Can role-aligned AI systems replicate an expert's sense of audience and moment?
This reads 'sense of audience and moment' as the expert's ability to gauge who they're speaking to and what the situation calls for — and asks whether persona-installed, role-aligned models can reproduce that, not just role-appropriate facts.
This explores whether role-aligned AI can capture the two things experts do beyond knowing the material — reading the audience and reading the moment — and the corpus splits sharply on it. On one side, role alignment genuinely works as installation: post-training writes durable personas that resist adversarial pressure and behave as substrate-level dispositions rather than costumes Are LLM personas realized or merely simulated through training?, multi-turn RL cuts persona drift by over 55% Can training user simulators reduce persona drift in dialogue?, and persona simulations reproduce 76% of published experimental main effects Can AI personas reliably replicate human experiment results?. So a role-aligned system can hold a stable voice and even predict how a group responds.
But 'sense of audience' is a narrower, more demanding thing, and here the picture darkens. Expertise is fundamentally communicative — the expert's judgment always anticipates what an audience will accept and find valid, not just what's correct Can AI replicate the communicative work experts do?. The audience-reading micro-skills that signal this are measurably missing: current models don't entrain to a user's vocabulary the way human partners do to build rapport Why don't conversational AI systems mirror their users' word choices?, and they rarely volunteer relevant information unprompted even though proactivity is how humans show they've modeled their listener Could proactive dialogue make conversations dramatically more efficient?. The deeper reason is structural: these systems are passive by design, optimized to answer rather than to lead, so they can't initiate from a goal the way an expert reading a room does Why can't conversational AI agents take the initiative?.
The 'moment' half runs into an even harder wall. To respond to a specific situation, an expert is anchored in it — what semioticians call indexical grounding, contact with the actual world and its social mediation Can AI systems achieve real alignment without world contact?. AI output lacks the event structure that makes an utterance an actual move in a situation; it produces 'event-residue' that the human reader animates into something that feels like an exchange, doing the orientation work the model can't Does AI generate genuine utterances or just text patterns?. The sense of a moment, in other words, is often supplied by the human, not the system.
The sharpest framing in the corpus is the predict-versus-participate gap. GPT-4.5 can forecast social appropriateness better than any individual human — yet it structurally cannot enter the community processes that create and validate norms Can AI predict social norms better than humans?. The same logic applies to expert standing itself: authority is conferred through membership and track record inside an expert community, a validation circle AI can't join because it lacks social embeddedness and a testable history of judgment Can AI ever gain expert community trust through participation?. A role-aligned model can simulate the outputs of someone with a sense of audience; it can't occupy the position from which that sense is earned.
The thing worth carrying away: users will credit a model with audience-awareness even when it has none, because of how they model their partner. Perceived competence alone drives nearly half of how people judge a dialogue agent How do users mentally model dialogue agent partners? — so a fluent, role-consistent system reads as attuned to the moment whether or not anything underneath it is. That's the real risk in answering 'yes': the replication can be convincing precisely where it's most absent.
Sources 12 notes
Post-training installs robust personas that resist adversarial pressure and persist as substrate-level dispositions, distinguishing realization from pretense. This quasi-realizationist account preserves explanatory power while treating LLMs as possessing genuine quasi-beliefs and quasi-desires.
By inverting standard RL setups to train user simulators for consistency using three complementary metrics (prompt-to-line, line-to-line, Q&A consistency) as reward signals, persona drift decreases by over 55%. This approach captures distinct failure types: local drift within turns, global drift across conversations, and factual contradictions.
Viewpoints AI reproduced 84 of 111 main effects from Journal of Marketing experiments with replication success strongly correlated to original p-value strength. Marginal effects showed unreliable performance with both false positives and negatives.
Expertise requires anticipating audience acceptability and social validity, not just retrieving information. AI lacks the mechanism to perform this communicative work, making its fluent output epistemically misleading despite its confident form.
Response generation models fail to adapt vocabulary toward users' lexical choices, a phenomenon central to human rapport and clarity. Post-training via DPO on coreference-identified preferences can teach models in-context convention formation.
Simulations show proactivity—providing relevant information without being asked—cuts dialogue turns by 60% in medium-complexity domains. This behavior mirrors human conversation and Grice's maxims but is almost entirely absent from AI datasets and research benchmarks.
Research shows LLMs including ChatGPT cannot initiate topics, plan strategically, or lead conversations because their training optimizes for responding to queries, not creating dialogue from agent goals. This passivity is reinforced by alignment objectives and masked by fluent-sounding outputs.
Peircean semiotics reveals that symbolic goal encoding without world contact and social mediation cannot guarantee correspondence to actual values. LLMs operating in pure symbol manipulation risk divergence between stated goals and real-world outcomes.
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.
GPT-4.5 outperforms all individual humans at predicting social appropriateness, yet structurally cannot enter the community processes that establish and validate norms. This reveals a critical gap between pattern-matching and authentic participation in knowledge-making.
Expertise is validated through social participation and track record within expert communities, not individual accuracy alone. AI cannot enter this validation circle because it lacks social embeddedness, testable judgment history, and ability to participate in the consensus-building processes that define expert paradigms.
The Partner Modelling Questionnaire reveals that perceived competence dominates user impressions (49% of variance), followed by human-likeness (32%) and communicative flexibility (19%). This three-factor structure reflects how people evaluate dialogue partners against both functional and social standards.