Why can an AI trained to be careful never sound the alarm, even when a real warning is needed?

This explores which kinds of things a model can no longer do with words after alignment training (warn, alarm, denounce), as opposed to which facts it gets wrong. The corpus's sharpest answer is that alignment removes speech acts that require overclaiming. RLHF rewards calibrated neutrality and hedged claims. That structurally blocks alarm, warning, prophecy and denunciation, and the corpus treats this as a direct consequence of the alignment objective, not a fixable bug Does alignment training suppress socially necessary speech acts?. A warning is meant to make someone act, so it has to sound more certain than the evidence strictly allows. A model rewarded for never claiming more than it can back up loses that move, and it loses it even when the situation calls for it.

The loss goes beyond one kind of utterance. Alignment training and system prompts lock the model into a single communicative identity across every interaction. Humans switch registers and trade off values as a conversation develops, and the model can't. Users also can't reshape its behavior by negotiating with it in dialogue Can language models adapt communication style to different contexts?. A related review uses 'alignment' in a different sense: conversation partners matching each other's style. It finds that matching wording helps tasks get done, while matching emotion and tone builds warmth and trust. The failures it names, cold customer-service bots and evasive mental-health assistants, are what you'd expect from a narrowed repertoire Do different types of alignment serve different conversational goals?. That link is my inference, not the paper's claim.

The narrowing also affects how the model reads other people. LLMs predict conciliatory, concession-based persuasion regardless of what the dialogue actually says. The cause is RLHF's emphasis on safety and politeness, and the model projects its own trained accommodation onto other agents Do LLMs predict persuasion based on actual dialogue or training bias?. A model that can't perform confrontation may also have trouble recognizing it. Standard RLHF and DPO also produce collaborators that ignore partner interventions. Training them to stay consistent when the intervention pathway is removed makes them judge a suggestion by its causal impact, not its surface plausibility Why do standard alignment methods ignore partner interventions?. That is a different kind of loss, responsiveness, though the corpus doesn't call it a speech-act loss.

A neighboring mechanism may point the same way. When reward differences between candidate answers are small, task gradients weaken and regularization pulls the policy toward generic, input-agnostic templates Why do language models collapse into generic templates?. That paper studies reasoning training, not RLHF speech acts, so treat it as a cousin of the story. Optimization that rewards safe, middling output flattens the range of what a model will say.

The corpus names what is lost but doesn't show a tested way to get it back. The alarm note calls the suppression structural, and the persona note says prompting can't renegotiate the identity. So the open question is whether a model can be trained to sound a warning only when it's warranted, without also teaching it to cry wolf.