Extended Cognition

Two related but categorically different claims are routinely conflated under the label "extended mind," and the conflation does real intellectual damage. The Extended Mind (EM) thesis — Clark & Chalmers (1998) — holds that mental states themselves can be partly constituted by external objects: Otto's belief about the museum location is literally in his notebook, not his brain. Extended Cognition (EC) holds the weaker and more defensible claim: cognitive processes — reasoning, computation, information integration — can be distributed across brain, body, and environment, without any external artifact becoming part of the mind. The mind remains bounded by the biological subject; external objects participate in cognitive processes without becoming part of the mind.

The distinction is not semantic. Under EM, wiping Otto's notebook destroys part of his mind. Under EC, it destroys a cognitive scaffold he was using — the notebook was never the location of his beliefs. This page defends EC and argues that EM is wrong.

Why the Extended Mind Fails

The coupling-constitution fallacy. Adams & Aizawa (2008) name the central error: being causally coupled to something during cognition does not make it constitutive of your mind. Digestion is causally coupled to food; food is not part of the digestive system. An LLM is causally coupled to a reasoning process; it does not thereby become part of the mind. Clark & Chalmers respond with the parity principle — if something plays the same functional role as an internal state, it is that state. But parity of functional role does not entail parity of metaphysical status. A prosthetic limb plays the functional role of an arm without being flesh. Functional equivalence is a criterion for engineering substitutability, not for ontological identity.

The phenomenology problem. Mind carries phenomenal character — there is something it is like to hold a belief. A notebook has no phenomenology. An LLM, whatever it is, does not have the subjective character that attaches to mental states. EC does not require external components to have phenomenology; EM appears to, or generates deeply uncomfortable questions about what it means for a belief to "be in" an artifact. If a belief can exist in a notebook with no experiential properties whatsoever, the concept of belief has been stretched past the point of usefulness.

Personal identity asymmetry. Mental states are mine in a way external objects are not. My memories cannot be physically removed and handed to someone else; a notebook can. This asymmetry is not incidental — it tracks something real about the difference between having a mind and using cognitive tools. EM would have to either deny this asymmetry or treat it as a contingent feature of how personal identity works, neither of which is satisfactory.

The dormancy problem. Otto is asleep. Under EM, his belief about the museum's location is currently sitting in the notebook — his mind is partly on the table. This requires either that minds persist during sleep in distributed form (deeply counterintuitive) or that the extension is toggled on and off (in which case what does "constitutive" mean, and what is the metaphysical status of the notebook between uses?). EC avoids this entirely: when the cognitive process stops, the notebook is an inert resource. Its information is available to future cognitive episodes, but it is not currently part of any mind.

Intrinsic versus derived content. Adams & Aizawa add a second argument: cognition has intrinsic, non-derived content — neural states mean things in virtue of what the biological system is, its causal history, its organismic organization. Notebooks and LLMs have derived content: they mean what users take them to mean. A notebook with addresses means nothing until someone uses it as a record. Therefore they are not cognitive in the constitutive sense, even when they participate in cognitive processes. Thompson (2007) makes a structurally similar point from the phenomenological tradition: mind is the ongoing self-organizational dynamics of a living organism — not a distributed computational process. A notebook can no more be part of a mind than a thermostat can be alive, because mind is a biological organizational property.

What Extended Cognition Does Claim

Rejecting EM does not require retreating to skull-bound cognition. EC makes substantive empirical claims that are well supported.

Hutchins' distributed cognition. Ship navigation is accomplished by a system of crew, instruments, and charts. The cognitive properties of navigation — the competence to get from A to B — belong to the system, not to any individual. No part of the system's competence is "in" anyone's head alone. Hutchins (1995) documents this ethnographically and formalizes it as distributed cognition: a proper unit of cognitive analysis that includes artifacts and social structure. The ship's navigation system is cognitively organized without any of its external components being parts of minds.

Vygotsky's mediated activity. Language, tools, and more capable others scaffold cognitive tasks children couldn't accomplish alone. The scaffolded process is cognitive; the scaffold does not become part of the child's mind. The key contribution: external scaffolds can be internalized over time, restructuring internal cognitive architecture. The ZPD is an EC concept — the cognitive process is extended, the mind is not.

Number as cognitive tool. Pica et al. (2004) documented that Munduruku speakers, whose language has no exact number words above approximately five, can do approximate arithmetic but not exact arithmetic on large quantities. Number words are not labels for pre-existing thoughts — they are cognitive tools that enable operations that are otherwise unavailable. See linguistic-relativity. The word participates in the cognitive process without becoming part of the mind; its absence makes certain cognitive processes impossible, not merely harder.

