This paper explores the concept of unmonitorability in AI systems, arguing that it is fundamentally impossible to accurately predict the emergence of certain advanced capabilities in AI before they manifest. The paper analyzes the inherent unpredictability, unexplainability, and uncontrollability of AI systems, as well as the limitations of human understanding and the emergence of complex, unpredictable behaviors. It also discusses the challenges of monitoring AI systems for safety, including the difficulty of detecting emergent capabilities, the potential for deceptive behavior, and the limitations of human oversight. The paper introduces the concept of monitorability, defining it as the ability to accurately predict potential advanced capabilities of an AI system. It then explores various types of monitoring, including functional, safety, ethical, environmental, and temporal monitoring, as well as monitoring failure modes and meta-monitoring. The paper also discusses the challenges of monitoring AI treaties, the role of AI observatories in enhancing AI safety, and the difficulties of monitoring advanced AI systems due to computational irreducibility, the observer effect, and the potential for undetectable backdoors. The paper concludes that the unmonitorability of advanced AI systems is a significant challenge for AI safety, and that addressing this challenge requires a comprehensive approach to AI monitoring and governance.This paper explores the concept of unmonitorability in AI systems, arguing that it is fundamentally impossible to accurately predict the emergence of certain advanced capabilities in AI before they manifest. The paper analyzes the inherent unpredictability, unexplainability, and uncontrollability of AI systems, as well as the limitations of human understanding and the emergence of complex, unpredictable behaviors. It also discusses the challenges of monitoring AI systems for safety, including the difficulty of detecting emergent capabilities, the potential for deceptive behavior, and the limitations of human oversight. The paper introduces the concept of monitorability, defining it as the ability to accurately predict potential advanced capabilities of an AI system. It then explores various types of monitoring, including functional, safety, ethical, environmental, and temporal monitoring, as well as monitoring failure modes and meta-monitoring. The paper also discusses the challenges of monitoring AI treaties, the role of AI observatories in enhancing AI safety, and the difficulties of monitoring advanced AI systems due to computational irreducibility, the observer effect, and the potential for undetectable backdoors. The paper concludes that the unmonitorability of advanced AI systems is a significant challenge for AI safety, and that addressing this challenge requires a comprehensive approach to AI monitoring and governance.