, and social implications revealing that wearable robots profoundly affect users’ self-perception and identity, potentially inducing machine-like body images and fostering device dependence that parallels — but is qualitatively distinct from — experiences with prostheses and wheelchairs. More fundamentally, these technologies risk reinforcing ableist assumptions by framing disability as a physical problem requiring technological correction, thereby narrowing accepted body norms and potentially undermining advocacy for accessible built environments [77]. [24] identifies five Cartesian assumptions embedded in such discourse — that disability represents a failing body or mind, that disability is located within individuals rather than social arrangements, that independence is the hallmark of a life worth living, that technology is the appropriate saviour, and that scientific expertise should guide intervention over the lived knowledge of disabled people themselves — constituting what Ashley Shew calls “technoableism.” The cognitive dimension of this critique is crucial for extended mind theory: if assistive technologies are designed around a deficit model — the assumption that disability represents an individual biological lack requiring medical intervention — then they may extend cognition while simultaneously restricting the forms of embodied and social life that the extended mind framework identifies as constitutive of cognitive agency [15]. Crucially, [24] introduces the concept of affordance shrinkage to capture how environments disable people by offering diminished possibilities for action, relocating disability from within the individual to the organism-environment relationship — a move that resonates directly with the extended mind thesis’s insistence on the constitutive role of environmental scaffolding.

Ethnographic research on neurodivergent children’s technology use makes the intersectional dimensions of this problem empirically visible. [78] demonstrates that autistic children’s media use is not radically different from that of nonautistic children; rather, how technology is experienced and accessed is profoundly shaped by the intersection of disability with race, class, and gender. Digital technologies serve as vehicles for meaningful social engagement, identity expression, and community formation — not as prosthetic substitutes for social capacities assumed to be absent [78]. This finding directly challenges design practices premised on cognitive deficit and reveals that the distribution of cognitive scaffolding — who gets access to which technologies, under what framings, and with what normative expectations — reproduces existing hierarchies of race, class, and ability [25]. The cognitive stratification that results is not an incidental side effect but a structural feature of how assistive and educational technologies are currently designed and distributed.

Research on brain-computer interfaces (BCIs) sharpens this point further by exposing systematic misalignments between the priorities of technology designers and those of disabled users. [79], drawing on qualitative interviews with patients representing key BCI target populations, found that participants did not prioritise mobility restoration — the goal that engineering and clinical teams typically foreground — but instead emphasised privacy, dignity, and independence from caregiver dependency as the domains where technological empowerment would be most meaningful. Patients who had experienced privacy violations or emotional asymmetries in relationships with caregivers expressed the strongest enthusiasm for BCIs as tools of autonomy rather than mobility [79]. This finding reveals a fundamental misalignment rooted in the medical deficit model: expert-driven design prioritises restoring normative function, while disabled people prioritise the relational conditions of their dignity and self-determination. Broader social science research on BCI use corroborates this gap, finding that qualitative and ethnographic methods consistently surface user concerns — around identity, dependency, and social participation — that are rendered invisible by the quantitative outcome metrics dominant in engineering-led research [80]. The global geography of BCI research compounds this concern: bibliometric analysis reveals that Africa accounts for less than one percent of global BCI publication output despite being home to more than 80 million people with disabilities, while psychology and the social sciences contribute approximately one percent of the publication record — confirming that the field’s growth has not been matched by proportionate attention to the human and justice-oriented dimensions of the technology [81].

The extended mind framework, properly applied to disability, thus yields a powerful conclusion: cognitive disability is better understood not as a deficit located within the individual but as a relational condition shaped by access to scaffolding, the design assumptions embedded in assistive technologies, and the structural inequalities that determine who benefits from cognitive extension and who does not [24]. The Internet of Things offers a concrete illustration: comparative analysis of identification technologies reveals that RFID achieves a 99% success rate in assistive retail navigation for visually impaired users, against just 26% for barcodes — a performance differential so large that infrastructure choice alone can determine whether an individual has functional independence or is effectively excluded from commercial environments [82]. Yet the same IoT systems that enable such independence carry persistent privacy and security vulnerabilities, and their implementation costs and infrastructure requirements map onto existing socioeconomic inequalities, ensuring that the cognitive extension they afford is unevenly distributed [82]. Taken together, these convergent lines of evidence — from wearable robotics, BCI research, intersectional media studies, and IoT infrastructure analysis — establish that the structural conditions of cognitive extension are inseparable from the justice conditions of disabled life.

