From Extended Minds to Coupling Flexibility: Cognitive Eco-Sourcing and Generative AI in Education

Synthesis: Di Paolo argues that debate about Generative AI in education is stuck in an enhancement-versus-diminishment oscillation that hides the question that actually matters: how cognitive labor is distributed when a learner recruits a technological resource. Building on the extended-mind tradition but moving past its concern with whether artifacts become part of the mind, the paper proposes cognitive eco-sourcing as the process of functionally recruiting distributed material, social and technological resources, and distinguishes four deliberately non-parallel, overlapping and time-variable forms of coupling: offloading, where demands are reorganized while the learner remains the principal performer; uploading, where a constructed niche stably supplies a function across agents and cohorts; delegation, where the resource executes an operation but the agent stays substantively inside the production loop with effective supervision and possible authorship; and outsourcing, which is external execution without effective supervision or authorship. These forms are analyzed through integration, execution, supervision and authorship, and judged against a developmental relevance criterion indexed to the target capacity and the learner's current competence. The central claim is conditional: no form is intrinsically beneficial or harmful. Education should therefore cultivate coupling flexibility, the developing ability to understand, evaluate, select and reorganize couplings as epistemic and educational aims change.

Key Findings

1. The taxonomy is of relations, not of technologies. Saying that a learner "uses AI" identifies a resource but not the form of recruitment. The same system may offload prospective memory, delegate sentence-level revision, or outsource an entire literature synthesis, and a knowledgeable user may supervise a GenAI system while someone unable to read a spreadsheet's formula outsources the calculation it performs. Forms can coexist inside one activity at different analytical scales and shift over time, so analysis must track both a coupling's organization at a moment and its trajectory.
2. Cognitive eco-sourcing has four forms, deliberately non-parallel. Offloading reduces or reorganizes information-processing demands (the Montessori Stamp Game preserves intermediate quantities); uploading is the creation of a new cognitive function taken on by a constructed niche and surviving the particular agents who made it (traffic signals coordinating right of way); delegation and outsourcing both involve external execution but differ in supervision and authorship.
3. Four dimensions do the discriminating work. Integration asks how stable, individualized and substitutable a resource is within a routine. Execution asks which operations are transferred. Supervision asks whether the agent has the competence and opportunity to understand, monitor, evaluate, redirect and revise an externally performed operation. Authorship asks whether the agent can intelligibly own and answer for the realized performance. Authorship is related to, but reducible to neither, execution nor supervision.
4. The delegation/outsourcing line is supervision, not reliance. Delegation holds when the agent remains substantively within the production process — think of a renowned chef whose brigade executes differentiated operations under a culinary conception the chef is answerable for. Outsourcing is present when a resource realizes a function outside the agent's supervision: accepting requirements-and-outcome control, as a restaurant customer controls the order, is not authorship. Delegation can decay into outsourcing without any change in technology or explicit decision to relinquish control, as reduced involvement erodes situational awareness.
5. Education changes the stakes because activity is formative. A coupling that improves present performance may alter what Learners can later perform, supervise or claim as their own. The proposed developmental relevance criterion asks three questions: what is the target capacity, what is the learner's current competence, and what trajectory does the coupling support. A productive trajectory need not end in independence — it may combine enduring supports with growing capacity to supervise, author and outsource selectively.
6. The novice supervision paradox. Learners are told to verify GenAI outputs while still acquiring the knowledge that verification requires; without that knowledge, "checking" becomes a ritual that leaves the system's apparent authority intact. The paradox does not justify prohibition: supervision can be scaffolded and distributed across a group or classroom, and participation in a jointly supervised activity can build the habits that later let learners assume supervision and authorship themselves.
7. Integration can become parasitic, and outsourcing can be capacity-displacing. High integration may coexist with weak supervision and diminished authorship, and what becomes parasitic is an arrangement persistently serving external interests at the learner's expense — not dependence as such, since skilled cognition routinely relies on enduring social and material resources. The problem is a trajectory in which dependence deepens without corresponding growth in competence, supervision or authorship.
8. Design, practice and Assessment are the levers. Because design shapes which operations are readily transferred and how visible the distribution stays, the paper argues for interfaces that elicit an initial contribution, reveal assistance progressively, signal uncertainty and keep contributions open to revision; for inspectable couplings that let learners revisit, compare and withdraw support; and for assessment that asks learners to reconstruct arguments, map premises, justify revisions or explain rejected alternatives rather than inspecting the finished text.
9. Coupling flexibility is the educational aim. The plural cognitive ecology it presupposes coordinates multiple routes — teachers, peers, books, physical and low-tech and multisensory tools, GenAI — around the target capacity rather than multiplying representations indiscriminately, preserving alternatives that both enable verification and reduce dependence on any single resource.

