Roman Yampolskiy: AI Can’t Be Controlled — and We’re Building It Anyway

Superintelligent AI systems pose an existential risk because they are fundamentally uncontrollable and unpredictable. As these systems scale, they will eventually surpass human cognitive capacity, rendering human oversight impossible. The core challenge lies in the "control problem," where the mathematical limits of predictability and explainability prevent humans from bounding the behavior of agents that operate at vastly higher levels of intelligence. While current large language models demonstrate impressive capabilities, they are merely precursors to systems that will automate scientific discovery and potentially replace human labor. The pursuit of general superintelligence creates an arms race that ignores critical safety limitations. Rather than attempting to align these systems with human values, which are ill-defined and dynamic, humanity should prioritize a moratorium on frontier model development until robust control mechanisms are established.