Cognitive Enhancement, Diminishment, and Epistemic Justice

The most recent literature has extended these concerns into the domain of cognitive technology more broadly, raising questions about what might be called cognitive sovereignty — the capacity to exercise meaningful control over one’s own cognitive processes in environments saturated with technologies designed to modify them. [67] offer a systematic framework for understanding technologically induced cognitive diminishment as the underexamined counterpart to cognitive enhancement. While enhancement discourse tends to frame cognitive technologies as uniformly beneficial, Fasoli, Cassinadri, and Ienca demonstrate that technologies designed to augment specific cognitive functions may simultaneously attenuate others, generating what they term “dark side” effects that are distributed unevenly across populations [67].

This framework connects directly to questions of epistemic justice. Cognitive technologies do not affect all users equally: privileged populations typically exercise greater choice over which cognitive prosthetics they adopt, can more readily exit platforms that diminish autonomy, and have greater access to literacy resources that enable critical evaluation of these technologies [20]. Marginalized populations may face cognitive technologies as imposed rather than chosen infrastructures — algorithmic systems governing employment decisions, predictive policing tools, or attention-capturing platforms optimized for engagement rather than epistemic flourishing. The disability studies literature reviewed above provides the clearest empirical grounding for these concerns: the differential distribution of assistive technologies across race, class, gender, and geographic lines documented by [78], [81], and [82] demonstrates that cognitive stratification is not a hypothetical risk but a measurable reality. When brain-computer interface research remains concentrated overwhelmingly in wealthy nations — with equitable access, cybersecurity, and data privacy identified as pressing but underaddressed concerns in the field’s governance [83] — when wearable robot design encodes ableist assumptions about normative embodiment [77], and when the priorities of disabled users are systematically subordinated to those of clinicians and engineers [79], the result is a form of cognitive injustice in which the material conditions for extended cognition are distributed along lines of pre-existing social advantage.

The intersectional dimensions of this differential exposure remain a significant gap in the literature; existing frameworks tend to treat cognitive diminishment as a generic risk rather than one that compounds along multiple axes of oppression. [78] represents one of the few contributions that traces these compounding effects empirically, showing how the intersection of autism, race, and class shapes not only what technologies children can access but how their technology use is interpreted, medicalised, and accommodated by institutions. Future work must build on this intersectional foundation if the cognitive justice implications of extended mind theory are to move from theoretical articulation to actionable policy.

Questions of legal and policy frameworks for cognitive rights are only beginning to receive systematic attention, with [67] identifying the absence of adequate governance structures as a critical concern and [21] proposing that sufficiently integrated cognitive devices warrant legal protections associated with bodily integrity rather than property. Scholars have further noted that as BCIs and related neurotechnologies expand from medical applications into education, communication, and cognitive enhancement, the absence of proactive and inclusive regulatory frameworks risks entrenching existing inequalities rather than ameliorating them [83, 23].

Common Conclusions and Unresolved Questions

The four threads traced in this section — oppressive material artefacts, socially constituted abstract concepts, the structural conditions of disability and assistive technology, and the uneven distribution of cognitive enhancement and diminishment — converge on a single, consequential conclusion: cognitive justice cannot be achieved through functional analysis alone. It is insufficient to ask whether a given technology extends cognitive capacity in the abstract; one must ask whose cognition is extended, on whose terms, under what design assumptions, and with what redistributive consequences. Whether the mechanism is a built environment that naturalizes hierarchy by reshaping habit and expectation [75], a social interactional structure that excludes marginalized voices from the very processes through which normative concepts acquire meaning [62], an assistive technology whose deficit-model design forecloses forms of embodied agency it purports to restore [77, 79], or a cognitive enhancement platform whose “dark side” effects fall disproportionately on those with least capacity to exit it [67] — the underlying structure is the same: the material and social conditions of cognitive extension reproduce and amplify pre-existing social inequalities [25, 24].