The extended-mind starting point and its limits

The paper opens from the extended mind thesis, which made explicit the question of whether resources beyond skin and skull can constitute components of a cognitive system, and from the claim that ordinary artifacts perform functions commonly associated with internal cognition: notebooks store, written marks stabilize intermediate results, tools transform the structure of problems. But the author follows Sterelny in treating the tighter extended-cognition cases as limiting cases inside a wider field of environmentally supported cognition, and reads the contrast as a difference in explanatory direction: scaffolding accounts begin with the environment and ask how available resources support competent activity, extended-cognition accounts begin with the agent and ask how resources become integrated into its cognitive economy.

What the starting point leaves open is the paper's actual question. Even where a resource is not incorporated in the strong extended-mind sense, couplings may still be cognitively consequential, and what matters is how recruitment is regulated, stabilized and made more or less available. The author therefore reserves scaffolding for the structured field of resources capable of supporting activity, and coupling for the functionally organized relation established when resources are recruited — a movement from Scaffolding to coupling that changes the explanatory focus rather than the metaphysics. This positions the argument against accounts of schools as cognitive ecologies and GenAI as an emerging epistemic infrastructure: the same author's related conceptual paper, Educating minds with generative AI, situates AI inside historically structured, multi-technological environments, whereas this paper examines the variable relations through which agents recruit resources within them.

Cognitive eco-sourcing: four forms, four dimensions

Cognitive eco-sourcing names the process of resource recruitment, and the concept carries three commitments: cognition is ecologically distributed across artifacts, other people, institutional routines, spatial arrangements and digital systems; distribution is selective, so agents recruit only a subset of what is available depending on task, skill, learning history and resource properties; and recruitment establishes a context-sensitive relation among agent, resource and function. Selection may be metacognitively guided or may reflect learned patterns of attention and action rather than explicit decisions.

The four forms sit at different temporal and organizational scales. Offloading reorganizes or reduces demands without prescribing how integrated the coupling is; the paper contrasts a substitutable GenAI checklist against the Stamp Game, where material participates in a coordinated sensorimotor process and is a strong candidate for extended cognition. Uploading is the stronger environmental reorganization in which the niche itself performs or reliably supplies a function that acquires some independence from its founders — traffic regulation being the illustration — and in education, curricula, timetables, assessment practices and pedagogical roles already upload functions that persist as teachers and students change. GenAI may contribute to uploading when explanation, Feedback, drafting or evaluation become routinely available through platforms, practices and expectations across cohorts; occasional chatbot use does not meet that condition.

Delegation and outsourcing turn on external execution. In delegation the agent monitors the developing contribution, evaluates it against the wider activity's aims, redirects or revises it, and determines its incorporation; supervision need not be continuous observation of every internal operation, only substantive opportunities to intervene, since a merely formal opportunity to accept or reject does not place the agent within production if automation bias defeats it. Authorship may be shared — collaborators who supply ideas or draft sections share authorship as well as execution — which is why a chef can remain answerable for a brigade's differentiated work. Outsourcing transfers execution and supervision together while excluding authorship of that function, though the agent may still author the wider activity that uses its output; a researcher who outsources transcription keeps authorship of a study's questions and claims, whereas a student who accepts an AI-generated literature review without being able to reconstruct how sources were selected has outsourced epistemic evaluation. The breadth of Large Language Models (LLMs) systems makes outsourcing widely available and less apparent, because a technology becomes transparent when agents act fluently through it and cease to attend to the technology itself.

The novice supervision paradox and parasitic integration

Education is formative, so the question is not whether cognitive operations stay inside the head — learning has always been scaffolded and is a concern for Learning Theories — but how particular distributions of labor relate to the capacity being cultivated and to the learner's current position. Three complications follow. First, target capacity must be specified, since an essay may develop factual knowledge, argument construction, source evaluation, disciplinary style or linguistic fluency, and different GenAI uses redistribute different parts of it. Second, effects depend on prior competence, so guidance useful to novices may become redundant as expertise develops while contributions experts can supervise may exceed a novice's capacity for interpretation and evaluation. Third, trajectories differ: some scaffolds are withdrawn, others remain integrated into skilled activity for life.

