What makes the Extended Mind thesis incompatible with internalism?
This explores the actual philosophical fault line: the Extended Mind thesis denies that cognition stops at any fixed boundary (skull, skin, or system), while internalism insists the mind is whatever happens inside such a boundary — so the corpus reads them as incompatible at the level of where you draw the edge of a mind.
This explores why the Extended Mind thesis and internalism can't both be true: they disagree about where a mind ends. Internalism says cognition is whatever happens inside the agent's boundary — the skull, or by analogy the closed AI system. The Extended Mind thesis, the 1998 Clark and Chalmers argument, denies there is any privileged boundary at all: if a notebook or an external tool plays the same functional role a memory would, it is literally part of the cognitive process, not just an input to it. The incompatibility isn't a disagreement about facts inside the head; it's that Extended Mind rejects the very line internalism needs to draw.
The sharpest way the corpus stages this is by catching the thesis's own co-author switching sides. In the note on Chalmers using his 1998 commitments against himself Did Chalmers abandon his own Extended Mind principles?, his 2026 account of the LLM 'interlocutor' quietly places that interlocutor *inside* the AI system — an internalist move. That only works if the boundary is real and meaningful, which is exactly what Extended Mind spent its argument dissolving. The note frames this as an internal inconsistency: you can hold the externalist thesis or the internalist localization, but not both without retracting one.
What makes this more than a bookkeeping problem is *why* the boundary matters. Other notes show the mind's edge doing real work. The argument that consciousness candidacy requires an embodied encounter in a shared world Can disembodied language models ever qualify as conscious? is itself an externalist claim — it says the relevant facts about a mind include its co-presence with others and its triangulation on shared objects, not just internal states. That's Extended Mind's spirit pushed further: meaning and mentality are constituted partly by relations to a world, so sealing them inside a system loses the very thing you were trying to explain.
The tension cuts the other way too. The note that LLMs operationalize Saussure's *langue* Can language models learn meaning without engaging the world? shows a system whose meanings are *purely* internal-relational — defined by their place in a web of text with no external referents. That's a strange hybrid: relational like the externalists want, but with the relations all kept inside the model rather than reaching into a shared world. It suggests internalism and externalism aren't a clean binary; what you count as 'external' depends on whether you mean outside-the-head or outside-the-symbol-system.
The deeper lesson, which the note on Chalmers redefining 'interlocutor' makes explicit Does Chalmers silently redefine what interlocutor means?, is that these debates often turn on quietly relocating a boundary while keeping the old vocabulary. An interlocutor used to be a social-normative role that lives *between* people, not inside one of them; redrawing it as a behavioral state inside a system imports internalist boundaries under an externalist-sounding word. So the real incompatibility isn't abstract metaphysics — it's that you can't smuggle the boundary back in and still claim the mind was never bounded.
Sources 4 notes
The 2026 virtual-instance account locates the LLM interlocutor inside the AI system, implicitly adopting internalist boundaries that the 1998 Extended Mind thesis explicitly rejected. This creates internal inconsistency unless the earlier thesis is retracted or the new application misapplies its principles.
Current disembodied LLMs cannot be candidates for consciousness because consciousness language originates from and applies only to entities sharing a world with us through co-presence and triangulation on shared objects.
Research shows LLMs learn culturally situated discourse patterns by compressing relational structure from text, demonstrating that fluent language generation requires no external referents or embodied grounding.
Chalmers replaces the classical concept of interlocutor—a social-normative communicative role—with a behavioral-functional definition compatible with LLMs, keeping the traditional word to import its philosophical authority while delivering an entity with none of its properties.