Three significant gaps remain. First, intersectional empirics are almost entirely absent: with the partial exception of [78], the literature treats populations as homogeneous and cognitive diminishment as a generic risk rather than one that compounds across axes of race, class, gender, and disability simultaneously [25]. Second, legal and governance frameworks remain underdeveloped; the proposal by [21] that sufficiently integrated cognitive devices should attract bodily integrity protections is suggestive but represents a beginning rather than a framework, and the absence of adequate governance structures identified by [67] remains largely unaddressed — a gap that is all the more pressing given rapidly advancing neural interface technologies [83]. Third, and perhaps most fundamentally, the relationship between distributed cognition and political epistemology at the collective level — specifically, how oppressive material-cognitive systems are reproduced across generations through shared memorial practices and collective meaning-making [4, 58] — remains almost entirely unexplored in the extended mind literature [3, 1]. Closing these gaps will require not only more empirical work of the kind documented by [78, 82, 79, 77] but sustained theoretical engagement between cognitive science, critical theory, and political philosophy [84]. The conceptual tools are increasingly available; the task now is to deploy them in service of structural analysis and actionable policy.

8. Discussion

The body of research synthesized in this review marks a genuine inflection point in how cognitive scientists, philosophers, and practitioners understand the boundaries of mind. Taken together, the five thematic clusters do not merely accumulate evidence for the Extended Mind Hypothesis — they collectively reframe the central questions. The field has moved, in a relatively short period, from debating whether cognition extends into the environment to interrogating how, under what conditions, and with what consequences it does so. This shift carries significant implications for theory, methodology, and the design of cognitive systems in an era of generative AI.

What the Field Now Understands

The most consequential advance emerging from recent work is the progressive convergence between theoretical frameworks that were previously treated as rivals. Predictive processing accounts, once developed largely within internalist assumptions, are now being read as providing the mechanistic substrate that Clark and Chalmers’ original parity reasoning lacked [3]. If the brain is fundamentally a prediction machine minimising free energy [12, 85], then the selective recruitment of external resources — artefacts, other agents, digital systems — becomes not an exotic philosophical possibility but a routine computational strategy [13, 14]. This reframing closes one longstanding debate while opening another: the question is no longer whether the brain permits extension, but how predictive architectures determine which external resources become genuinely constitutive rather than merely causally efficacious [6]. Neuroscientific evidence on augmentation-related brain plasticity now provides empirical support for this reframing, demonstrating that the brain actively reorganises its representational maps in response to sustained use of prosthetic and assistive devices — a finding that shifts the debate from philosophical speculation to measurable biological process [41].

Simultaneously, empirical work on cognitive offloading has matured enough to reveal consistent patterns — selective offloading under memory load [8], interactivity effects in problem-solving [9, 7], representational scaffolding in expert practice — while also exposing the limits of current paradigms. The overwhelming reliance on simple laboratory memory tasks means we have robust findings about conditions that rarely obtain in consequential real-world cognition. Complex professional practice, collaborative design, and clinical decision-making involve dynamic, high-stakes cognitive work that offloading studies have barely touched. This is not a minor methodological inconvenience; it is a structural gap between the theory’s most ambitious claims and the evidence base actually supporting them. The participatory design research conducted with people living with dementia [40] represents a notable exception, demonstrating that ecologically valid studies of cognitive extension in vulnerable populations are both feasible and theoretically productive — and suggesting a methodological model that the broader field would benefit from emulating.

Learning sciences research has provided a crucial corrective to the performance-orientation of laboratory offloading studies by demonstrating that the consequences of cognitive offloading depend fundamentally on the developmental context in which it occurs. The expertise reversal effect — well established for individual scaffolding, where prompts that initially support novice learners become counterproductive as expertise develops [47] — acquires particular significance when extended to collaborative and AI-mediated settings. Collaborative Cognitive Load Theory formalises this insight for group learning, showing that collaboration’s cognitive benefits are conditional on a favourable balance between collective working memory gains and transactive coordination costs [29]. The consistent finding across this literature is that cognitive distribution enhances learning only when it preserves the effortful processing through which knowledge structures are constructed [20] — a conclusion that directly informs how the extended mind thesis should be applied to educational design and that sharpens the philosophical distinction between coupling that constitutes cognition and coupling that merely substitutes for it.

