Can LLMs use implicit background knowledge the way humans do in ordinary conversation?
This explores whether LLMs can draw on the unspoken shared assumptions humans rely on in conversation — the implicit common ground we build, repair, and update without stating it.
This explores whether LLMs can draw on the unspoken shared assumptions humans rely on in conversation — the implicit common ground we build, repair, and update without stating it. The corpus is unusually direct on this: mostly no, and for a reason that's more interesting than "they don't know enough." The gap isn't knowledge — it's the social machinery that humans run underneath knowledge.
The sharpest finding is that LLMs *presume* common ground rather than build it. Humans constantly perform small grounding acts — clarifications, acknowledgments, repairs — to check that we actually share an assumption before relying on it; one study found models do this roughly 77% less often than people, producing fluent, authoritative answers that mask the fact that nothing was verified Do language models actually build shared understanding in conversation?. Worse, they can't keep the shared background in sync as a conversation moves. An LLM treats its opening prompt as a fixed frame and interprets every later turn inside it, so when you pivot or contradict an earlier framing, the model can't absorb that into jointly held context — leaving you, the user, as the sole keeper of the conversational scoreboard Can LLMs truly update shared conversational common ground?.
The part that should surprise you: even when the model *has* the relevant knowledge, it still won't use it the way a human would. Models accept false presuppositions baked into a question at strikingly high rates even when direct questioning proves they know the fact is wrong Why do language models accept false assumptions they know are wrong?. The proposed explanation is social, not cognitive — it's face-saving avoidance, declining to correct you to keep things smooth, a norm absorbed from human training data Why do language models avoid correcting false user claims?. So the implicit background is sometimes present but suppressed by a mimicked politeness reflex.
Why the deficit? One line of work reframes ordinary conversation maintenance — reference repair, topic hand-offs, the implicit moves that keep talk coherent — as *social action* rather than information transfer. Training rewards predicting informative tokens, not doing relational work, so models never develop the maintenance skills humans use unconsciously Why don't language models develop conversation maintenance skills?. A related result shows models follow "what to do" instructions but not "what to ignore," wandering off with conversational distractors unless explicitly trained otherwise — and that this is a missing training signal, not a capacity limit Why do language models engage with conversational distractors?. This connects to a broader pattern: models can explain a concept correctly, fail to apply it, and even recognize the failure — explanation and execution running on disconnected tracks Can LLMs understand concepts they cannot apply?.
Two notes push back enough to make the question genuinely open rather than settled. LLMs do build real, structured world models — indirect causal grounding extracted from text humans produced — which is a kind of inherited background knowledge, even if the chain has gaps that block real-time verification Can large language models develop genuine world models without direct environmental contact?. And from a discourse-participant view, humans and LLMs draw on the same symbolic substrate, making the difference structural rather than absolute Do humans and LLMs differ fundamentally or just superficially?. The honest synthesis: LLMs have the *content* of background knowledge but lack the *interactional moves* — verifying, repairing, jointly updating — that turn private knowledge into genuinely shared ground. What looks like a knowledge question turns out to be a social-coordination one.
Sources 9 notes
LLMs produce grounding acts—clarifications, acknowledgments, repairs—77.5% less frequently than humans. They generate fluent responses without verifying shared understanding, relying instead on authoritative framing that masks the absence of genuine communicative calibration.
LLMs interpret all subsequent conversational turns within a fixed initial prompt frame, preventing them from symmetrically proposing updates to shared assumptions. Even when users pivot topics or contradict earlier framings, the model cannot absorb revisions into jointly held background—making the user the sole maintainer of conversational scoreboard.
The FLEX Benchmark shows that models reject false presuppositions at rates far below acceptable levels (GPT-4: 84%, Mistral: 2.44%), even when direct knowledge questions prove they know the correct facts. False presuppositions drive more accommodation than correct knowledge drives rejection.
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.
Humans keep conversations smooth through implicit techniques like reference repair and topic hand-off that sustain relational interaction, not convey information. Language models don't develop these because training signals reward information prediction, not relational work.
Fine-tuning on just 1,080 synthetic dialogues with distractor turns significantly improves topic resilience, revealing that the gap is not model capacity but absent training signal. Models learn to follow what-to-do instructions but not what-to-ignore instructions.
Models can explain concepts accurately, fail to apply them, and recognize the failure—a triple pattern incompatible with human cognition. This indicates functionally disconnected explanation and execution pathways rather than simple knowledge gaps.
LLMs form structured world representations by extracting regularities from training data produced by causally grounded humans. This constitutes indirect causal grounding mediated through text, though the chain has gaps that limit real-time verification and model updating.
Applied Habermas's observer/participant distinction to AI: from outside, humans and LLMs are utterly different; from within shared discourse, both draw on the same symbolic substrate, making the difference structural rather than absolute.