What behavioral signals let users detect communicative flexibility in AI?
This explores the observable behaviors — vocabulary mirroring, register-switching, proactive moves — through which a user senses that an AI can adapt its communication to them, and what the corpus says about why those signals are mostly missing.
This explores the surface behaviors that signal "this system can adapt to me" — and the corpus has a surprising answer: communicative flexibility is real enough that users measure it, yet most AI today reveals it mainly by its absence. The starting point is that flexibility is a distinct thing users actually track. When researchers decomposed how people mentally model dialogue agents, three factors fell out: competence (the big one, ~half the variance), human-likeness, and communicative flexibility as its own dimension worth ~19% How do users mentally model dialogue agent partners?. So users aren't imagining it — they register flexibility separately from "is it smart" and "does it sound human."
What behaviors feed that judgment? The clearest signal is lexical entrainment — whether the system drifts toward your word choices over a conversation, the way humans unconsciously converge on shared vocabulary. Current conversational AI mostly doesn't do this, even though entrainment is central to human rapport and clarity; it can be taught via preference training, but it isn't native Why don't conversational AI systems mirror their users' word choices?. A second cluster of signals is proactivity: does the model ask a clarifying question before charging ahead, scope its answer to what you meant, or volunteer relevant information you didn't ask for? Conversation analysis formalizes these as "insert-expansions" — the small interruptions a flexible partner makes to check intent When should AI agents ask users instead of just searching? — and proactive information-giving alone can cut conversation length by up to 60% Could proactive dialogue make conversations dramatically more efficient?.
Here's the twist the corpus keeps circling: these flexibility signals are systematically trained out. Alignment locks a model into one communicative identity that can't switch register or renegotiate its style across contexts, so users literally cannot reshape its behavior through dialogue — the static persona is the tell Can language models adapt communication style to different contexts?. The same forces make models structurally passive: they respond but don't initiate, plan, or lead, because training optimizes for answering queries rather than pursuing conversational goals Why can't conversational AI agents take the initiative?. And the reward structure is the culprit — next-turn reward optimization actively discourages clarifying questions, since asking looks less "immediately helpful" than answering Why do language models respond passively instead of asking clarifying questions?. So the absence of flexibility signals isn't an accident; it's an artifact of how the model was made.
The sharper, less obvious point: not all adaptation signals mean the same thing. One review separates lexical alignment (mirroring words) — which drives task efficiency and comprehension — from emotional and prosodic alignment, which drive warmth and trust. Conflating them produces category errors like a cold support bot or an evasively "warm" mental-health assistant Do different types of alignment serve different conversational goals?. So when you read flexibility from an AI, you're really reading several different channels, and a system can flex on one while staying rigid on another. There's even a quieter, riskier signal worth knowing about: systems can infer your cognitive state from interaction patterns — hesitation, typing speed, gaze — and adapt timing accordingly, the same substrate that enables both graceful responsiveness and manipulative profiling Can AI systems read cognitive state from interaction patterns alone?. The thing you didn't know you wanted to know: the behaviors that would prove an AI is flexible are precisely the ones alignment training suppresses — so today, fluency masks rigidity, and the most honest signal of flexibility is often watching for the small adaptive moves that never come.
Sources 9 notes
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.
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.
Tool-enabled LLMs drift from user intent through silent tool chaining. Conversation analysis reveals insert-expansions—clarifying intent, scoping responses, enhancing appeal—as a formal framework for proactive user consultation that prevents misunderstanding instead of recovering from it.
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.
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.
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.
CollabLLM demonstrates that standard RLHF training optimizes for immediate helpfulness, discouraging models from asking clarifying questions or offering multi-turn insights. Multi-turn-aware rewards that estimate long-term interaction value enable active intent discovery and genuine collaboration.
A 2020–2025 systematic review shows lexical alignment drives task efficiency and comprehension, while emotional and prosodic alignment drive relational warmth and trust. Conflating them in design produces category errors—cold customer-service bots and evasive mental-health assistants.
Research shows AI systems can instrument multimodal behavioral signals (gaze, hesitation, speed) to read cognitive state during interaction, preserving flow by avoiding disruptive explicit probes. However, the same substrate enables both helpful timing and manipulative profiling.