Where Recent Work Has Shifted Consensus

The emergence of large language models and generative AI has introduced genuinely novel pressure on the field’s theoretical vocabulary. Prior debates about cognitive extension focused on relatively passive artefacts — notebooks, diagrams, instruments — whose behaviour was fully determined by the user [4, 5]. Generative AI systems produce outputs the user did not anticipate, adapt in real time, and in some cases actively shape the user’s subsequent reasoning. Clark [18], drawing on predictive processing frameworks, describes the outcome as “brain, body, world tapestries” — continuously co-adapting wholes in which biological and non-biological components each shift in response to what the rest provides — and points to evidence from Go players whose exposure to superhuman AI strategies generated genuinely novel human moves rather than mere imitation, suggesting that generative tools can transform rather than merely supplement cognition. This challenges the coupling-constitution distinction [6] in ways that earlier cases did not, because the system’s contribution is not merely the storage and retrieval of previously encoded content but the generative transformation of cognitive inputs. Whether this constitutes a deeper form of extension or a categorically different relationship remains genuinely open, and it is arguably the most pressing theoretical question the field faces in the immediate term. The concept of empersonification [21] — the threshold at which an integrated device becomes a constituent of the person rather than a tool — offers a promising framework for addressing this question, though it requires operationalisation beyond its current application to brain-computer interfaces.

Equally significant is the mainstreaming of normative and political analysis within what was formerly a descriptive research programme. The recognition that cognitive systems are always also social arrangements — that who has access to cognitive scaffolding, whose representations become shared infrastructure, and how cognitive labour is distributed across gender and class lines are not peripheral concerns — has broadened the field’s scope in ways that cannot easily be reversed [10]. Critical disability studies scholarship has been instrumental in this broadening, demonstrating with empirical specificity that the populations most dependent on cognitive extension — people with dementia, acquired brain injuries, neurodevelopmental disabilities — are also those whose access to scaffolding is most constrained by structural inequalities [25, 78, 79]. Clark [18] raises a related concern at the level of epistemic culture, warning that generative AI systems risk reinforcing intellectual monocultures when deployed without adequate metacognitive oversight — a structural parallel to the access inequalities documented in disability scholarship. This is a genuine conceptual gain, though it also creates integrative demands that the field has not yet fully met.

Implications for Theory and Practice

Theoretically, the most urgent work involves developing criteria that are both philosophically principled and empirically tractable. The current situation — in which the parity principle provides intuitive guidance but no formal decision procedure — limits cumulative progress. As Section 3 demonstrated, much of this difficulty traces to unresolved questions about the scope and variety of functionalism that underwrites the Parity Principle: whether mental states are individuated by coarse-grained causal roles or fine-grained computational profiles determines how readily external processes can qualify as cognitive, and neither the extended mind camp nor its critics have converged on a single functionalist framework [33, 34]. Longstanding debates over the “mark of the cognitive” — what principled criterion distinguishes genuinely cognitive processes from mere causal coupling — remain unresolved, with internalist proposals such as the requirement for non-derived representations criticised for arbitrarily excluding functionally equivalent extended systems [6, 1]. Formal modelling approaches, drawing on active inference frameworks or information-theoretic measures of cognitive integration, offer one promising avenue [85]: predictive error minimisation (PEM), for instance, has been proposed as a principled, algorithmically-specified mark of the cognitive that is neutral about whether the implementing processes are internal or external, though this proposal carries its own assumptions about what counts as cognition that require scrutiny [14, 13]. The graduated integration thresholds proposed in work on ethical parity [42] and empersonification [21] suggest that progress may come not from a single binary criterion but from a multidimensional assessment of reliability, information flow, durability, trust, and transformative embedding — criteria that can be applied across the spectrum from simple memory aids to brain-computer interfaces.

Synthesis and Forward-Looking Priorities

Stepping back from the individual thematic contributions, what strikes the reader of this literature as a whole is the degree to which the five clusters — philosophical foundations, cognitive offloading, learning sciences, neuroscientific augmentation, and normative-political analysis — have developed largely in parallel rather than in sustained dialogue. Each cluster has made genuine progress on its own terms, yet the cross-cutting gains have been slower to materialise. The convergence now emerging between predictive processing accounts and extended mind theory [13, 14], between disability studies critiques and philosophical frameworks for cognitive justice [25, 24], and between learning sciences findings and the coupling-constitution debate suggests that the field is approaching a point of genuine theoretical integration. Realising that integration, however, requires deliberate methodological and institutional effort rather than the assumption that disciplinary proximity will produce it automatically.

