A new study found that an AI “safety net” reduced diagnostic errors by 16% and treatment errors by 13% — but also created new challenges.