Familiar Style and Calibrated Correction in Repeated AI Use. A framework for style recognition discrimination and substantive correction

Abstract

Repeated users of an artificial-intelligence agent may learn both its competence and how it sounds. This paper proposes that acquired style congruence can change acceptance independently of evidential quality, while any benefit to error management may arise through more than one route. Learned style–correctness contingencies can guide classification; reduced processing demands may also free attention for substantive checking when style itself is non-diagnostic. The framework distinguishes recognition, response criterion, classification sensitivity, competence learning, defect identification, successful correction, and final accuracy. Five competing mechanisms are specified: familiarity halo, anomaly vigilance, diagnostic cue learning, competence learning, and processing-efficiency learning. Building on appropriate-reliance research, calibration symmetry is treated as a requirement for identification and measurement: both familiar-looking wrong answers and unfamiliar-looking correct answers must be examined. It does not imply equal frequency, cost, or normative importance. The contribution is a bounded synthesis for acquired, source-specific style, not a new general principle of reliance or a demonstrated familiarity effect. Study 1 crosses acquired style congruence with correctness while preserving randomized assignment and measuring inferred source or version change. Study 2 varies cue diagnosticity and distinguishes classification through style from identifying and repairing substantive defects. Outcome feedback, trial-level learning, delayed unaided performance, and transfer are specified separately from assisted performance. Conversational affirmation and linguistic accommodation are adjacent pathways, not evidence of acquired-style effects. Governance implications remain provisional pending direct evidence.