The Shoggoth of AI Risk

AI risk is not a single, monolithic problem but a collection of distinct issues from different sources. The main sources of risk are categorized as adversaries, unforced errors, misaligned goals, dependencies, societal impact, and emergent behavior. Each category presents unique challenges, from intentional attacks and simple mistakes to an AI optimizing for the wrong metric or relying on fragile external systems. By clearly defining and understanding each type of risk, it becomes possible to move from panicked reactions to precise, strategic responses.