Can pacing strategies make conversational agents feel more human?

In a between-subjects study (N=50) across two text-based supportive scenarios — career support and relationship support — the authors compared an LLM-based conversational agent that dynamically selects among five pacing strategies against a static-pacing baseline agent. "The context-aware agent scored higher than static-pacing control on perceived human-likeness, smoothness, and interactivity, supporting deeper self-disclosure and higher engagement." Deeper self-disclosure and higher engagement were evidenced by increased emotional language, first-person pronouns, and conversation turns. In the career-support scenario specifically, the context-aware agent also yielded higher perceived listening quality and affective trust.

The authors attribute this to two failure modes they name in current CA design. The first, "the fallacy of uniform immediacy," treats any response delay as latency to be eliminated, even though "in supportive contexts where both advice and emotional comforting are needed, consistently immediate responses can be perceived as perfunctory, emotionally detached, and affectively untrustworthy" — instant replies fail to perform the social function of deliberation. The second, "conversational dominance through continuous output," describes agents that "talk too much," flooding the interaction with lengthy answers that raise cognitive load and shift the exchange "from a collaborative dialogue to a passive information dump," which discourages self-disclosure. Against both, strategies like Holding Space and Facilitative Silence — where the agent strategically "shuts up and listens" — gave users room to process and elaborate, which the authors link to the observed engagement gains.

This gives controlled evidence for a claim When should AI systems choose to stay silent? raises architecturally but does not test empirically: that silence itself, independent of what is said, can raise perceived listening quality. The authors report that semantic content analysis confirmed "strong semantic equivalence" between conditions, so the measured effect is attributed to timing rather than wording. That also sits in tension with Does chatbot empathy actually make users ask more questions?, where empathetic cues in a health chatbot lengthened replies but did not raise question-asking or disclosure; here, manipulating pacing rather than cue content did move disclosure, suggesting the lever for disclosure in supportive chat may sit more in timing than in lexical or affective framing.

The finding is scoped to a single prototype across two support scenarios with free-flowing generative replies that the authors found infeasible to strictly control... turn-by-turn, and they explicitly caution against generalizing "our specific results or five-strategy framework to all populations," citing user background and traits as critical factors. The excerpt does not establish whether pacing effects hold in task-oriented, non-support, or voice-based interactions; the result supports pacing as a design-worthy variable in text-based emotional support contexts specifically, not as a general law of conversational agent design.