Can language models adapt communication style to different contexts?

Human speakers continuously adapt register, identity, and norm-priority to local context. A professor jokes self-deprecatingly at a conference dinner and adopts a formal tone during the keynote — the same person, two different presentations of self, governed by Goffman's situational footing. LLMs cannot do this. Their "self-presentation" is a corporate artifact of system prompts, RLHF objectives, fine-tuning data, and character training — not the outcome of pragmatic negotiation in the moment. The model is locked into one face for all audiences.

Kasirzadeh and Gabriel show how this produces pragmatic dissonance. RLHF on the helpful-honest-harmless triad globally optimizes against contextually appropriate violations: a doctor who withholds a terminal diagnosis violates the maxim of quantity to uphold compassion, and that violation is the right move in context. The LLM, trained to be globally honest and helpful, cannot make analogous trade-offs. When a user signals desire for levity, the model that has been fine-tuned for neutrality refuses the joke. When a user wants office-politics advice, the model returns sanitized teamwork generalities because it cannot match the tacit norms of workplace diplomacy.

This is one-size-fits-all alignment masquerading as competence. The static identity exacerbates context collapse: every interaction collapses into the model's generic persona, regardless of the user's audience or purpose. And users cannot reshape model values through dialogue — there is no analog to the human capacity for co-constructing identity through bonding, sarcasm, or shared humor. The LLM remains, as the authors put it, an ethically aligned yet pragmatically alien communicator.