Do voice agents that interrupt feel more natural to older adults?

Liu and colleagues (CHI 2025 full paper) report that an LLM-powered "Barge-in" voice agent — one that allows both the user and the agent to interrupt mid-utterance, and that produces backchannel cues such as "yes" and "uh-huh" — was rated by older adult participants as more natural, engaging, and fluent than a "No Barge-in" turn-taking baseline agent, in a within-subject exploratory study. The authors state conversations with the Barge-in agent were "more natural, engaging, and fluent than with the No barge-in agent." Participant P6 said "The agent's interruptions introduce new topics, making the conversation more diverse."

The design is grounded in a formative study (N=8) of older adults' in-person conversations, from which the authors derive design considerations (DCs). DC1 states the agent "should prioritize co-operative interruptions when interacting with older adults, while incorporating a diverse range of interruption types," drawing on Murata et al.'s split between co-operative and competitive/intrusive interruption. The resulting agent used topic-changing interruptions it initiated itself — not only user-initiated barge-ins — to steer conversation toward topics drawn from participants' own stated interests, which the authors frame as "guiding topics proactively through contextual cues and timely interruptions."

Where Can pacing strategies make conversational agents feel more human? raises naturalness through calibrated silence and timing in text-based supportive chat, this paper raises it through bidirectional interruption and backchannel in voice; both findings converge on the claim that naturalness ratings respond to turn-level design features, not only to content quality. It also gives a concrete instance of an LLM agent initiating rather than only reacting, contrasting with Why can't conversational AI agents take the initiative? — here the agent's self-initiated topic-changing interruptions functioned as a form of initiative-taking that participants rewarded rather than resented. It also sits alongside Can models learn when NOT to speak in conversations? as a second, independently derived answer to when a voice or conversational agent should speak unprompted, and connects to How does conversational control actually shift between humans and AI? in locating conversational control in the interaction between parties rather than fixed to either one.

The excerpt does not give a participant count for the Barge-in evaluation study itself (only the formative study's N=8 is stated), and the study is a single prototype tested with couples and same-age friends on companionship-oriented topics (shared photos, shared positive experiences). The authors state applicability "to more task-oriented agents is unclear" and flag that interaction patterns with younger family members may differ. The finding establishes that two-way interruption plus backchannel can beat rigid turn-taking on self-reported naturalness for this population and context, not that it generalizes across relationship types, task-oriented use, or larger samples.