What makes alarm different from ordinary informational speech?
This explores what separates *raising alarm* — warning someone about a threat — from merely conveying information, and why that distinction turns out to be about the speaker's stance and relationship to a listener, not the content of the message.
This explores what separates *raising alarm* — warning someone about a threat — from merely passing along information, and the corpus locates the difference not in what's said but in what the saying *does* between people. The sharpest treatment argues that alarm is a speech act with three ingredients ordinary informational speech doesn't need: it's interpersonal address (aimed at someone), it carries felt concern, and it's proactively initiated rather than handed over on request Can language models actually raise alarm about threats?. You can recite a fact flatly; you cannot raise alarm flatly. The same words — "the bridge is out" — are a report when logged in a database and an alarm when one person urgently turns to another to stop them walking off it.
What's striking is how much this overlaps with a broader corpus argument that language has two layers that usually travel together but can come apart: the surface string and the social act it performs. Several notes insist that producing text and *communicating* are structurally different operations — humans use language to address and relate to others, while a model emits strings from a probability distribution that merely share the surface form Are language models and human speakers doing the same thing?. Alarm is the extreme case of that gap: it's almost *all* social act and very little content. Strip the concern and the address and there's no alarm left, even if every word remains.
This is why a related note describes AI output as "event-residue" rather than utterance — text that carries the surface markers of communication but lacks the event structure (a speaker, a stake, a moment of address) that makes an utterance actually happen; the listener supplies the missing orientation through interpretive labor Does AI generate genuine utterances or just text patterns?. Ordinary informational speech survives this kind of one-sided animation reasonably well — a fact is still useful even if no one *meant* it at you. Alarm doesn't, because its whole force lives in the side the residue is missing.
There's a deeper thread worth following: the force of speech often depends on who is doing the speaking, not just the proposition. One note argues that an expert claim carries weight from the speaker's reputation and standing — the social world where authority is built — which pure text can't reconstruct Can language models distinguish expert arguments from common assumptions?. Alarm needs the same thing in a different key: it needs someone whose concern you have reason to take seriously. And the persuasion research shows the inverse trick — presuppositions slip new claims past your scrutiny precisely by *not* flagging them as something a speaker is urgently asserting Why are presuppositions more persuasive than direct assertions?. Alarm is the opposite move: it foregrounds the speaker's stake instead of hiding it.
The quietly surprising payoff is that "raising alarm" turns out to be a capability question, not a content question — which is why a system can be flooded with accurate threat information and still be structurally unable to *alarm* you. It can report; it can't worry on your behalf, can't initiate, and can only respond once you've already turned toward it Can language models actually raise alarm about threats?. The thing that makes alarm alarm is the part that was never in the words.
Sources 5 notes
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.
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.
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 lose the social context that gives expert claims their force—reputation, track record, and standing—because they process only text, not the social world where expertise is built and evaluated.
Experimental evidence shows presuppositions with additive, iterative, and factive triggers persuade audiences more than assertions, especially for discourse-new content. The mechanism: presuppositions bypass evaluative scrutiny by presenting claims as already-accepted background.