Clinical AI models are increasingly being outperformed by, or failing to improve upon, traditional statistical methods and established clinical formulas in critical predictive tasks.