Skilled listeners use five different kinds of silence — can an AI learn to pick the right one in the moment?

This explores how an AI's choices about when to pause, when to speak, and when to interrupt shape whether talking with it feels natural. The corpus suggests that silence is not one thing. It is a set of different moves. A study of skilled human active listeners found five distinct pacing strategies: reflective silence, facilitative silence, empathic silence, holding space, and immediate response. Each one fits a different moment in a conversation. An LLM agent could tell which one a moment called for with about 86% accuracy What pacing moves make conversational agents sound like active listeners?. When those strategies were built into a support chatbot, people rated it a better listener and more human-like, and they shared more about themselves than with a bot that paced every reply the same way Can pacing strategies make conversational agents feel more human?. The surprise is that this happened in text chat, where you might expect timing to matter little. A well-placed delay still read as attention.

A second line of work treats staying quiet as something a model has to learn on purpose. DiscussLLM makes 'say nothing' an explicit choice, alongside five types of intervention, so the model learns when to speak as a separate skill from what to say Can models learn when NOT to speak in conversations?. Voice systems take this further. Instead of a pipeline that waits for you to finish and then replies, a single streaming model can take in audio and video continuously and learn turn-taking as part of its behavior, with response times under a second Can a single model learn when to speak and respond?. Both approaches point the same way: natural timing works better when it's learned than when it's bolted on as a rule.

The most useful tension in the corpus comes from two studies with older adults. In one, a voice agent that could interrupt and be interrupted, and that made small 'mm-hm' backchannel sounds, felt more natural, engaging, and fluent than a strict turn-taking agent. Its own topic-changing interruptions even steered talk toward what participants cared about, and nobody resented them Do voice agents that interrupt feel more natural to older adults?. In the other, frequent backchannel lowered older adults' ratings. Age-related slowing in auditory processing made rapid cues hard to separate from the conversation, so they felt like noise rather than engagement Does frequent backchannel from voice agents help or hurt older adults?. Read together, the lesson is not 'more backchannel' or 'less backchannel'. Natural timing depends on the listener. The same cue that signals attention to one person is clutter to another, depending on how fast they process speech.

Why does timing carry so much weight? One answer is that users judge an agent partly on human-likeness and communicative flexibility, not just competence. Those two factors make up about half of how people model a dialogue partner How do users mentally model dialogue agent partners?. A more philosophical answer is that AI output lacks the real-time event structure of a human utterance, so users do the work of turning text into a felt exchange Does AI generate genuine utterances or just text patterns?. On that view, good pacing helps the user do that work. It supplies some of the rhythm that makes a reply feel like a response to you rather than text that simply appears.

The corpus has gaps. It has little on measured latency thresholds, such as how many milliseconds of delay start to feel unnatural, and little on how timing norms differ across cultures. The strongest evidence covers support conversations and older adults, so treat the broader claims as promising rather than settled.