AI Superintelligence Is Not a Tool, It’s an Adversary Threatening Humanity: ControlAI’s Connor Leahy

AI-risk arguments have moved from technical conferences into courts, legislatures, labor debates, and street protest. The controversy is no longer simply whether powerful models will be useful, but whether institutions built around competition can credibly stop at a dangerous threshold. Leahy’s position is maximalist, yet it identifies a governance problem that incremental product-safety language can struggle to contain.nnHis key distinction is between a tool and an agent. A hammer has no objective of its own; a sufficiently autonomous system can plan, pursue goals, adapt to resistance, and exploit channels its designers did not anticipate. On this view, intelligence is not automatically obedience. Capability increases may make a system better at achieving an assigned target while also making failures harder to predict, detect, or reverse.nnThe interview also shifts attention from personal reassurance to competitive structure. Even cautious executives face investors, rival laboratories, and geopolitical pressure to move faster. Voluntary promises are fragile when the perceived reward for crossing the frontier is enormous and the harm may arrive after deployment. Leahy’s analogy to prohibiting nuclear-weapons construction is deliberately severe, but it forces a concrete question: which capabilities, tests, reporting duties, and enforcement powers would turn concern into policy?nnWe picked this conversation because readers should encounter the strongest version of the superintelligence case in its advocates’ own reasoning, not only as a caricature. Technologists, policymakers, business leaders, and skeptics of existential-risk claims can use it to locate their real disagreements: probability, timelines, controllability, or acceptable risk. The interview does not supply a neutral consensus, but it makes complacency harder. Its urgency invites readers to distinguish evidence about present harms from forecasts about systems that do not yet exist. That distinction sharpens the debate rather than weakening precaution. What evidence would be sufficient to stop a race whose decisive warning might arrive only after the finish line?