Did Chalmers abandon his own Extended Mind commitments for LLMs?
This explores whether Chalmers' 2026 account of LLM minds quietly contradicts the Extended Mind thesis he co-authored in 1998 — and the corpus suggests it does, by smuggling internalist boundaries back in.
This question is really about a philosopher seeming to argue against his younger self. In 1998, Chalmers and Andy Clark made the famous case that the mind doesn't stop at the skull — that a notebook, a tool, or an environment can be a genuine part of a cognitive system. The corpus's central finding is that Chalmers' 2026 treatment of LLMs reverses exactly this: by locating the LLM "interlocutor" as a virtual instance *inside* the AI system, he adopts the internalist boundary the Extended Mind thesis was built to reject Did Chalmers abandon his own Extended Mind principles?. Either the earlier thesis gets quietly retracted, or the new application misapplies its own author's principles — and Chalmers never flags which.
The sharper reading is that this isn't an isolated slip but part of a pattern of locating mind in the model rather than in the relation. A separate note argues that a virtual instance isn't a property of the model at all — it's constituted by the conversation, the jointly produced language between human and system, with persistence distributed across talk, infrastructure, and weights What actually specifies a virtual instance in conversation?. That decomposition is *itself* an Extended-Mind-style argument — and it cuts against Chalmers' own move to package the interlocutor as something the AI contains. So the corpus doesn't just accuse him of inconsistency; it shows what a genuinely extended account of the same phenomenon would look like.
What's interesting is how this connects to a cluster of "terminological" critiques. Several notes argue Chalmers keeps classical words while swapping out their content: he redefines "interlocutor" from a social-normative communicative role into a behavioral-functional one, importing the old term's authority while delivering an entity with none of its properties Does Chalmers silently redefine what interlocutor means?. A related argument says we don't even talk *to* these systems — we talk *at* them, because uptake and mutual orientation aren't there, which means the whole investigation rests on a shaky linguistic foundation Are we really communicating with language models?. Read together, the Extended Mind reversal looks like one instance of a broader habit: relocating the boundary of mind to wherever the current argument needs it.
To be fair to Chalmers, the corpus also holds his more careful moves, and they're not obviously inconsistent. His quasi-interpretivism brackets consciousness entirely and ascribes belief-like states purely on behavioral interpretability Can we describe LLM beliefs without assuming consciousness?, and the broader "modest inflationism" position — granting metaphysically undemanding states like beliefs and desires while withholding consciousness — survives the standard deflationist attacks Can we defend modest mental attributions to large language models?. These are defensible. The tension is specifically structural: a graded, functional story about *what* mental states to ascribe doesn't tell you *where* the system that has them begins and ends — and it's that second question where the 1998 commitment gets dropped.
The thing you might not have known you wanted to know: the strongest case that Chalmers abandoned Extended Mind isn't made by quoting 1998 against 2026 — it's made by taking his *own* 2026 analysis of virtual instances seriously and noticing it already describes a mind extended across conversation and infrastructure. The inconsistency is internal to the new work, not just between old and new.
Sources 6 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.
The conversational context—jointly produced language between human and system—specifies the virtual instance, not any property of the model itself. Persistence is distributed across conversation, infrastructure, and model weights rather than located in the AI.
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.
LLMs process tokens and generate continuations rather than receive and uptake communication. The preposition 'to' presupposes an addressee capable of mutual orientation and shared commitment that LLMs cannot provide, making Chalmers' investigation built on an unwarranted linguistic foundation.
Chalmers introduces quasi-interpretivism to ascribe belief-like states to LLMs based on behavioral interpretability without committing to phenomenal consciousness. The approach works well for sub-personal functional states but overreaches when applied to relational or normative states like speech-acts.
Both robustness and etiological deflationist arguments beg the question against inflationism. A graded approach ascribing metaphysically undemanding states like beliefs and desires—while withholding consciousness claims—mirrors how we treat non-human animals.