The novice supervision paradox sits at the junction of the first two. Fluent and plausible GenAI output may contain inaccurate, fabricated or unsupported claims, so learners are asked to verify what they are simultaneously still learning to understand; without appropriate knowledge and methods, checking can become ritual. Where GenAI performs a developmentally relevant operation the learner cannot assess or redirect, the coupling may constitute outsourcing immediately, and becomes capacity-displacing when it removes the practice that would make supervision possible. The temporal dimension is that couplings formed before learners can supervise them can stabilize into their cognitive ecology. A tightly integrated technology may support a user's aims, or it may direct attention and behavior towards external interests that conflict with them — the paper's notion of parasitic cognition — and capacity-displacing outsourcing is a distinct but overlapping problem that becomes parasitic when providers or institutions benefit from encouraging dependence. Integration becomes parasitic when the arrangement persistently serves external interests at the learner's expense.

What follows for education: designing and teaching the coupling

The plural cognitive ecology is the paper's structural response: teachers, peers, books, physical activities, low-tech tools, multisensory devices and GenAI offer different forms of access to the same object of learning, letting learners verify AI-generated information, compare representations and methods, and consider which support suits the target knowledge. But plurality must be coordinated around the target capacity rather than multiplied indiscriminately, since redundant representations can produce cognitive overload; the worked example is the equivalence 3 + 3 = 6 represented symbolically, with physical objects, and in a virtual environment, so learners identify what remains invariant rather than binding the relation to one medium. Coordination does not require a single route, and varied forms of representation and participation may reduce barriers for learners with different sensory and cognitive profiles.

Design is the second lever, because the coupling that emerges depends on the interaction as well as the user's choice. An interface that immediately supplies complete responses facilitates several transfers at once, whereas one that responds with questions, hints, alternatives or targeted feedback keeps contributions bounded and open to supervision; the argument is that design does not determine the coupling but makes some forms more likely, and that learners and educators should both participate in deciding how systems are incorporated. Policy should regulate which functions may be transferred within a particular activity rather than tools as a whole, and AI Literacy should be cultivated as scaffolded, reflective practice rather than technical proficiency. Because these relations cannot be inferred from the finished product — fluent prose may conceal whether a learner understands an argument — assessment should examine the learner's relation to the work by asking for reconstruction, premise-and-conclusion mapping, justification of revisions, or reasons an alternative was rejected. Metacognition is where this lands developmentally: the aim is neither independence from technology nor integration for its own sake, but learners who understand what different resources contribute, can supervise transferred operations, and can decide when particular couplings serve their aims. That is coupling flexibility, and the author extends Clark's "canny cognizer" and Hutchins on cultural organization while insisting that cognitive economy cannot be education's sole criterion.

What this means for practice

• Instructors. Name the target capacity before choosing a GenAI use, because the same essay assignment develops factual knowledge, argument construction, source evaluation, disciplinary style or linguistic fluency, and each use redistributes a different part of it.
• Designers. Configure assistance to keep contributions bounded and open to supervision — questions, hints, alternatives and targeted feedback rather than complete responses — since the interface makes unsupported transfers more likely without determining the coupling.
• Instructors. Assess the learner's relation to the work rather than the finished product: ask for reconstruction, premise-and-conclusion mapping, justification of revisions, or reasons an alternative was rejected, because fluent prose does not reveal what the learner supervised.
• Administrators. Keep a plural cognitive ecology in the curriculum — teachers, peers, physical materials and low-tech tools alongside GenAI — and avoid organizing a program around one pervasive technology or default coupling, which is the paper's practical conclusion under uncertainty.
• Researchers. Code observed human–AI activity with the four dimensions (integration, execution, supervision, authorship) and track trajectories over time, since a coupling's form can shift without any change in the technology.

Limitations

• The four forms may coexist within an activity, admit borderline cases and change over time; the authors reply that this does not undermine the taxonomy because the forms are neither mutually exclusive nor positions on a single continuum, and borderline cases may reflect genuine uncertainty about how labor is organized or genuine transitions, which are common in learning.
• "Delegation" and "outsourcing" anthropomorphize AI, and the reply is that the vocabulary is functional, describing the organization of cognitive labor, so a system may execute a transferred operation without possessing intentions, authority, responsibility or authorship; responsibility remains with human and institutional actors.
• It is a conceptual argument, not an empirical study: it generated and analyzed no data and claims no findings about learning outcomes, and evidence on the long-term developmental effects of GenAI remains preliminary, so claims about cognitive debt, deskilling and parasitic integration should be investigated rather than treated as settled diagnoses; the taxonomy is offered as a way to direct that research.
• It has normative limits — it cannot determine which capacities education ought to cultivate or how competing aims should be ranked — though it is not normatively neutral, since foregrounding competence, supervision, authorship and capacity development identifies features whose distribution requires justification; it is Philosophy of AI in Education and Theory Development in AI in Education work rather than an intervention study.

Citation

Di Paolo, L. D. (2026). From Extended Minds to Coupling Flexibility: Cognitive Eco-Sourcing and Generative AI in Education. Topoi (manuscript submitted for publication).