Key Figures Defending EC Against EM

Robert Rupert — Cognitive Systems and the Extended Mind (2009, Oxford). The most systematic philosophical treatment. Argues for HEMC — the Hypothesis of Embedded Cognition: cognitive systems are embedded in and shaped by environments without extending into them. EM requires too much promiscuity about what qualifies as cognitive; nearly any environmental feature that influences behavior would qualify, which drains the concept of content.

Fred Adams & Ken Aizawa — The Bounds of Cognition (2008, Wiley-Blackwell). Source of the coupling-constitution fallacy critique and the intrinsic/derived content argument. The most rigorous philosophical attack on EM's foundations.

Evan Thompson — Mind in Life (2007, Harvard). From the phenomenological tradition (building on Merleau-Ponty and Husserl via Varela and Maturana). Mind is a biological organizational property; it cannot be distributed into artifacts any more than life can. The boundary of the mind is the boundary of the living organism, not the boundary of the causal process the organism engages in.

Alva Noë — Out of Our Heads (2009). Noë is not straightforwardly an EC defender — he thinks neuroscience badly misunderstands mind — but he is explicitly not an EM defender either. Mind is a relational property of organism-environment coupling, but it's not in external objects. Perceptual experience happens in active engagement with the world — not in the brain alone, and not in the objects. The environment contributes to mind without constituting it.

EC as the Right Frame for Human-AI Collaboration

The stakes are not merely philosophical. EM and EC generate different predictions and normative frameworks for thinking about human-AI systems.

Under EM, human cognitive contribution becomes analytically optional. A sufficiently capable LLM and a passive, endorsing user constitute a cognitive system — the human is along for the ride, their internal states redundant if the LLM's outputs are good enough. This is the wrong frame empirically: the Dell'Acqua et al. "jagged frontier" finding shows that the human's evaluative competence is decisive in determining whether AI assistance produces good or bad outcomes. It is also normatively wrong: it obscures the cognitive autonomy questions that matter for AI policy. If the LLM is part of the user's mind, questions about the LLM's influence on cognition become questions about what is permissible to do to someone's mind — a frame that applies equally to education, drugs, and propaganda, but that is incoherent applied to an external tool.

Under EC, the human mind is always the subject of extended cognitive processing. The LLM extends the reach of that cognition; it does not replace its source. What the human brings internally — representational architecture, evaluative capacities, accumulated conceptual structure — determines what kind of extension is possible and how productive it is. The cognitive process is distributed; the mind is not.

Literacy as Coupling Mechanism

EC provides the theoretical grounding for why literacy predicts AI productivity in ways that exceed "literate people can follow instructions." Literacy is an internal cognitive capacity that determines the bandwidth of extended cognitive processing. A person with high compositional literacy brings richer internal representations to the coupling: they can specify problems precisely, recognize when outputs are wrong, decompose complex goals, and integrate LLM outputs into larger cognitive structures. The extended cognitive process achieves correspondingly more.

This is not about tool skill — it is about the prerequisite internal architecture for productive distributed cognition. The same LLM coupled to a more cognitively capable human produces qualitatively different extended cognitive processes. This connects to language-of-thought: the LOT is not in the LLM. It is distributed across the human-LLM system. The human provides functional grounding; the LLM provides compositional fluency at scale. The system's LOT-like properties depend on the quality of the human side. See lot-llm-paradox.

The Vygotsky Alignment

Vygotsky's Zone of Proximal Development maps cleanly onto the human-LLM coupling. With the right scaffold, a child accomplishes cognitive tasks that are impossible alone. The scaffold participates in the cognitive process. But the scaffold does not become part of the child's mind — it gets internalized over time, restructuring internal cognitive architecture. LLMs work analogously: scaffolds for extended cognitive processes that, used well, can build stronger internal models through practice. Used badly — passively accepted without engagement — they produce the opposite: atrophy of the internal capacities that make productive coupling possible.

The normative implication is direct. If the LLM is an external scaffold in the EC sense, the relevant question is not "is the LLM capable enough?" but "is the human-LLM coupling structured to strengthen or weaken the human's internal cognitive architecture?" EM dissolves this question; EC makes it central.

Related

language-of-thought · lot-llm-paradox · linguistic-relativity · agentic-workflows · mcp

Sources

• Clark, A. & Chalmers, D. (1998) "The Extended Mind," Analysis 58(1), 7–19
• Clark, A. (2008) Supersizing the Mind, Oxford University Press
• Adams, F. & Aizawa, K. (2008) The Bounds of Cognition, Wiley-Blackwell
• Rupert, R. (2009) Cognitive Systems and the Extended Mind, Oxford University Press
• Thompson, E. (2007) Mind in Life, Harvard University Press
• Hutchins, E. (1995) Cognition in the Wild, MIT Press
• Extended Mind — Stanford Encyclopedia of Philosophy