A Clalit and Ben-Gurion study finds six leading AI models diverge on unnecessary testing precisely where clinicians themselves disagree.