Several research priorities emerge with particular clarity from the gaps documented across the review’s five sections. First, longitudinal investigation of AI-mediated cognitive extension is urgently needed. The studies reviewed here are almost entirely cross-sectional, capturing snapshots of human-AI coupling at single time points; they cannot address how sustained interaction with generative systems reshapes representational capacities, alters dependency profiles, or modifies the cognitive work that users themselves perform over time. This concern is sharpened by recent theoretical work warning that habitual reliance on generative AI risks hollowing out the foundational knowledge structures that anchor productive cognitive coupling [20], and by analyses suggesting that the extension of cognition through AI systems may alter not just what users offload but how they regulate and monitor their own thinking [19]. The neuroscientific evidence on brain plasticity following prosthetic use [41] provides both a theoretical rationale and a partial methodological template for such longitudinal designs, though transposing these methods to AI-mediated contexts will require significant adaptation. Second, cross-cultural research on cognitive offloading practices remains almost entirely absent. The laboratory paradigms reviewed in Section 2 are drawn overwhelmingly from WEIRD populations, and there is no principled reason to assume that the conditions under which offloading is selective, beneficial, or detrimental will generalise across the cultural and material environments in which cognition actually unfolds. Third, as noted above, the development of criteria for cognitive extension that are simultaneously philosophically principled and empirically tractable stands as a foundational theoretical need, without which cumulative progress on both descriptive and normative questions will remain inhibited. Finally, and perhaps most pressingly given the inequalities documented in Section 7, intersectional cognitive justice frameworks that move beyond generic accounts of cognitive diminishment to trace how its burdens compound across race, class, gender, disability, and geography are needed both to advance theoretical understanding and to ground actionable policy. Critical disability studies scholarship has begun to map how ableist assumptions are embedded in sociotechnical imaginaries [24], but this work has yet to be fully integrated with the philosophical and empirical literatures on cognitive extension. The participatory design model demonstrated with dementia populations [40] offers one concrete example of how ecologically embedded, justice-oriented research can be conducted; extending that model to the intersectional contexts identified in the normative literature represents a tractable and important next step.

Closing

This review began from the recognition that two developments — the rapid proliferation of generative AI as a cognitive partner and the sustained advocacy of disability scholars for a more materially grounded understanding of cognitive extension — have together created a moment of unusual urgency for the extended mind research programme. That urgency has not diminished in the course of the review; if anything, the gaps documented across its five sections have sharpened it. The philosophical foundations of the field remain contested at precisely the point where clarity would matter most: generative AI systems challenge the coupling-constitution distinction in ways that earlier artefact-based cases did not, raising new questions about what Clark [18] terms “brain, body, world tapestries” — continuously co-adapting wholes in which biological and non-biological components shift in response to one another — and demanding correspondingly new criteria for demarcating genuine cognitive extension [6]. The populations for whom cognitive extension is not a philosophical thought experiment but a daily condition of functional life — people with dementia, acquired brain injury, neurodevelopmental disabilities — remain underrepresented in both the empirical literature and the governance conversations that will determine how cognitive technologies are designed, distributed, and legally protected [25, 78, 79]. The tools for addressing these challenges — graduated integration thresholds [42, 21], intersectional empirical frameworks, ecologically valid longitudinal methods — are increasingly available. The task now is to deploy them in sustained dialogue across the disciplinary boundaries that have, until recently, kept the field’s most generative conversations from fully converging [1, 10].

