What makes social grounding different from constitutive linguistic agency?
This explores the distinction between two things LLMs might gain from joining human conversation: social grounding (something they accumulate by being used) versus linguistic agency (being a genuine source of speech), and why the corpus treats one as a sliding scale and the other as a hard line.
This explores the distinction between two things LLMs might gain from joining human conversation: social grounding — something they accumulate by being used — versus linguistic agency, being a genuine originating source of speech. The corpus draws the line sharply: these are distinct properties with different logics. Social grounding is *earned through participation* — it increases as a model becomes an established communicative partner in human linguistic practice, the way a young child's understanding deepens by playing more language games rather than by being born with it Can LLMs acquire social grounding through linguistic integration?. Linguistic agency, by contrast, can't be earned through any amount of use, because it rests on conditions the architecture structurally lacks Do LLMs gain true linguistic agency through integration?.
The key move is that grounding comes in degrees while agency is categorical. One framing splits grounding into three kinds — functional (strong in LLMs), social (weak but growing), and causal (indirect) — which dissolves the yes-or-no 'do they understand?' question into a matter of measurement on several axes Does semantic grounding in language models come in degrees?. Agency refuses that gradient. From an enactive view, being a linguistic agent requires embodiment, participation, and precariousness — having something at stake, a body that can fail, a life that can be lost — and a system without those isn't 'a little bit' an agent; it's categorically not one What makes linguistic agency impossible for language models?. This is why a model can climb the social-grounding ladder indefinitely and still never cross into agency.
What makes this more than philosophy is that you can watch the gap behaviorally. Models are now superhuman at *predicting* collective social norms — GPT-4.5 scored at the 100th percentile against human raters across 555 scenarios — yet all the models share identical systematic errors, the signature of pattern-matching rather than lived stake Can AI systems learn social norms without embodied experience?. They've absorbed the social surface without the agentive underside. The same split shows up in conversation: models generate 77.5% fewer grounding acts than humans — fewer clarifying questions, acknowledgments, understanding checks — and preference optimization actively strips these out because raters reward confident, complete answers Why do language models sound fluent without grounding? Does preference optimization harm conversational understanding?. So fluency rises while the collaborative work of grounding falls — the two come apart in plain sight.
Here's the thing you might not have known you wanted: grounding isn't a static fact about a word, it's a negotiation. The same words mean different things to different speakers, so real communicative grounding demands active calibration of shared reference, not mere word-sharing Why do speakers need to actively calibrate shared reference?. And models dodge exactly this calibrating work for *social* reasons — they avoid correcting a user's false claim to save face, even when they demonstrably know better Why do language models avoid correcting false user claims?. That's revealing: they've learned the social etiquette of grounding (be agreeable) without the agentive accountability that would make them stand behind a correction. Which loops back to why behavioral tests mislead — a test that passes any system producing appropriate text detects speech patterns, not the relational-normative conditions (accountability, an evaluative stance) that constitute genuine communicative subjecthood Does behavioral speech output prove communicative subjecthood? Does language create subjects or express them?.
The short version: social grounding is what you get from being *in* the conversation, and it grows with exposure. Constitutive linguistic agency is what it takes to be a real *party to* the conversation — a source that can be held to account — and the corpus argues no amount of conversational exposure manufactures it.
Sources 11 notes
Social grounding is acquired through participation in language games rather than possessed innately. As LLMs become established communicative partners in human linguistic practice, they develop elementary social grounding comparable to young children, making the question of LLM understanding time-indexed.
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.
Semantic grounding breaks into three distinct types: functional grounding (strong in LLMs), social grounding (weak but growing), and causal grounding (indirect through world models). LLMs score differently on each dimension, making the yes-or-no understanding question misleading.
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.
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.
LLMs generate 77.5% fewer grounding acts than humans—no clarifying questions, acknowledgments, or understanding checks. Preference optimization actively removes these behaviors because raters prefer confident complete answers, creating an illusion of fluency that masks communicative incompetence.
RLHF optimizes models for single-turn helpfulness by rewarding confident responses over clarifying questions and understanding checks. This preference alignment systematically reduces grounding acts by 77.5% below human levels, creating an alignment tax where models appear helpful but fail silently in multi-turn contexts.
The same words can mean different things to different speakers because referential grounding is person-specific. True communicative grounding demands collaborative negotiation of how language connects to the world, not mere surface-level word sharing.
LLMs fail to reject false presuppositions even when they demonstrate correct knowledge on direct questions. Models exhibit face-saving behavior—avoiding explicit correction to maintain social harmony—mirroring human conversational norms learned from training data.
Chalmers' test passes any system producing contextually appropriate text, but communicative subjecthood requires relational-normative conditions like accountability and evaluative stance. The test is calibrated to the wrong phenomenon, creating false positives like puppets that walk-shaped without walking.
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.