ABSTRACT
AI systems and the societies they operate within exhibit many of the hallmarks of complex systems. This has important implications for ensuring AI safety. Complex systems display emergent properties that arise from interactions. These properties do not exist in any of the components in isolation and cannot easily be derived from reductive analysis of the system. Unintended side effects often result from interventions in complex systems, which means that cautious trial and error will likely be required in making AI systems safe. To develop effective strategies for AI safety, it is crucial to adopt approaches that account for the unique properties of complex systems and enable us to anticipate and address AI risks. This often requires digging beneath the most obvious problems to identify systemic factors that can be influenced.
