Generalization is a major challenge for responsible AI in patient-facing clinical applications. AI systems must be able to apply their knowledge to new data that may differ from their training data. Current debate in bioethics proposes selective deployment as a solution. This paper explores data-based reasons for generalization challenges and looks at how selective predictions might be implemented technically, focusing on clinical AI applications in real-world healthcare settings. Generalization is a core challenge for real-world impact in all areas of human-centric AI. There are currently limited technical solutions that work for generalization challenges in patient-facing clinical applications of machine learning. To address this, recent work in bioethics advocates selective deployment of AI in healthcare and provides a thorough analysis of the ethical implications. "Selective deployment" suggests that algorithms should not be deployed for groups underrepresented in their training datasets due to risks around poor or unpredictable algorithm performance. Why is generalization a challenge in clinical AI? Expressive ML models, especially deep neural networks, are prone to overfitting, i.e., they over rely on low-level features and learn spurious correlations in a dataset. Furthermore, training data reflecting societal prejudices or lacking diversity can result in algorithmic biases that can cause models to generalize less well to underrepresented groups. These problems are exacerbated in clinical applications, where datasets are high dimensional, contain the inherent uncertainties of biological systems, are often small and noisy, contain large numbers of missing values, and may not be representative of the target population. ML models that do not generalize may fail silently, i.e. perform significantly worse on new samples or individuals unnoticed, especially if not externally validated. Ignoring these challenges and applying ML models in the clinic regardless is irresponsible as it may harm patients from underrepresented groups. The paper discusses the ethical considerations of selective deployment in clinical AI, highlighting the need for responsible AI that balances practicality with equity. It also explores technical approaches for achieving trustworthy predictions, including data-centric and model-centric methods. The paper concludes that while selective deployment is a potentially contentious option, it may represent the most ethical tradeoff between competing considerations around utility, safety, and equity.Generalization is a major challenge for responsible AI in patient-facing clinical applications. AI systems must be able to apply their knowledge to new data that may differ from their training data. Current debate in bioethics proposes selective deployment as a solution. This paper explores data-based reasons for generalization challenges and looks at how selective predictions might be implemented technically, focusing on clinical AI applications in real-world healthcare settings. Generalization is a core challenge for real-world impact in all areas of human-centric AI. There are currently limited technical solutions that work for generalization challenges in patient-facing clinical applications of machine learning. To address this, recent work in bioethics advocates selective deployment of AI in healthcare and provides a thorough analysis of the ethical implications. "Selective deployment" suggests that algorithms should not be deployed for groups underrepresented in their training datasets due to risks around poor or unpredictable algorithm performance. Why is generalization a challenge in clinical AI? Expressive ML models, especially deep neural networks, are prone to overfitting, i.e., they over rely on low-level features and learn spurious correlations in a dataset. Furthermore, training data reflecting societal prejudices or lacking diversity can result in algorithmic biases that can cause models to generalize less well to underrepresented groups. These problems are exacerbated in clinical applications, where datasets are high dimensional, contain the inherent uncertainties of biological systems, are often small and noisy, contain large numbers of missing values, and may not be representative of the target population. ML models that do not generalize may fail silently, i.e. perform significantly worse on new samples or individuals unnoticed, especially if not externally validated. Ignoring these challenges and applying ML models in the clinic regardless is irresponsible as it may harm patients from underrepresented groups. The paper discusses the ethical considerations of selective deployment in clinical AI, highlighting the need for responsible AI that balances practicality with equity. It also explores technical approaches for achieving trustworthy predictions, including data-centric and model-centric methods. The paper concludes that while selective deployment is a potentially contentious option, it may represent the most ethical tradeoff between competing considerations around utility, safety, and equity.