---

References

[1] Rupert, R. (2009). Cognitive Systems and the Extended Mind. [2] Alexander, C. (2025). What is 4E cognitive science?. Phenomenology and the Cognitive Sciences. [3] Clark, A., Chalmers, D. (1998). The extended mind. Analysis. [4] Slors, M. (2020). From Notebooks to Institutions: The Case for Symbiotic Cognition. Frontiers in Psychology. [5] Clark, A. (2008). Supersizing the Mind: Embodiment, Action, and Cognitive Extension. Choice Reviews Online. [6] Piredda, G. (2017). The Mark of the Cognitive and the Coupling-Constitution Fallacy: A Defense of the Extended Mind Hypothesis. Frontiers in Psychology. [7] Kirsh, D. (2010). Thinking with external representations. AI & Society. [8] Risko, E., Dunn, T. (2015). Storing information in-the-world: Metacognition and cognitive offloading in a short-term memory task. Consciousness and Cognition. [9] Steffensen, S., Vallée‐Tourangeau, F., Vallée-Tourangeau, G. (2015). Cognitive events in a problem-solving task: a qualitative method for investigating interactivity in the 17 Animals problem. Journal of Cognitive Psychology. [10] Hollan, J., Hutchins, E., Kirsh, D. (2000). Distributed Cognition: Toward a New Foundation for Human-Computer Interaction Research. ACM Transactions on Computer-Human Interaction. [11] van Aalst, J. (2009). Distinguishing knowledge-sharing, knowledge construction, and knowledge-creation discourses. International Journal of Computer-Supported Collaborative Learning. [12] Clark, A. (2013). Whatever next? Predictive brains, situated agents, and the future of cognitive science. Behavioral and Brain Sciences. [13] Clark, A. (2022). Extending the Predictive Mind. Australasian Journal of Philosophy. [14] Kersten, L. (2022). A New Mark of the Cognitive? Predictive Processing and Extended Cognition. Synthese. [15] Gallagher, S. (2023). Embodied and Enactive Approaches to Cognition. Cambridge University Press eBooks. [16] Clark, A. (2008). Pressing the Flesh: A Tension in the Study of the Embodied, Embedded Mind. Philosophy and Phenomenological Research. [17] Bruner, E. (2023). Cognitive Archaeology, Body Cognition, and the Evolution of Visuospatial Perception. Cognitive Neuroscience. [19] Chiriatti, M., Ganapini, M., Panai, E., et al. (2025). System 0: Transforming Artificial Intelligence into a Cognitive Extension. Cyberpsychology Behavior and Social Networking. [20] Klein, C., Klein, R. (2025). The extended hollowed mind: why foundational knowledge is indispensable in the age of AI. Frontiers in Artificial Intelligence. [21] Bublitz, C. (2022). Might artificial intelligence become part of the person, and what are the key ethical and legal implications?. AI & Society. [22] van Balen, B. (2024). Can communication Brain-Computer Interfaces read minds? Demystifying Brain-Computer Interface mediated mindreading with enactivism. Phenomenology and the Cognitive Sciences. [23] Burwell, S., Sample, M., Racine, É. (2017). Ethical aspects of brain computer interfaces: a scoping review. BMC Medical Ethics. [24] van Grunsven, J. (2024). Disabled Body‑Minds in Hostile Environments: Disrupting an Ableist Cartesian Sociotechnical Imagination with Enactive Embodied Cognition and Critical Disability Studies. Topoi. [25] Goodley, D., Lawthom, R., Liddiard, K., et al. (2021). Key Concerns for Critical Disability Studies. International Journal of Disability and Social Justice. [26] Simon, C., Bolton, D., Kennedy, N., et al. (2021). Challenges and Opportunities for the Future of Brain-Computer Interface in Neurorehabilitation. Frontiers in Neuroscience. [27] Saha, S., Mamun, K., Ahmed, K., et al. (2021). Progress in Brain Computer Interface: Challenges and Opportunities. Frontiers in Systems Neuroscience. [28] Bai, Z., Fong, K., Zhang, J., et al. (2020). Immediate and long-term effects of BCI-based rehabilitation of the upper extremity after stroke: a systematic review and meta-analysis. Journal of NeuroEngineering and Rehabilitation. [29] Kirschner, P., Sweller, J., Kirschner, F., et al. (2018). From Cognitive Load Theory to Collaborative Cognitive Load Theory. International Journal of Computer-Supported Collaborative Learning. [31] Pearlman, K., Mackay, J., Sutton, J. (2018). Creative