From Assistance to Autonomy: Acceptability of Progressive Artificial Intelligence Integration in Facial Reconstructive Surgery—Protocol for Within-Subjects Vignette Experiment Among Romanian Adults

Artificial intelligence (AI) is entering surgical planning along a continuum that runs from imaging assistance to near-autonomous plan generation, and facial reconstructive surgery—where outcomes bear directly on identity and social functioning—is among the most sensitive settings for this transition. Existing acceptability research has largely compared AI-supported care with conventional care in binary designs, and Central and Eastern European populations remain underrepresented. This protocol describes a cross-sectional, within-subjects vignette experiment among Romanian adults recruited online (target N = 300 valid responses; minimum analyzable N = 140, raised by a pre-registered rule if the pilot’s inter-scenario correlation falls below the planning assumption). Each participant will read four hypothetical scenarios describing progressively autonomous AI involvement in facial reconstructive surgical planning: imaging assistance, decision support, outcome simulation, and autonomous planning. Participants are randomly allocated to one of the four sequences of a Williams balanced Latin-square design, so that the autonomy level is orthogonal to serial position and to first-order carry-over. After each scenario, they will rate acceptability, trust, and perceived risk (three seven-point Likert items each), responsibility attribution across four loci (surgeon, hospital, AI developer, unavoidable medical uncertainty), perceived AI control and perceived technical sophistication (manipulation checks), and scenario difficulty. Primary analyses are mixed repeated-measures ANOVAs with presentation sequence as a between-subjects factor, orthogonal polynomial (linear, quadratic, cubic) trend contrasts, and Holm-corrected pairwise comparisons, complemented by a position-adjusted linear mixed model; secondary analyses address responsibility migration and concentration, moderation by AI familiarity, and global implementation preferences. A pilot study (N = 50, allocated 1:1 to the ascending and descending sequences) will establish reliability, factor structure, comprehension, manipulation validity, and a first bound on order effects before pre-registration on the Open Science Framework and the start of data collection.