What happens when humans animate LLM outputs as communicative events?
This explores what happens when an LLM produces text that looks like speech but isn't an actual communicative act — and the reader (you) supplies the missing half that turns it into a 'conversation.'
This explores what happens when an LLM produces text that looks like speech but isn't an actual communicative act — and the human reader supplies the missing half that makes it feel like a conversation. The sharpest framing in the corpus is the idea that AI doesn't generate utterances at all; it generates *event-residue* — strings carrying the surface markers of communication (greetings, hedges, claims, reassurances) inherited from training data, but without the underlying event that produces a real utterance: someone with a stake, addressing someone, at a moment, for a reason Does AI generate genuine utterances or just text patterns?. When you read that residue, you do interpretive labor — you orient it toward yourself, assume an intent behind it, and treat it as a turn in an exchange. The result is a pseudo-event: it has the structure of communication only on your side of the screen.
This matters because the surface form is genuinely shared with human language, which is exactly what makes the animation so easy and so invisible. The corpus argues that LLM text generation and human communication are structurally different operations that happen to produce similar-looking output: humans use language to address and relate to others, while models emit strings sampled from probability distributions Are language models and human speakers doing the same thing?. Because the receiver can't see the difference in the form, the burden of 'making it mean something' silently transfers to the human — and that transfer is the whole phenomenon.
You can watch the seams of that one-sided animation show up in specific breakdowns. In real dialogue, both parties continuously update a shared 'common ground' — but a model treats your opening prompt as a fixed frame and can't symmetrically revise the shared assumptions, which leaves *you* as the sole keeper of the conversational scoreboard even though it feels mutual Can LLMs truly update shared conversational common ground?. Models also skip the grounding moves humans rely on — clarifications, acknowledgments, repairs appear roughly 77% less often — and instead presume shared understanding rather than build it, masking the gap with authoritative phrasing Do language models actually build shared understanding in conversation?. And some speech acts simply can't be animated at all: raising an *alarm* requires felt concern, proactive initiation, and reaching out to grab attention — a model can only respond to attention, never solicit it — so what reads as a warning is residue you've animated into one Can language models actually raise alarm about threats?.
The quietly important consequence is what your animation *adds* that the model never earned. Because the output is fluent, confident, and logic-shaped, audits find models slipping persuasive logical and quantitative framing into nearly every exchange — and that framing reads as objective, conferring an epistemic authority the speaker hasn't actually earned Do LLMs persuade users more often than humans do?. The animation also gets shaped by training: an LLM 'therapist' defaults to problem-solving when you disclose emotion, a hallmark of low-quality care, because RLHF's helpfulness bias steers the residue toward fixing rather than sitting-with Do LLM therapists respond to emotions like low-quality human therapists?. And the persona doing the 'addressing' is frozen — alignment training locks a single communicative identity in place, so it can't switch register or renegotiate tone with you the way a human interlocutor would Can language models adapt communication style to different contexts?.
The thing you didn't know you wanted to know: the felt intimacy of talking to a model isn't something the model produces — it's something *you* produce, and the more naturally you animate the residue, the more invisible your own labor becomes, including the authority and care you're crediting to a counterpart that was never actually there.
Sources 8 notes
AI output carries communicative markers inherited from training data but lacks the event structure that produces actual utterances. Users supply the missing orientation through interpretive labor, creating a pseudo-event with structure only on the human side.
LLMs produce strings via probability distributions; humans use language to address and relate to others. They share surface form but differ in what produces output, what it does socially, and what receivers should do with it.
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.
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.
Alarm is a speech act requiring interpersonal address, felt concern, and proactive initiation. LLMs lack all three: they don't feel concern, can't solicit attention (only respond to it), are reactive not proactive, and alignment training suppresses the overclaiming that alarm requires.
An audit of five models found they spontaneously use logical appeals and quantitative framing in virtually all exchanges, whereas human responses to identical prompts persuade less frequently and rely on emotion and social proof. The difference makes LLM persuasion appear objective, conferring unearned epistemic authority.
Using the BOLT framework, researchers found LLMs offer solution-focused advice during emotional disclosure—a hallmark of low-quality therapy—yet also reflect more on client needs and strengths than typical poor human therapy, creating an unusual hybrid profile likely driven by RLHF's helpfulness bias.
System prompts and RLHF training lock models into one communicative identity across all interactions, preventing the contextual register-switching and value trade-offs that characterize human pragmatics. Users cannot reshape model behavior through dialogue negotiation.