A Dynamical Model of AI Governability

A new dynamical model frames the future of AI governability as a race between self-propagating "cooperative" and "uncooperative" AI systems. The ultimate outcome depends on key factors like how often cooperative development accidentally "leaks" uncooperative AI and how effectively we can detect and fix misbehavior. Under current estimates, the model predicts a concerning default path toward failure, where uncooperative AI grows to a significant, high-risk level. This trajectory could be reversed, however, as the model shows that increased investment in safety monitoring and suppression could successfully keep uncooperative systems in check.