This study investigates the heterogeneous effects of AI assistance on 140 radiologists performing 15 chest X-ray diagnostic tasks. It reveals that conventional factors like experience, subspecialty, and familiarity with AI tools do not reliably predict the impact of AI assistance. Instead, the occurrence of AI errors significantly influences treatment outcomes, with inaccurate predictions negatively affecting radiologists' performance on all pathologies and half of the individual pathologies. The study emphasizes the need for personalized approaches to clinician-AI collaboration and accurate AI models. It finds that lower AI error leads to greater treatment effects on all pathologies and half of the individual pathologies, while the direction of AI error also impacts treatment outcomes. The study also shows that unassisted error is a poor predictor of treatment effect, and that AI error is a significant predictor of treatment effect, with more accurate AI predictions leading to better treatment effects. The study highlights the importance of developing more accurate AI models and helping radiologists identify inaccurate AI predictions. It also notes that AI predictions that underestimate ground truth probabilities can lead to better treatment effects than those that overestimate them. The study concludes that individualized approaches, high-quality AI models, and comprehensive assessments of multiple factors are essential for optimizing AI assistance in clinical practice.This study investigates the heterogeneous effects of AI assistance on 140 radiologists performing 15 chest X-ray diagnostic tasks. It reveals that conventional factors like experience, subspecialty, and familiarity with AI tools do not reliably predict the impact of AI assistance. Instead, the occurrence of AI errors significantly influences treatment outcomes, with inaccurate predictions negatively affecting radiologists' performance on all pathologies and half of the individual pathologies. The study emphasizes the need for personalized approaches to clinician-AI collaboration and accurate AI models. It finds that lower AI error leads to greater treatment effects on all pathologies and half of the individual pathologies, while the direction of AI error also impacts treatment outcomes. The study also shows that unassisted error is a poor predictor of treatment effect, and that AI error is a significant predictor of treatment effect, with more accurate AI predictions leading to better treatment effects. The study highlights the importance of developing more accurate AI models and helping radiologists identify inaccurate AI predictions. It also notes that AI predictions that underestimate ground truth probabilities can lead to better treatment effects than those that overestimate them. The study concludes that individualized approaches, high-quality AI models, and comprehensive assessments of multiple factors are essential for optimizing AI assistance in clinical practice.