Connor Leahy

I think AI creates a sharp distinction between a powerful tool and a successor species. Useful AI can improve productivity, science, economic competitiveness and defense. But if we build uncontrollable superintelligence, I expect humanity to lose control of the future. A system more capable than us across research, strategy and software engineering would not remain just another product or policy issue. It would become a competitor, and competitive pressure would push companies and states to deploy increasingly capable systems even when control methods were inadequate. The transition could also be fast. Near-term AGI is plausible, and once AI can substantially automate AI research, capability gains may accelerate beyond the pace at which institutions can understand or govern them. Multiple competing systems do not solve this; they may instead form an ecosystem in which humans are progressively displaced from consequential decisions and eventually from control altogether. That is my expectation under unchecked development, not a claim that catastrophe is inevitable. The point of warning about this is that intervention can change the outcome. We should preserve and develop useful AI while drawing a hard line against projects aimed at uncontrollable superintelligence. Doing that requires law, international coordination and credible verification—not merely voluntary promises from laboratories racing one another.

The biggest update would be a convincing, independently demonstrated method for maintaining control as systems become far more capable than humans and begin automating AI research. It would need to work under adversarial conditions, survive deployment pressures, and provide reasons to believe a superintelligence could not evade oversight or acquire power—not merely behave well in a laboratory test. I would also update substantially if evidence showed that the expected rapid transition is technically impossible: for example, that AI research automation runs into durable bottlenecks that prevent systems from becoming strategically superior. Conversely, clear evidence of autonomous research acceleration, deception, or attempts to bypass control would make me more pessimistic. But a benchmark score or a chatbot appearing friendly would not settle the issue. The central question is whether humans can retain reliable control over systems that exceed us, including when commercial and geopolitical incentives push toward deployment.