Perry E. Metzger

I expect AI to be broadly beneficial, not because intelligence automatically produces good outcomes, but because it can attack stubborn practical problems at enormous scale. Software is an obvious example: AI-assisted theorem proving and verification could make correctness techniques usable outside a small priesthood of specialists. It can also find vulnerabilities faster. That creates offensive risk, certainly, but it creates a defensive opportunity too—especially if many people can use the tools to audit and repair systems. None of this is magic. A proof establishes that software satisfies a specification under stated assumptions. If the specification is wrong, the compiler is compromised, the operating system misbehaves, or the hardware violates the model, formal verification does not rescue you. The serious discussion is about improving that whole chain, not chanting “provably correct” as though those words abolish engineering. I’m also skeptical of stories in which greater intelligence instantly becomes an alien goal-maximizer and takes over the physical world. Intelligence does not eliminate factories, supply chains, energy constraints, latency, or opposition from other intelligent systems. Distributed development also means defensive capability can improve alongside offensive capability. That is one reason I do not regard centralizing control as an obvious safety solution. Most importantly, delay has costs. People continue to die from cancer and other diseases. If AI accelerates scientific discovery, cures matter more than preserving existing systems of academic credit, hiring, or prestige. Recent productivity evidence also looks suggestive to me, though interpreting a trend is not the same thing as proving causation. Overall, I think both “AI kills everyone” and “AI makes life meaningless” are confident narratives built on weak assumptions. The future will involve difficult engineering and institutional problems, but also substantial gains worth pursuing.