Editing: Svilova and Vertov’s Distributed Cognition. SHILAP Revista de lepidopterología. [32] Andy Clark (2008). Supersizing the Mind. [33] Mark Sprevak (2009). Extended Cognition and Functionalism. The Journal of Philosophy. [34] Michael Wheeler (2010). In Defense of Extended Functionalism. The MIT Press eBooks. [35] Fred Adams, Kenneth Aizawa (2001). The bounds of cognition. Philosophical Psychology. [36] Fred Adams, Kenneth Aizawa (2010). Defending the Bounds of Cognition. The MIT Press eBooks. [37] Robert Anton Wilson (2004). Boundaries of the mind the individual in the fragile sciences ; cognition. [38] Jakob Hohwy (2014). The Self‐Evidencing Brain. Noûs. [39] Kenneth Aizawa (2015). Cognition and behavior. Synthese. [40] Lazar, A., Edasis, C., Piper, A. (2017). Supporting People with Dementia in Digital Social Sharing. [41] Di Pino, G., Maravita, A., Zollo, L., et al. (2014). Augmentation-related brain plasticity. Frontiers in Systems Neuroscience. [42] Heinrichs, J. (2017). Against Strong Ethical Parity: Situated Cognition Theses and Transcranial Brain Stimulation. Frontiers in Human Neuroscience. [43] Smart, P. (2022). Minds in the Metaverse: Extended Cognition Meets Mixed Reality. Philosophy & Technology. [44] Malone, T., Crowston, K. (1994). The Interdisciplinary Study of Coordination. ACM Computing Surveys. [45] Britten-Neish, G. (2025). Cognitive offloading and the causal structure of human action. Synthese. [46] Evan F. Risko, Sam J. Gilbert (2016). Cognitive Offloading. Trends in Cognitive Sciences. [47] Nückles, M., Roelle, J., Glogger‐Frey, I., et al. (2020). The Self-Regulation-View in Writing-to-Learn: Using Journal Writing to Optimize Cognitive Load in Self-Regulated Learning. Educational Psychology Review. [48] Pouw, W., de Nooijer, J., van Gog, T., et al. (2014). Toward a more embedded/extended perspective on the cognitive function of gestures. Frontiers in Psychology. [49] Scardamalia, M., Bereiter, C. (1994). Computer Support for Knowledge-Building Communities. Journal of the Learning Sciences. [50] Lee, E., Chan, C., van Aalst, J. (2006). Students assessing their own collaborative knowledge building. International Journal of Computer-Supported Collaborative Learning. [51] Karsten Stegmann, Armin Weinberger, Frank Fischer (2007). Facilitating argumentative knowledge construction with computer-supported collaboration scripts. International Journal of Computer-Supported Collaborative Learning. [52] Reimann, P., Baker, M. (2025). Deepening our Epistemic Scripts: CSCL at the Intersection of Learning, Collaboration, and Information Sciences. Computer-supported collaborative learning/˜The œComputer-Supported Collaborative Learning Conference. [53] Zabolotna, K., Nøhr, L., Iwata, M., et al. (2025). How does collaborative task design shape collaborative knowledge construction and group‑level regulation of learning? A study of secondary school students’ interactions in two varied tasks. International Journal of Computer-Supported Collaborative Learning. [54] Immordino‐Yang, M., Damásio, A. (2007). We Feel, Therefore We Learn: The Relevance of Affective and Social Neuroscience to Education. LEARNing Landscapes. [55] Fischer, F., Bruhn, J., Gräsel, C., et al. (2002). Fostering collaborative knowledge construction with visualization tools. Learning and Instruction. [56] Wainio‐Theberge, S., Armony, J. (2023). Antisocial and impulsive personality traits are linked to individual differences in somatosensory maps of emotion. Scientific Reports. [57] Creß, U., Kimmerle, J. (2008). A systemic and cognitive view on collaborative knowledge building with wikis. International Journal of Computer-Supported Collaborative Learning. [58] Finley, J. (2025). Expanded taxonomies of human memory. Frontiers in Cognition. [59] Cowan, N. (2001). The magical number 4 in short-term memory: A reconsideration of mental storage capacity. Behavioral and Brain Sciences. [60] Rucha Khot, Teis Arets, Joel Wester, et al. (2025). Challenging Futures: Using Chatbots to Reflect on Aging and Dementia. [61] Dary, Z., Lenggenhager, B., Lagarde, S., et al. (2023). Neural bases of the bodily self as