Bastani and colleagues’ PNAS article (June 25, 2025) studied mathematics review practice at one Turkish school in fall 2023. The authors report classroom assignment through scheduling/balance software as randomized; the inspected methods do not specify its chance step. GPT-4 chat typically supplied solutions. Its classes scored higher in practice comparisons that were not preregistered, but lower on the preregistered, same-session closed-book exam than book-and-notes classes. Exam problems closely matched practice. A tutor given hidden teacher solutions and told to provide hints showed no clear exam difference, which does not prove equality. Changes were bundled; these immediate results do not measure long-term loss of established expertise.