AI-based diagnostic systems promise to improve medical decision-making, yet little is known about how the quality of their recommendations shapes behavioral reliance and users' confidence in their own expertise. We report a between-subjects online experiment in which 74 medical professionals interpreted 15 X-ray images with support from a simulated AI whose accuracy was manipulated (14 vs. 8 of 15 correct recommendations). Accurate AI advice increased reliance and diagnostic performance but lowered participants' confidence in their unaided judgment, consistent with social comparison theory. When the previously reliable AI produced a false negative, two thirds of participants who had initially identified the abnormality abandoned their correct diagnosis. A PLS-SEM analysis shows that domain self-efficacy raises diagnostic confidence, which in turn reduces calibrated diagnostic performance under accurate AI. The findings establish reliance and confidence as separable outcomes that can move in opposite directions and identify a boundary condition of self-efficacy in human-AI collaboration.