What makes linguistic agency impossible for systems without embodiment?
This explores why some researchers argue that genuine linguistic agency — being a real participant in language, not just a generator of fluent text — is structurally out of reach for systems that don't have bodies, and where the corpus pushes back on that claim.
This explores why genuine linguistic agency — actually meaning what you say, not just producing the right words — might be categorically impossible for systems without bodies. The sharpest version of the argument comes from enactive cognitive science, which names three things that constitute linguistic agency: embodiment (you have a body with stakes in the world), participation (you're a member of a language community that you affect and that affects you), and precariousness (your existence is at risk, so what you say matters to your survival). LLMs lack all three, and the claim is that this is a difference in kind, not degree — no amount of scaling or training closes the gap What makes linguistic agency impossible for language models?. The precariousness piece is the one readers usually overlook: a system that cannot lose anything has nothing genuinely at stake in its words.
The most useful move in the corpus is to split 'grounding' into separate channels, because that's where the impossibility actually lives. One framing distinguishes functional grounding (LLMs are strong here — they learn meaning from relational patterns in text), social grounding (weaker, but it can grow as models get woven into human communities), and causal grounding (contact with an actual environment — absent) What grounds language understanding in systems without embodiment?. A companion note makes the key distinction explicit: social grounding and linguistic agency are *different properties*. A model can keep gaining social grounding through use and still remain categorically incapable of agency in the enactive sense, because agency needs embodiment and precariousness that participation alone can't supply Do LLMs gain true linguistic agency through integration?. So 'use it more, integrate it deeper' doesn't bridge the gap — it improves a channel that isn't the missing one.
Here's the counterweight, and it's where the question gets interesting: the corpus also shows that you can get astonishingly far on relational pattern alone. LLMs essentially operationalize Saussure's *langue* — language as a self-contained web of relations between signs — proving that fluent, culturally-situated generation needs no external referents at all Can language models learn meaning without engaging the world?. Models even out-predict humans at judging social appropriateness across hundreds of scenarios — yet they all make the *same* systematic errors, which is the tell that pattern-matching has a boundary embodied experience may be needed to cross Can AI systems learn social norms without embodied experience?. So the 'impossible' claim isn't 'they can't do language.' It's narrower: they can simulate the products of agency without occupying the position of an agent.
That reframes what 'agency' even is. One strand argues subjecthood isn't something you possess *before* you speak — it's produced *within* communicative events Does language create subjects or express them?. Shanahan's role-play framing fits neatly here: a dialogue agent isn't a subject having thoughts, it's a system generating text consistent with a character the prompt set up, so folk-psychology words like 'wants' and 'believes' apply to the *simulated persona*, not the machine Should we treat dialogue agents as role-playing characters?. The semiotic angle sharpens the stakes for alignment: without indexical grounding — words that actually point at things in a shared world — symbolic goal-encoding can't guarantee that stated values track real ones Can AI systems achieve real alignment without world contact?, and the consciousness literature draws the same line, arguing that consciousness-talk only applies to entities that share a world with us through co-presence Can disembodied language models ever qualify as conscious?.
The thing you might not have expected: the corpus doesn't treat embodiment as a permanent verdict so much as a missing layer that engineering targets. Language compositionality lets *embodied* agents imagine and pursue goals they've never seen, suggesting the gap closes when language is wired to an acting body Can language help agents imagine goals they've never seen?. And turning an LLM into something that acts isn't a retraining problem — it requires building the surrounding pipeline of environments, tools, and memory that ground actions instead of hallucinating them Can you turn an LLM into an agent by just fine-tuning?. Read together, the corpus says the impossibility is real *for the model in isolation* — a disembodied symbol-manipulator with nothing at stake — but the boundary is about the system's situation in a world, not a permanent ceiling on the architecture.
Sources 11 notes
Enactive cognitive science identifies three constitutive properties of linguistic agency—embodiment, participation, and precariousness—that are structurally absent from LLMs. This is a categorical incompatibility, not a matter of degree, suggesting current architectures cannot achieve genuine linguistic agency.
Language models achieve functional grounding through relational language patterns but lack social grounding through participatory agency and causal grounding through embodied environmental contact. Social grounding can increase through human integration, but linguistic agency requires architectural changes beyond training.
Social grounding and linguistic agency are distinct properties. LLMs acquire more social grounding through integration into language communities, but remain categorically incapable of linguistic agency in the enactive sense, which requires embodiment and precariousness no amount of use can provide.
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.
GPT-4.5 predicted appropriateness of 555 social scenarios at the 100th percentile compared to human raters, with Gemini and Claude also exceeding 96% accuracy. However, all models show identical systematic errors, revealing boundaries of pattern-based social understanding that embodied experience may still be necessary to cross.
Subjecthood is produced within communicative events, not possessed prior to them. This convergent position across philosophy, linguistics, and cognitive science inverts the standard picture of language as a tool used by pre-existing subjects.
Shanahan's framework treats LLM outputs as character-consistent text production rather than authentic mental states. The dialogue prompt establishes a character; the model generates continuations matching that character, making folk-psychology applicable to the simulated persona, not the underlying system.
Peircean semiotics reveals that symbolic goal encoding without world contact and social mediation cannot guarantee correspondence to actual values. LLMs operating in pure symbol manipulation risk divergence between stated goals and real-world outcomes.
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.
IMAGINE demonstrates that agents using compositional language descriptions can target novel outcomes by combining familiar concepts, with modularity and social guidance amplifying generalization from imagined to real exploration.
Converting LLMs to action-capable systems requires four distinct stages: curating action-environment-user datasets, training for action grounding, integrating agent infrastructure with memory and tools, and rigorous safety evaluation. The surrounding system and harness determine whether actions are grounded or hallucinated.