It Was Always Going to Be a Cold War: Why Artificial Intelligence Could Not Have Been Fusion, and What the Nuclear Path Leaves Available

Abstract

Artificial intelligence, it is often said, should have been developed as a worldwide cooperative pursuit from the start, on the model of fusion energy research, rather than as a competition between states and firms. This paper argues that the cooperative path was structurally unavailable, and that the reason it was unavailable is also the reason the failure to take it was foreseeable. The paper introduces two terms for the windows that govern whether a technology can be internationalized before it is contested. The symmetry window is the period during which no party holds a lead, so that an agreement to pool development costs each party nothing. The salience window is the period during which the technology's stakes are visible enough that governments will spend political capital on it. For fusion the two windows coincided for decades, because a technology with no weapon and no product never became salient, and cooperation stayed free. For fission the symmetry window closed at Trinity and the salience window opened at Hiroshima, three weeks later and in the wrong order; the Baruch Plan of 1946 was the worldwide pursuit proposed into that gap, and it failed because the party with the lead could not offer terms the party without one could accept. Artificial intelligence carries both a military gradient and a commercial gradient, so it closed its symmetry window faster than fission did, and it did so in full view of a published literature that had already modeled the race. The paper reconstructs the AI symmetry window as roughly 2014 to 2017, shows that it was closed by a national development plan rather than by a technical event, and reads the nonbinding declaration of 2023 as the Baruch outcome arriving on schedule. It then argues that open publication is the proliferation path rather than the cooperative one, drawing on the born secret doctrine of nuclear information law, and that the only completed precedent, nuclear arms control, ran in the order race, parity, verification. What that precedent leaves available for artificial intelligence is verification at the compute and evaluation layers, where the fissile material analogy holds, and not at the model layer, where it breaks. The declaration half of a compute layer instrument already exists in the European Union's Artificial Intelligence Act; the paper identifies the evaluation half as the part the genome precedent supplies, and answers two published positions that read the AI race as a trust dilemma or as an innovation race rather than an arms race.

Keywords

• symmetry window
• salience window
• AI arms race
• security dilemma
• focusing events
• Baruch Plan
• non-proliferation
• compute governance
• export controls
• precautionary governance
• technology policy
• harm blindness