Calibrated Trust
A graded user disposition in which the degree of acceptance granted to an AI output is scaled to its stakes: high-consequence outputs receive more independent verification, routine outputs receive proportionally lighter checking. Trust varies by context rather than being uniformly extended or withheld.
Operational Definition
STATE. Two trained raters score, for each AI output a user acts upon, (a) the stakes of the output on a 3-point scale (low/medium/high consequence of error) and (b) the observed verification effort the user applied before acting (none / light cross-check / independent confirmation). Calibrated trust is operationalised as the rank-order alignment between stakes and verification effort across a session; higher alignment indicates greater calibration.
Measurement Schema
Proposed measurement protocol (not yet empirically validated): Rater-coded verification-effort vs. stakes alignment per acted-upon output; calibration quantified as the Goodman-Kruskal gamma rank correlation between the stakes scale and the effort scale within a participant. Inter-rater reliability for both coded dimensions could be reported as weighted Cohen's kappa on a held-out 20% double-coded subset; participant-level self-report supplement via a 5-point Likert item ('I checked this more carefully because it mattered more').