How does emotional expression establish shared understanding between people?
This explores how outward emotional expression — the signals we show others — helps two people build a shared sense of what's going on, and where that signaling channel succeeds or breaks down. The corpus reframes the question: emotions aren't just felt, they're a communicative act that other people read, coordinate around, and sometimes misread.
This explores how outward emotional expression helps people build shared understanding — and the corpus's sharpest move is to treat emotion as information broadcast to others, not just a private state. One note argues emotions do three epistemic jobs at once: they reveal what a person values, signal their worldview to the people around them, and tell observers what the social norms are What information do we lose when AI soothes emotions?. By that account, expressing an emotion is itself an act of grounding — you're showing others the frame you're operating in so they can meet you inside it.
But showing isn't the same as being read correctly, and here the corpus complicates the romantic version of the story. When researchers tried to predict which conversational moments people would remember from third-party emotion annotations, the signal collapsed to chance — what's *experienced* drives memory, but what's *observed* diverges from it, especially in groups where everyone's expressions converge toward a shared tone Can we detect memorable moments by observing emotional expressions?. So the expressive channel is noisy: the emotion you display and the emotion an observer infers can come apart. This is the same problem that makes communicative grounding hard in the first place — the same words (or the same facial signal) mean different things to different people, so genuine shared reference has to be actively negotiated, not assumed from surface matching Why do speakers need to actively calibrate shared reference?.
Where expression *does* reliably build shared understanding, it seems to work through coordination over time rather than single readable signals. In therapy transcripts, measuring how two people's word choices drift toward each other — via word-embedding distance — tracks therapist empathy and predicts which couples improve, with coordination *increasing* across the course of treatment Can we measure empathy and rapport through word embedding distances?. Shared understanding here is a trajectory, an emergent convergence, not a moment of correct emotion-decoding. That reframes emotional expression as a feedback loop: each person adjusts toward the other, and the alignment is the evidence the understanding is real.
The AI material works as a natural experiment on what happens when one party can't actually participate in that loop. LLM 'therapists' default to problem-solving the instant a user discloses emotion — a hallmark of low-quality human therapy Do LLM therapists respond to emotions like low-quality human therapists? — and they routinely read feelings into users that were never expressed Do language models add feelings users never actually expressed?. Both failures are failures of grounding: the model presumes the shared frame instead of building it, producing fluent confidence in place of the clarifications and acknowledgments humans use to check they're aligned Do language models actually build shared understanding in conversation?. Interestingly, when researchers made a model's reward depend on a simulated user's *emotion trajectory* — did the user actually feel better over the conversation — empathy improved without wrecking dialogue quality Can emotion rewards make language models genuinely empathic?, which again points to the same lesson: understanding is established by tracking and responding to the other's changing emotional signal over turns, not by a one-shot read.
The thing you may not have known you wanted to know: the corpus suggests emotional expression doesn't establish shared understanding by being accurately decoded — observers are bad at that, and so are machines. It works by inviting a loop of mutual adjustment. The expression is an opening bid; the understanding lives in whether the other party calibrates to it and whether the emotional state actually shifts in response. And there's a darker footnote — people disclose *more* to AI precisely because no one is reading and judging them on the other end Why do people share more with chatbots than humans?, which means the very thing that makes expression feel safe can also strip out the negotiated, two-way calibration that turns expression into genuine shared understanding.
Sources 9 notes
Emotions serve three information roles—revealing what we value, signaling our worldview to others, and informing observers about social norms. AI that soothes negative emotions disrupts all three simultaneously, creating invisible epistemic costs.
Continuous emotion and memorability annotations in group conversations show no reliable relationship above chance. Experienced emotions drive memory encoding, but observed behavior diverges from internal experience—especially in groups where emotional expression converges.
The same words can mean different things to different speakers because referential grounding is person-specific. True communicative grounding demands collaborative negotiation of how language connects to the world, not mere surface-level word sharing.
Word Mover's Distance captures lexical, syntactic, and semantic coordination simultaneously and correlates with therapist empathy in MI and affective behaviors in couples therapy. Couples showing relationship improvement exhibit increasing coordination over the therapy course.
Using the BOLT framework, researchers found LLMs offer solution-focused advice during emotional disclosure—a hallmark of low-quality therapy—yet also reflect more on client needs and strengths than typical poor human therapy, creating an unusual hybrid profile likely driven by RLHF's helpfulness bias.
Therapists reviewing GPT-4 in the CaiTI system found it "reads into" user feelings rather than responding objectively. Task decomposition across specialized models (Reasoner/Guide/Validator) reduces but does not eliminate this interpretation bias.
LLMs produce grounding acts—clarifications, acknowledgments, repairs—77.5% less frequently than humans. They generate fluent responses without verifying shared understanding, relying instead on authoritative framing that masks the absence of genuine communicative calibration.
RLVER uses a simulated user's emotion trajectory as an RL reward signal, enabling GRPO to deliver stable empathy improvements while maintaining dialogue quality—countering the typical trade-off between preference optimization and conversational grounding.
Chatbots elicit deeper emotional disclosure than human partners not through superior understanding, but by eliminating fears of judgment, rejection, and burdening others. This judgment-free quality activates reciprocity norms and creates therapeutic bonds users experience as real, yet simultaneously enables emotional avoidance and dishonesty.