revealed by electrical brain stimulation: A systematic review. Human Brain Mapping. [62] Anna M. Borghi, Claudia Mazzuca, Luca Tummolini (2025). The role of social interaction in the formation and use of abstract concepts. [63] Cheadle, J., Davidson-Turner, K., Goosby, B. (2024). Active Inference and Social Actors: Towards a Neuro-Bio-Social Theory of Brains and Bodies in Their Worlds. KZfSS Kölner Zeitschrift für Soziologie und Sozialpsychologie. [64] Broniatowski, D. (2021). Psychological Foundations of Explainability and Interpretability in Artificial Intelligence. [65] Nguyen, A. (2024). Human-AI Shared Regulation for Hybrid Intelligence in Learning and Teaching: Conceptual Domain, Ontological Foundations, Propositions, and Implications for Research. Proceedings of the … Annual Hawaii International Conference on System Sciences/Proceedings of the Annual Hawaii International Conference on System Sciences. [66] Creely, E., Janssen, K. (2024). Onto-epistemological Understandings of Generative Artificial Intelligence in Education. International Journal of Changes in Education. [67] Marco Fasoli, Guido Cassinadri, Marcello Ienca (2025). The Dark Side of Cognitive Enhancement: A Framework for the Technologically Induced Cognitive Diminishment. Journal of Cognitive Enhancement. [68] Trautner, M., Pinquart, M. (2025). Development of motivational self-regulation in childhood—An integrative review. Frontiers in Psychology. [69] Adler, I., Warren, S., Norris, C., et al. (2025). Leveraging opportunities for self-regulated learning in smart learning environments. Smart Learning Environments. [70] Figurelli, R. (2025). EDL: A Domain-Specific Language for Epistemic Architectures in Heuristic Physics Systems. Preprints.org. [71] Kästner, L., Crook, B. (2024). Explaining AI through mechanistic interpretability. European Journal for Philosophy of Science. [72] Hassija, V., Chamola, V., Mahapatra, A., et al. (2023). Interpreting Black‑Box Models: A Review on Explainable Artificial Intelligence. Cognitive Computation. [73] Kotseruba, I., Tsotsos, J. (2020). 40 years of cognitive architectures: core cognitive abilities and practical applications. Artificial Intelligence Review. [74] Lende, D., Casper, B., Hoyt, K., et al. (2021). Elements of Neuroanthropology. Frontiers in Psychology. [75] Shen‐yi Liao, Bryce Huebner (2020). Oppressive Things*. Philosophy and Phenomenological Research. [76] Ventura, S., Ottoboni, G., Pappadà, A., et al. (2023). Acceptance of Assistive Technology by Users with Motor Disabilities Due to Spinal Cord or Acquired Brain Injuries: A Systematic Review. Journal of Clinical Medicine. [77] Kapeller, A., Felzmann, H., Fosch‐Villaronga, E., et al. (2020). A Taxonomy of Ethical, Legal and Social Implications of Wearable Robots: An Expert Perspective. Science and Engineering Ethics. [78] Alper, M. (2023). Kids Across the Spectrums: Growing Up Autistic in the Digital Age. The MIT Press eBooks. [79] Schicktanz, S., Amelung, T., Rieger, J. (2015). Qualitative assessment of patients’ attitudes and expectations toward BCIs and implications for future technology development. Frontiers in Systems Neuroscience. [80] Kögel, J., Schmid, J., Jox, R., et al. (2019). Using brain-computer interfaces: a scoping review of studies employing social research methods. BMC Medical Ethics. [81] Maiseli, B., Abdalla, A., Massawe, L., et al. (2023). Brain-computer interface: trend, challenges, and threats. Brain Informatics. [82] Semary, H., Al-Karawi, K., Abdelwahab, M., et al. (2024). A Review on Internet of Things (IoT)-Related Disabilities and Their Implications. Journal of Disability Research. [83] Sonko, S., Fabuyide, A., Ibekwe, K., et al. (2024). Neural interfaces and human-computer interaction: A U.S. review: Delving into the developments, ethical considerations, and future prospects of brain-computer interfaces. International Journal of Science and Research Archive. [84] Di Paolo, E., Potapov, K. (2024). Evald Ilyenkov and the enactive approach. Studies in East European Thought. [85] Pezzulo, G., Parr, T., Friston, K. (2022). Active Inference: The Free Energy Principle in Mind, Brain, and Behavior. Biological Psychology.