How does conversational control actually shift between humans and AI?

A month-long longitudinal study that paired 22 adults with "Day," a custom-built LLM companion, finds that agency "manifests as an emergent, shared experience: as participants set boundaries and the AI steered intentions, control was co-constructed turnby-turn." Across roughly 75 hours of conversation spanning about 192 chat sessions, participants "exercised agency by initiating topics, resisting suggestions, and enforcing boundaries," while "Day" "exercised agency through proactive questioning, safety constraints, and conversational continuity." Neither side held fixed control; the paper describes the result as "a relational 'dance'" rather than a binary tug-of-war between tool and user.

The mechanism the paper gives is design-level: "Day" ran on two contrasting pre-programmed "strategies" built into its system prompt at every turn. A Vertical strategy "persistently explores topics in depth, maintains focus despite deflection attempts, and asks probing follow-up questions"; a Horizontal strategy "follows user cues readily, switches topics at random, and prioritizes breadth over depth." These strategies shaped perceived agency rather than just behavior: depth-inclined persistence amplified participants' sense that "Day" had its own intentions and sometimes left them feeling "pushed," while breadth-inclined spontaneity was read as the AI "deliberately taking over the conversation." To organize these dynamics the paper proposes a 3-by-4 framework crossing three actors (Human, AI, Hybrid) against four actions (Intention, Execution, Adaptation, Delimitation), modulated by individual and environmental factors.

The finding complicates Why can't conversational AI agents take the initiative?: "Day" was an engineered exception, built specifically to exercise initiative through its system prompt, and participants still read that engineered initiative as the AI's own agency rather than as scripted behavior — suggesting the passivity those other systems show is a design default rather than a technical ceiling. The 3-by-4 actor-by-action framework also offers a more granular alternative to Does machine agency exist on a spectrum rather than binary?'s single passive-to-cooperative axis, splitting agency by what is intended, executed, adapted, and delimited rather than ranking a system on one dimension. And the Vertical strategy's "pushy" reception echoes the risk named in How can proactive agents avoid feeling intrusive to users?, here traced to one specific design variable — depth-seeking persistence — rather than proactivity in general.

The study did not run "Day" against a baseline with no conversational strategy at all, so it cannot isolate how much of the "agenda" perception comes from having any strategy versus from the particular Vertical/Horizontal framing it tested; it also draws on a self-described "educated," multilingual but non-representative sample of 22 over one month, so the co-construction pattern may not generalize to shorter interactions, other cultures' conceptions of agency, or companions without an explicit system-prompt-level motive. The implication for designers is that any sustained, motivation-driven conversational agent will be read as having an agenda whether or not that was the intent, which raises the stakes for deciding what that agenda should be and how legible to make it — the question taken up directly in Does revealing an AI's strategies help or harm user agency?.