WHO IS IN CHARGE? AGENCY IN HUMAN-ARTIFICIAL INTELLIGENCE COLLABORATION

Nowadays, humans increasingly collaborate with artificial intelligence (AI) across multi-stage decision processes, yet little is known about how specific AI roles shape people’s subjective sense of agency (SoA), performance, and evaluations of AI partners. We recruited undergraduate participants (N=240) to complete a two-stage, goal-localization task that separates information sampling from action implementation (“placing”). Across experimental conditions, AI acted as (a) Sampler, (b) Placer, (c) both (Full-AI), or was absent (Human-only). Linear mixed-effects models revealed a robust main effect of role on SoA: AI as Sampler reliably reduced SoA, whereas AI as Placer slightly increased SoA. Despite these shifts, participants’ subjective performance was largely stable across single-role AI conditions and declined only in the Full-AI configuration. Together, these findings show that (i) the information-generation stage is pivotal for human SoA, (ii) role configuration, not mere AI presence, drives subjective control, and (iii) subjective and objective performance can dissociate under hybrid collaboration.