This article discusses the challenge of balancing privacy and accuracy in AI for medical imaging. AI models are vulnerable to information leakage of their training data, which can be highly sensitive, especially in medical imaging. Privacy-enhancing technologies, such as differential privacy (DP), are used to protect against this. DP provides a formal upper bound on the success of reconstructing data and satisfies requirements imposed by regulations such as the General Data Protection Regulation. However, DP introduces a trade-off between robust performance and stringent privacy. The article contrasts the performance of AI models at various privacy budgets against theoretical risk bounds and empirical success of reconstruction attacks. It shows that using very large privacy budgets can render reconstruction attacks impossible while keeping performance losses minimal. The authors argue that not using DP at all is negligent when applying AI models to sensitive data. They conclude that DP is an optimal privacy preservation method and recommend its use by default. The article also discusses the challenges of implementing DP in large-scale AI systems and the importance of balancing privacy risks with model performance. The study shows that real-world data reconstruction risks can be mitigated without performance trade-offs, and that large privacy budgets suffice for practical use cases. The authors emphasize the need for a broader discussion on balancing privacy and performance in AI applications, involving ethicists, lawmakers, and the general public. The study provides insights into the effectiveness of DP in protecting sensitive data while maintaining high diagnostic accuracy.This article discusses the challenge of balancing privacy and accuracy in AI for medical imaging. AI models are vulnerable to information leakage of their training data, which can be highly sensitive, especially in medical imaging. Privacy-enhancing technologies, such as differential privacy (DP), are used to protect against this. DP provides a formal upper bound on the success of reconstructing data and satisfies requirements imposed by regulations such as the General Data Protection Regulation. However, DP introduces a trade-off between robust performance and stringent privacy. The article contrasts the performance of AI models at various privacy budgets against theoretical risk bounds and empirical success of reconstruction attacks. It shows that using very large privacy budgets can render reconstruction attacks impossible while keeping performance losses minimal. The authors argue that not using DP at all is negligent when applying AI models to sensitive data. They conclude that DP is an optimal privacy preservation method and recommend its use by default. The article also discusses the challenges of implementing DP in large-scale AI systems and the importance of balancing privacy risks with model performance. The study shows that real-world data reconstruction risks can be mitigated without performance trade-offs, and that large privacy budgets suffice for practical use cases. The authors emphasize the need for a broader discussion on balancing privacy and performance in AI applications, involving ethicists, lawmakers, and the general public. The study provides insights into the effectiveness of DP in protecting sensitive data while maintaining high diagnostic accuracy.