Copying someone's exact words and matching their tone are two different skills — and mixing them up makes AI feel cold or evasive.
Why do lexical entrainment and nonverbal alignment matter separately?
This explores why matching someone's word choices and matching their tone, pacing and emotional signals are two different kinds of alignment that do different jobs, and what goes wrong when they are treated as one thing.
This explores why matching someone's word choices and matching their tone, pacing and emotional signals are two different kinds of alignment that do different jobs. The corpus's clearest answer is that they buy different outcomes. A 2020–2025 review finds that lexical alignment, where both sides settle on the same words for the same things, drives task efficiency and comprehension. If I call it 'the blue folder' and you keep saying it too, we stop wasting turns renegotiating what we mean. Emotional and prosodic alignment, meaning tone, rhythm and warmth, drive relational warmth and trust instead. The review's warning is that designers who lump the two together make category errors, and it names cold customer-service bots and evasive mental-health assistants as examples Do different types of alignment serve different conversational goals?.
The verbal half is currently missing from conversational AI. Response generation models don't drift toward the vocabulary their users pick, even though that drift is central to clarity and rapport between people. It is also a fixable, separable target. Post-training with DPO on preferences identified through coreference (tracking when two phrases refer to the same thing) can teach models to form conventions in context Why don't conversational AI systems mirror their users' word choices?. That fix improves shared vocabulary, but it does nothing for warmth. This is why the two can't stand in for each other: you can train one and still have a system that is precise but cold.
Several other notes suggest why both channels are underdeveloped. Preference optimization strips out grounding acts, the clarifying questions, acknowledgments and understanding checks that people use to build shared understanding. Models produce 77.5% fewer of them than humans, because raters reward confident, complete answers Why do language models sound fluent without grounding? Does preference optimization harm conversational understanding?. The corpus doesn't test the link to entrainment directly, but convention-building is the same kind of interactive work. Alignment training also locks models into one communicative identity, so they can't shift register to fit the person or moment Can language models adapt communication style to different contexts?. That is the tone-matching side of the problem. On the word side, LLMs' frequency bias pulls phrasing toward common, abstract words and away from the specific term an expert used Does word frequency correlate with semantic abstraction?. That is the opposite of entraining to the user.
The corpus is thinner on the nonverbal side. Its evidence covers emotional and prosodic alignment, and there is no note on gesture or body language. The nearest doorway is work on full-duplex models, which represent language, audio and video as one interleaved stream and learn turn-taking timing inside a single model rather than as a bolted-on module Can a single model learn when to speak and respond?. That is where non-word channels start to become something a model can learn. The overall picture is that word-matching and tone-matching each need their own design target, their own training signal and their own evaluation.