Can language models actually raise alarm about threats?

Alarm is a peculiar speech act. The informational content is often minimal — "danger," "fire," "stop." What does the work is the addressing: someone is reaching for the listener's attention, claiming priority, asserting that this matters now. Strip the addressing and the content becomes inert. The envelope is the message; the message-as-information barely exists.

This makes alarm fundamentally interpersonal. It is addressed to specific people in a specific moment by a specific source whose authority to raise an alarm is part of what makes the alarm function. The person raising the alarm is staking themselves on it — claiming that this rises to the level of warranted concern. The receiver attends partly because of the alarm-content but largely because of the alarm-source: someone competent took this seriously enough to address them.

LLMs cannot perform this speech act, for three structural reasons. First, LLMs do not feel concern. They cannot be alarmed about anything because there is no internal state of alarm to express. Whatever alarm-shaped output an LLM produces is mimicry, not expression. Second, LLMs cannot appeal to attention in the interpersonal sense. The output is generated in response to a prompt; it is not a reaching-for-attention from a source to a receiver. The attention that consumes the output is supplied by the prompter, not solicited by the LLM. Third, LLMs are reactive. Alarms are proactive — someone notices a threat and raises the alarm without being prompted. LLMs do not notice threats and do not generate without prompting; they cannot produce the unprompted address that alarm requires.

There is a fourth, training-side reason. Alarm-phrasing — direct, urgent, authoritative — runs counter to the calibration RLHF and alignment training enforce. Models are trained toward hedged, qualified, neutral output that satisfies users across contexts. A model trained to never overclaim cannot raise alarm, because alarm is overclaim relative to a baseline of calm description. The alignment that makes models socially acceptable in most contexts makes them constitutively unable to perform alarm.

The implication for AI in information ecosystems: AI is structurally unable to take on the social function alarm performs. In journalism, expert commentary, public health, civic life, alarm has historically been a way that authoritative sources alert publics to threats requiring response. AI cannot do this work — not because it lacks information, but because the speech act requires what AI structurally cannot do. Public information ecosystems that rely on AI for analysis will need to preserve human alarm-raisers explicitly, because the AI will not produce alarms even when warranted.

The strongest counterargument: AI can produce alarming-sounding text when prompted to summarize alarming information. True, but the alarm-act in such cases is performed by the prompter (selecting the alarming framing) and the receiver (treating the output as warning). The AI itself remains unable to raise alarm; it is being used as a content-channel for an alarm that some human is raising through it.

Source: LLMs don't get alarmed