America’s AI Doctrine – Move Fast, Until the Pentagon Notices

Following recent announcements from the US government, the US approach to AI is beginning to resemble a constitutional magic trick. Domestically, Washington wants to sweep away onerous state level safeguards in the name of innovation and competative advantage. The message to Silicon Valley is reassuringly simple; build, scale, dominate and try not to let individual state regulation slow you down with tiresome questions about transparency, bias, safety or public accountability.

Then comes the second act. The same frontier model that is apparently too delicate to endure state level regulation suddenly becomes so strategically sensitive that the federal government may restrict access, monitor release, control foreign use or intervene on national security grounds. So the model is not dangerous enough for democratic safeguards but it is dangerous enough for national security handling. Light touch, it seems, now means lightly touched by everyone except Washington.

This is where the Article 36 question becomes awkward. Article 36 of Additional Protocol I to the Geneva Conventions requires states, when studying, developing, acquiring or adopting a new weapon, means or method of warfare, to determine whether its use would be prohibited under international law. It was designed for weapons review, not ordinary commercial software. However, it begs the question that if a frontier model is restricted because it may materially affect cyber operations, intelligence, military capability or strategic conflict, then the US cannot have it both ways. Either it is a commercial product being unfairly burdened by nervous regulators or it is a dual use capability requiring serious pre-deployment review.

The irony deepens when open weights enter the picture. Open Weights refer to the final weights and biases of a trained neural network. These values, once locked in, determine how an AI model interprets input data and generates outputs. When AI developers share these parameters under the likes of an OSI Approved License, they empower others to fine tune, adapt or deploy the model for their own projects.

If Washington restricts access to Western frontier models on national security grounds but the global market can increasingly turn to capable Chinese open weight alternatives, the policy may not reduce strategic risk so much as redirect dependency and increase it. The West/US then faces an uncomfortable asymmetry; its strongest models are closed, licensed and politically interruptible, while some of the most attractive open weight alternatives come from jurisdictions that raise their own sovereignty, transparency and influence concerns. That is not a settled market. It is a geopolitical arbitrage opportunity.

The global market will I suspect see the contradiction. America is not offering deregulation. It is offering innovation at full throttle, followed by a sovereign emergency hand brake turn. For allies and customers, the question is no longer whether US AI is powerful. It is whether buying or licensing it means quietly accepting US strategic discretion as part of the agreement.

In short, America is telling the world to trust its AI because it is lightly regulated, that is it now appears, unless it becomes important, at which point it is heavily restricted and everyone else is invited to enjoy the Chinese open weight alternative.