Do users want to adjust AI proactivity during conversations?

Houde and colleagues built "Koala," an LLM-based (Llama 2) conversational agent that joined Slack channels as a group-chat participant, and ran two user studies of group brainstorming with different levels of AI participation: no Koala, a reactive variant that "only responded when directly addressed," and a proactive variant that "made its own determinations of when to contribute." Participants in the first study preferred completing the brainstorming exercise with Koala present and liked the ideas it generated, but they "raised significant concerns about feeling overwhelmed and interrupted by the proactive variant" and "expressed numerous ways in which they desired to control its behavior." Some participants "directly asked Koala to reduce its volume of brainstorm posts" — while, separately, participants pushed the reactive variant to contribute more than it was.

The authors read this two-directional push (throttle the proactive variant, boost the reactive one, and both in real time, mid-session) as evidence against two existing accounts of proactivity in the mixed-initiative literature. McComb et al. treat proactive and reactive behavior as a binary — assistants like Alexa or Siri react, autonomous vehicles act proactively. Shneiderman's framework improves on this by treating human control and automation as independent dimensions, but still "appears to treat proactivity as a fixed and static attribute" of a given product. Houde and colleagues argue their data contradicts both: "proactivity is not a simple binary property of an AI agent," and critically, "proactivity may need to be modified in the midst of usage" — a capability neither framework anticipates, since both describe where an agent sits on an initiative scale rather than whether that position needs to move while people are using it.

This sharpens How can proactive agents avoid feeling intrusive to users?: that note frames civility as something designed in upfront; Koala's participants instead wanted a live dial, suggesting civility (or its absence) is partly a runtime-adjustment problem, not only a pre-deployment design problem. It also extends Why can't conversational AI agents take the initiative?: Koala is a rare case of an LLM agent that does take initiative in a group setting, and the finding is that too much initiative is its own failure mode, not just too little. The pattern also echoes How does conversational control actually shift between humans and AI?, which found conversational control co-constructed turn-by-turn in one-on-one AI companionship; Houde's groups show the same turn-by-turn negotiation occurring collectively, with multiple humans simultaneously trying to retune one agent's participation level.

The excerpt does not establish how far this generalizes: it is a single organization's internal Slack prototype, one underlying model (Llama 2), and one task type (ideation/brainstorming), so it is not shown whether the same mid-session retuning desire holds for other task types, larger groups, or newer models. The excerpt also does not describe what controls, if any, Study 1 participants were actually given to act on this desire — that is addressed separately in the paper's Study 2 taxonomy. At the strength the evidence allows, the implication is that initiative level is better treated as a live, user-adjustable setting than as a fixed design-time choice, at least for LLM agents embedded in real-time group chat.