Do different levels of machine agency activate different interaction scripts?
This explores whether the *amount* of initiative a machine takes — from passive tool to proactive collaborator — changes the mental playbook people reach for when interacting with it, rather than humans running one fixed script for all AI.
This reads the question as connecting two threads the corpus keeps separate: that machine agency comes in degrees, and that humans run learned scripts when they interact. Put them together and the answer leans yes — agency level looks like a script trigger. The foundational move is recognizing agency isn't on/off. Does machine agency exist on a spectrum rather than binary? lays out five rungs — passive, semi-active, reactive, proactive, cooperative — and crucially says users judge each through a 'machine heuristic,' a mental shortcut for 'this is a machine, treat it accordingly.' That heuristic is the script in embryo.
The script idea gets its sharpest statement in Do humans apply human-human scripts to AI interactions?, which overturns the old CASA assumption that people just recycle human-human social habits onto machines. Longitudinal work shows people build a *second, machine-specific* script system through repeated use and apply it mindlessly — exactly the behavior you'd expect if different machine behaviors cue different responses. So the raw materials for the question's 'yes' are there: a spectrum of agency, plus evidence that humans grow dedicated scripts for machine agents rather than one universal one.
Where it gets interesting is the mechanism on the machine's side. Why do AI agents fail to take initiative? shows that initiative isn't a capability gap — next-turn reward optimization structurally trains it *out* of models, so most agents sit at the passive end by design, not necessity. When you deliberately train proactivity up (the note cites a jump from 0.15% to 73.98%), you're not just adding a feature — you're moving the agent to a different rung of the spectrum, which by the logic above should recruit a different user script. And the note's core tension — proactivity must be balanced against civility to avoid feeling intrusive — is really a statement that users *expect* a certain script and punish violations of it.
The reader-facing payoff is in how people model their partner. How do users mentally model dialogue agent partners? finds users evaluate dialogue agents along three axes — competence (49% of the variance), human-likeness (32%), and communicative flexibility (19%). A higher-agency, more proactive agent loads differently on these than a passive lookup tool, meaning the *same person* is running a different evaluative script depending on what the machine does. Can meta-learning prevent dialogue policies from collapsing? mirrors this from the machine side: a policy that can't adapt its behavior across user types collapses to one dominant action — the machine equivalent of running a single script for everyone, which fails. The thing you didn't know you wanted to know: the script-matching has to happen on *both* sides. Users adapt scripts to agency level, but agents that can't vary their own behavior across user profiles break — so interaction quality is really a negotiation between the user's machine-specific script and the agent's ability to hold more than one of its own.
Sources 5 notes
Research shows machine agency ranges across five levels—passive, semi-active, reactive, proactive, and cooperative—rather than existing as a binary choice. Users experience and judge these interactions through a 'machine heuristic' mental shortcut, and the mismatch between what AI can do and what workers want reveals deployment opportunities.
Extended CASA research shows humans develop and mindlessly apply interaction scripts specifically tailored to media agents rather than simply reusing human-human social scripts. Longitudinal studies demonstrate systematic changes in responses upon repeated AI interaction, revealing a coexisting second script system.
Research shows next-turn reward optimization structurally removes initiative from models, but proactive behaviors like critical thinking and clarification-seeking are trainable (0.15% to 73.98% with RL). The core challenge is balancing proactivity with civility to avoid intrusion.
The Partner Modelling Questionnaire reveals that perceived competence dominates user impressions (49% of variance), followed by human-likeness (32%) and communicative flexibility (19%). This three-factor structure reflects how people evaluate dialogue partners against both functional and social standards.
Without MAML, hierarchical RL for Motivational Interviewing phases collapses to a dominant action regardless of user type. Meta-learning enables the master policy to maintain variability and adapt across diverse user profiles.