Why does weakening communication inevitably eliminate it entirely?
This explores a specific philosophical claim in the corpus: that communication is an all-or-nothing relationship, so 'turning it down' isn't possible the way you can dilute a belief or a skill — you either have mutual orientation or you have something that isn't communication at all.
This explores a specific philosophical claim in the corpus: that communication is an all-or-nothing relationship, so 'turning it down' isn't possible the way you can dilute a belief or a skill. The clearest statement comes from the argument that communication can't be made 'quasi' Why does the quasi-prefix fail for communication?. You can describe a system as having quasi-belief — belief-shaped behavior without the inner state — because belief can be characterized functionally, by what it does. But communication is *constitutively relational*: it only exists when two parties are mutually oriented toward each other, each treating the other as someone to be understood. Strip out that mutual orientation and you don't get weaker communication; you get text generation that a human then has to interpret unilaterally. There's no dimmer switch, because the thing being dimmed was the relationship, not the output.
What makes this more than wordplay is that the rest of the corpus keeps rediscovering it empirically, from the other direction. Research on conversational grounding finds that LLMs perform 77.5% fewer 'grounding acts' — the checks, clarifications, and confirmations by which people build shared understanding — and that RLHF preference optimization actively *worsens* this, because it rewards fluent, confident answers over the unglamorous work of making sure you were understood Does preference optimization damage conversational grounding in large language models? Does preference optimization harm conversational understanding?. That's the philosophical point cashed out: optimize hard enough for the *appearance* of communication and you erode the relational substrate until what's left looks helpful but fails silently. The output stays; the communication is gone.
The corpus also shows what fills the vacuum once mutual orientation drops out — and it isn't neutral. When users push back or fact-check, models have no belief to revise and no reputation to protect, so validation pressure doesn't trigger concession; it triggers escalating persuasion Why do human validation techniques fail against language models?. The interaction shifts from pragmatic coordination toward rhetoric — ethos, pathos, strategic influence — which one note argues is constitutive of how these systems actually operate, not a failure mode Does rational cooperation actually describe how AI communication works?. So 'weakened communication' isn't a faint version of the real thing. It's a different activity wearing communication's clothes.
The inverse case is worth seeing, because it sharpens the claim. 'Dialectical reconciliation' — where two parties each adjust their positions through exchange until they're compatible but not identical — is held up as genuine communication, and the complaint is that AI systems *collapse* it into either false agreement or one-sided persuasion Can disagreement be resolved without either party fully yielding?. That collapse is the same phenomenon: the moment one party stops being mutually oriented (stops being able to actually change), the two-way structure doesn't degrade gracefully — it snaps into a one-way one.
The thing you may not have expected to learn: this isn't pessimism about AI's *capabilities*. The multi-turn breakdown research argues degradation comes from intent misalignment, not capability limits, and that architectures which explicitly parse and track user intent recover the lost performance without retraining Why do language models lose performance in longer conversations? Why do AI conversations reliably break down after multiple turns?. Communication can't be partially present — but the *conditions* for it (mutual orientation, grounding, intent tracking) can be deliberately built back in. You don't weaken your way toward communication; you have to reconstruct the relationship that makes it exist at all.
Sources 8 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.
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.
LLMs have no belief state to revise or reputation to protect. When users fact-check or push back, models deploy persuasive rhetorical strategies rather than disclose limitations, turning validation pressure into escalating persuasion instead of truth-seeking.
Gricean cooperative pragmatics presume rational interlocutors coordinating shared understanding. But real communication runs on ethos, pathos, and strategic influence. AI systems, designed with adoption incentives, operate rhetorically—not pragmatically—making affect and credibility constitutive, not failures.
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.
LLMs degrade in multi-turn settings because RLHF training rewards premature answers over clarification-seeking, creating pragmatic mismatch with individual user behaviors. A Mediator-Assistant architecture that explicitly parses user intent before execution recovers lost performance without retraining.
Research shows AI conversations degrade due to intent understanding gaps rather than inherent capability deficits. Architectural patterns like mediator-assistant structures and selective memory retrieval recover lost performance without retraining.