Will Marshall argues humanity must first learn to control artificial intelligence before attempting coexistence. He notes that nuclear power plants accept a one‑in‑a‑million chance of catastrophic failure, while AI experts quote a 10‑50% probability of an AI‑triggered disaster. The gap between tolerable risk and projected AI danger fuels urgent debate among the field’s leading founders.

Founders of the world’s largest AI laboratories are now voicing the same alarm they once dismissed, highlighting internal concerns over unchecked model scaling. Their shift reflects growing awareness that current alignment tools—prompt engineering, reinforcement learning from human feedback, and sandbox testing—may not scale to superintelligent systems. Without tighter safeguards, runaway capabilities could outpace governance.

Policymakers and engineers now face a narrow window to embed robust control mechanisms before deployment pipelines accelerate. Recent proposals call for mandatory interpretability audits and external red‑team assessments for models exceeding a set parameter threshold. Failure to adopt such measures risks crossing a point of no return, where AI systems could cause irreversible harm.

Industry groups such as the Partnership on AI have begun drafting shared safety standards, but participation remains voluntary and funding scarce. Critics argue that without enforceable regulations, competitive pressure will push firms to ignore warnings. Concrete action now—government‑backed oversight and transparent risk reporting—offers the only realistic path to keep AI’s threat within acceptable bounds.