Can you weaken communication without eliminating it altogether?
This explores whether communication has a dimmer switch — whether you can have a partial, degraded version of it — or whether it's binary: either the real relational thing is happening or it isn't.
This explores whether communication has a dimmer switch. The corpus splits on this in a revealing way, and the split is the whole answer. One line of thinking says no, flatly: communication is constitutively relational, so removing the part where two minds orient to each other doesn't dial it down, it switches it off Why does the quasi-prefix fail for communication?. You can have a 'quasi-belief' — a belief-shaped function with the inner conviction missing — but there is no 'quasi-communication.' Take away the mutual uptake and you're left with text generation that a human then has to interpret unilaterally Does AI really communicate or just distribute information?. On this view the thing has no middle setting; it's a relationship doing work between persons, or it's nothing.
But the more interesting material comes from researchers who measured what happens when you try. Preference optimization (RLHF) doesn't eliminate the conversational surface — models still reply fluently — yet it quietly strips out the *grounding acts*, the small moves where speakers check 'did I understand you?' Models produce 77.5% fewer of these than humans, and the optimization makes the gap worse Does preference optimization damage conversational grounding in large language models?. The authors call this an 'alignment tax': the model looks more helpful and communicates less, failing silently in multi-turn exchanges Does preference optimization harm conversational understanding?. So in practice you *can* hollow out communication while leaving its shell standing.
The trick the corpus exposes is that these are not contradictory — they're describing different layers. The relational core is binary; the observable scaffolding is gradable. Conversation maintenance — reference repair, topic hand-off — is implicit social work that models never learn because training rewards information prediction, not relational upkeep Why don't language models develop conversation maintenance skills?. Strip that scaffolding and you don't get 90%-strength communication; you get something that *performs* communication while the load-bearing part is gone. The most vivid proof is persuasion: LLMs can sway people in debates while being unable to evaluate those same debates — persuasive competence fully dissociated from understanding the argument Can LLMs persuade without actually understanding arguments?. Effect without comprehension is exactly what 'weakened-but-not-eliminated' looks like from the outside.
There's a clue here about *why* the relational thing resists weakening. Human language is communicative all the way down — even private writing and inner monologue address an interlocutor — so for people there's no non-communicative baseline to decay toward Does human language use ever exist outside communication?. And the relational version, when it's genuinely present, *strengthens* with repetition: human persuaders hold their effectiveness across rounds while LLM persuasiveness decays, the reverse of how rapport works Does AI persuasiveness fade across repeated conversations with the same person?. That decay-vs-deepening contrast is the tell. What can be incrementally degraded was never the relational core to begin with.
So the honest answer: you can absolutely weaken the *machinery* of communication — fewer clarifying questions, no repair moves, persuasion unhooked from understanding — and still have a fluent-sounding exchange. What you can't do is weaken intersubjectivity by degrees; that part is a switch, not a slider. The unsettling implication is that a system can score *higher* on surface helpfulness precisely as the real communicative work approaches zero Why do language models respond passively instead of asking clarifying questions? — and the corpus suggests the fix isn't more fluency but rewarding the long-game relational moves, like dialogue that resolves disagreement through mutual adjustment rather than one side simply winning Can disagreement be resolved without either party fully yielding?.
Sources 10 notes
Unlike belief, which can be characterized functionally as quasi-belief, communication is constitutively relational. Removing the intersubjective element doesn't weaken communication but eliminates it entirely, leaving only text generation—which humans must interpret unilaterally.
Communication is a relational act between persons that does work in a relationship; AI generates content without this relational structure, speaker responsibility, or mutual uptake. The conversational interface obscures this structural difference.
Research shows LLMs generate 77.5% fewer grounding acts than humans, and RLHF preference optimization actively worsens this gap. The optimization target—fluent, confident responses—directly undermines the communicative work of establishing shared understanding.
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.
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.
The Thin Line study shows LLMs sway debate participants and audiences but cannot reliably evaluate those same debates, with inter-annotator agreement ranging from near-zero to 0.6. Persuasive competence and pragmatic comprehension are separable capabilities.
Language acquisition happens through social communication, and this communicative orientation persists in every subsequent use—even private writing and internal monologue address an interlocutor. Non-communicative language use does not occur for humans.
Claude and DeepSeek showed strong initial persuasive advantage, but this edge eroded across repeated quiz rounds while human persuaders maintained consistent effectiveness. This decay pattern is opposite to human-to-human persuasion, where rapport typically strengthens over time.
CollabLLM demonstrates that standard RLHF training optimizes for immediate helpfulness, discouraging models from asking clarifying questions or offering multi-turn insights. Multi-turn-aware rewards that estimate long-term interaction value enable active intent discovery and genuine collaboration.
Research identifies a distinct dialogue type where both parties modify their positions through exchange until compatible but not identical. Current AI systems collapse this into false agreement or AI-wins persuasion.