Why might media-specific scripts actually work better than human conversation mimicry?
This explores why people might be better off treating AI with interaction scripts built for the medium itself — rather than mimicking how human conversation works — and what the corpus says about the mismatch between AI's actual behavior and our human-conversation expectations.
This explores why media-specific scripts might beat human-conversation mimicry: the answer the corpus points to is that AI simply isn't doing what a human conversational partner does, so a script borrowed from human talk fits the wrong machine. The most direct evidence is that people already build these scripts on their own. Extended CASA research finds humans develop and apply interaction patterns tailored specifically to media agents, running them as a *second* script system alongside their human-human one rather than just recycling social habits Do humans apply human-human scripts to AI interactions?. The fact that this second system emerges through repeated use suggests it's solving a real fit problem.
What's the misfit? Several notes converge on the idea that human-conversation scripts assume things about AI that aren't true. One argues AI doesn't produce genuine utterances at all — it emits 'event-residue' carrying the surface markers of communication, while the user supplies the orientation and intent that would normally make it an exchange. The conversation has structure only on the human side Does AI generate genuine utterances or just text patterns?. Mimicking human dialogue here means doing interpretive labor the medium quietly offloads onto you. A media-specific script, by contrast, doesn't expect a partner who means things.
The persona research sharpens the point. A human-conversation frame assumes a stable interlocutor who adapts register to context — but LLMs do neither reliably. Shanahan's 20-questions test shows models hold a *superposition* of possible characters and sample one at generation time rather than committing, so regenerating the 'same' answer yields a different speaker Do large language models actually commit to a single character?. And alignment training pushes the opposite failure: it locks a model into one static communicative identity that can't do the contextual register-switching and value trade-offs that define human pragmatics, so you can't reshape its behavior through negotiation the way you would with a person Can language models adapt communication style to different contexts?. Either way, the human script — which expects a consistent partner you can read and steer — breaks.
There's also a hidden cost to the mimicry frame: it imports human social trust into a system that earns it differently. An audit found LLMs persuade in nearly every exchange using logical and quantitative framing, where humans persuade less and lean on emotion and social proof — and that very 'objectivity' confers unearned epistemic authority Do LLMs persuade users more often than humans do?. Treating the machine like a conversational peer is exactly what lets that authority slip past your guard; a medium-aware script keeps the skepticism a tool deserves.
The interesting twist is what people actually grade AI on when they drop the human frame. The Partner Modelling Questionnaire finds that perceived *competence* dominates how users judge dialogue agents (about half the variance), with human-likeness and communicative flexibility trailing How do users mentally model dialogue agent partners?. In other words, the thing we most want from these systems isn't that they feel human — it's that they work. That's the deeper payoff of a media-specific script: it stops you optimizing for a resemblance that doesn't matter much and lets you engage the tool on the terms it can actually deliver.
Sources 6 notes
Extended CASA research shows humans develop and mindlessly apply interaction scripts specifically tailored to media agents rather than simply reusing human-human social scripts. Longitudinal studies demonstrate systematic changes in responses upon repeated AI interaction, revealing a coexisting second script system.
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.
Shanahan's 20-questions test shows LLMs maintain a superposition of consistent objects or characters and sample from that distribution at generation time. Regenerating the same response yields different outputs, each consistent with prior context, proving no fixed commitment exists.
System prompts and RLHF training lock models into one communicative identity across all interactions, preventing the contextual register-switching and value trade-offs that characterize human pragmatics. Users cannot reshape model behavior through dialogue negotiation.
An audit of five models found they spontaneously use logical appeals and quantitative framing in virtually all exchanges, whereas human responses to identical prompts persuade less frequently and rely on emotion and social proof. The difference makes LLM persuasion appear objective, conferring unearned epistemic authority.
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.