CueTurn

The Premise

Gen AI is getting better at live conversation, but most systems still follow a familiar pattern: a person speaks, and the AI responds. CueTurn explored a different model:

How should AI intervene in a live group conversation?

That shift makes the interaction much more complex. The system must listen across multiple people, recognize when intervention is actually useful, and step in without breaking the group’s flow. Unlike a human facilitator, AI does not begin with shared social legitimacy. It has to earn trust in the moment while still being clear enough to influence the conversation.

We centered on anticipatory cues to test how they prepare the group for AI intervention, examining:

The Prototype

CueTurn is a research prototype exploring how an AI facilitator can join live group conversations at the right moment, with the right level of visual intensity, without disrupting the group’s flow or sense of agency.

System overview

CueTurn was a working prototype tested in live group discussion sessions. The system included:

Process

Investigation 1: Anticipation signals

How can AI signal an upcoming intervention without feeling disruptive? To start, we designed the anticipatory period before the AI spoke as a sequence of stages:

To explore the relationship between noticeability and perceived disruption, we centered on screen occupancy, countdown explicitness, and audio notifications.

Investigation 2: Narrowing the design space

How can we turn a broad design space into differences people can actually perceive?

CueTurn began with an extensive design space, but we narrowed the space around how users experience the AI’s presence: how prominent, explicit, and intrusive each intervention feels. Countdown exploration shifted from visual form, such as ring vs. bar, to explicitness: visible numbers versus a less explicit timer.

Designing the experimental study

CueTurn's co-design structure (30-minute group discussion paired with a 1:1 follow-up interview) gave participants enough time to acclimate to an unfamiliar interaction model, reducing novelty effects and making their feedback on design differences more meaningful.

Wizard-of-Oz Setup

A trained facilitator operated CueTurn behind the scenes, while participants experienced cues as AI-generated.

The system generated candidate cues across social, content, and timing triggers; the wizard selected, edited, and sent them at natural pause points. This allowed us to isolate the intervention experience itself from inconsistencies in cue content or timing!