**Safety Cases: How to Justify the Safety of Advanced AI Systems** As AI systems become more advanced, companies and regulators must decide whether they are safe to train and deploy. To prepare for these decisions, developers need to construct a "safety case," a structured rationale that AI systems are unlikely to cause a catastrophe. This report proposes a framework for organizing a safety case and discusses four categories of arguments to justify safety: total inability to cause a catastrophe, sufficiently strong control measures, trustworthiness despite capability to cause harm, and deference to credible AI advisors. It evaluates examples of arguments in each category and outlines how they can be combined to justify the safety of AI systems. The report introduces the concept of a safety case, which is a method used in six UK industries to present safety evidence. A safety case is a structured rationale that a system is unlikely to cause significant harm if deployed in a specific setting. The report explains how developers can construct a safety case in the context of catastrophic AI risk. Section 2 provides an executive summary, outlining a framework for structuring an AI safety case. Section 3 defines essential terminology. Section 4 lists recommendations for institutions using safety cases to make AI deployment decisions. Section 5 groups arguments that AI systems are safe into four categories: inability, control, trustworthiness, and deference. It describes examples of arguments in each category and rates them on practicality, maximum strength, and scalability. Section 6 explains how to combine the arguments from section 5 into a holistic safety case. The report discusses four categories of safety arguments: inability, control, trustworthiness, and deference. Inability arguments claim that AI systems are incapable of causing unacceptable outcomes in any realistic setting. Control arguments assert that AI systems are incapable of causing unacceptable outcomes given existing control measures. Trustworthiness arguments claim that AI systems are safe because they consistently behave in a desirable way. Deference arguments assert that AI advisors are credible and can be relied upon to determine the safety of other AI systems. The report also discusses the practicality, maximum strength, and scalability of each argument category. It provides examples of arguments in each category and explains how they can be used to justify the safety of AI systems. The report concludes with recommendations for institutions using safety cases to make AI deployment decisions, emphasizing the importance of continuous monitoring, risk assessment, and the development of hard standards and formal processes.**Safety Cases: How to Justify the Safety of Advanced AI Systems** As AI systems become more advanced, companies and regulators must decide whether they are safe to train and deploy. To prepare for these decisions, developers need to construct a "safety case," a structured rationale that AI systems are unlikely to cause a catastrophe. This report proposes a framework for organizing a safety case and discusses four categories of arguments to justify safety: total inability to cause a catastrophe, sufficiently strong control measures, trustworthiness despite capability to cause harm, and deference to credible AI advisors. It evaluates examples of arguments in each category and outlines how they can be combined to justify the safety of AI systems. The report introduces the concept of a safety case, which is a method used in six UK industries to present safety evidence. A safety case is a structured rationale that a system is unlikely to cause significant harm if deployed in a specific setting. The report explains how developers can construct a safety case in the context of catastrophic AI risk. Section 2 provides an executive summary, outlining a framework for structuring an AI safety case. Section 3 defines essential terminology. Section 4 lists recommendations for institutions using safety cases to make AI deployment decisions. Section 5 groups arguments that AI systems are safe into four categories: inability, control, trustworthiness, and deference. It describes examples of arguments in each category and rates them on practicality, maximum strength, and scalability. Section 6 explains how to combine the arguments from section 5 into a holistic safety case. The report discusses four categories of safety arguments: inability, control, trustworthiness, and deference. Inability arguments claim that AI systems are incapable of causing unacceptable outcomes in any realistic setting. Control arguments assert that AI systems are incapable of causing unacceptable outcomes given existing control measures. Trustworthiness arguments claim that AI systems are safe because they consistently behave in a desirable way. Deference arguments assert that AI advisors are credible and can be relied upon to determine the safety of other AI systems. The report also discusses the practicality, maximum strength, and scalability of each argument category. It provides examples of arguments in each category and explains how they can be used to justify the safety of AI systems. The report concludes with recommendations for institutions using safety cases to make AI deployment decisions, emphasizing the importance of continuous monitoring, risk assessment, and the development of hard standards and formal processes.