Background: Existing literature has characterized medical student attitudes toward generative artificial intelligence (GenAI), but less is known about tasks, contexts, and behaviors. Educational effects may depend on whether students use GenAI to extend reasoning or accept outputs in place of it. Cognitive Load Theory and self-regulated learning theory suggest that uncritical AI use may bypass the processing and evaluation that durable learning requires. Objective: To characterize GenAI use across medical training, examine how and why students verify outputs, and explore augmentative versus substitutive use. Methods: We conducted an explanatory sequential mixed-methods study at one US MD-granting institution. A 37-item survey was administered to MS1-MS4 students (N=166; preclinical n=89, clinical n=77). Seventy-two prespecified comparisons were tested with Benjamini-Hochberg correction. Four focus groups (N=19) contextualized survey findings and were analyzed in NVivo 15. Results: GenAI use was near-universal (93.4%), but 86.1% reported no formal training. Clinical students used GenAI more during rotations (75.0% vs 6.0%; V=.709), whereas preclinical students used it more for post-class review (90.4% vs 55.6%; V=.397). Verification appeared to follow a stakes-based heuristic; students reported checking outputs more when errors could affect patient care and less when perceived consequences were low. Use fell along an augmentation-to-substitution continuum; some described AI explanations that felt sufficient in the moment but did not consolidate into durable learning. Conclusions: GenAI may weaken foundational knowledge and self-regulatory habits when it replaces rather than extends independent reasoning. This risk is amplified when learners rely on stakes-based verification while still developing the knowledge needed to evaluate outputs critically. These findings support a phased, longitudinal curriculum focused on verification and self-regulated AI use.