The Government Just Published Its AI War Doctrine

The document dropped on October 24th, quiet as a policy memo can be, and it contains sentences that would have sounded like speculative fiction five years ago.

The White House's National Security Memorandum on AI — NSM-33 if you want to be boring about it — is the Biden administration's formal declaration that the United States government is now in the AI race as a national security matter, not a technology policy matter. These are different things. Technology policy is NIST standards and interagency task forces. National security means the NSA, lethal force decisions, nuclear weapons, and export controls wielded like a weapon.

All of that is in here.

---

The most remarkable sentence in the document is about nuclear weapons. The memo explicitly prohibits autonomous AI from making decisions about nuclear weapons use. Read that again: the United States government had to write down, in a formal presidential directive, that artificial intelligence cannot decide to launch nuclear weapons on its own. The fact that this needed to be said — that the scenario was concrete enough to require explicit prohibition — is its own kind of landmark.

They didn't write that sentence because it was a hypothetical.

---

The NSA's AI Security Center, which most people didn't know existed two years ago, gets elevated here to something approaching a regulatory body with authority over AI security standards across the entire national security apparatus. The NSA has historically operated in a posture of "we regulate nothing, we absorb everything" — having it now push standards outward, over agencies that have never been eager to accept NSA oversight, is a quiet structural shift that will matter more than almost anything else in this document.

The memo also, in language that is careful and precise and obviously deliberate, endorses AI-assisted offensive cyber operations. NSA directed to use AI for "cybersecurity operations." The government definition of cybersecurity operations has always included offense. This is the first time something this formal has said it out loud at this level.

---

China gets named directly, repeatedly, as the primary driver of all of this. Not as a diplomatic aside — as the organizing threat. The memo frames the entire national security AI project as a response to Chinese AI development, and it directs concrete countermeasures: tighter export controls on advanced chips, denied access to frontier model weights, active monitoring of Chinese AI programs by the IC, and aggressive counter-positioning in international AI standards bodies.

The document is, in its bones, a cold war document — the kind you write when you've decided that a competition has already started and you're catching up on formally acknowledging it.

---

What the note I scrawled about "sovereign federated AI" was pointing at: the memo gestures, without naming it directly, at the shape of what allied AI infrastructure needs to look like. Partner nations building AI capacity. Interoperability requirements. Data sharing arrangements that only work with federated architectures. The US cannot both export-control its way to AI dominance and expect its allies to build nothing — so the memo quietly carves out space for something like an allied AI bloc, where each nation maintains sovereignty over its systems but they compose into something larger.

Nobody has built that yet. The memo is asking for the foundation before the building exists.

---

The human-in-the-loop requirement for lethal force decisions gets formally codified here too — no autonomous AI authorizes lethal force, period, human judgment required. This sounds obvious. It is not obvious. It is a commitment that the United States is making at the same moment it is directing the NSA to use AI offensively, and the tension between those two things is going to define the next decade of defense AI policy in ways that the people who wrote this document almost certainly understand.

They wrote it anyway. That's the job.

---

The part that hits different when you read it: this document is the product of an administration that is leaving office in three months. They knew that. They wrote it anyway — not as a set of policies that will definitely be implemented, but as a record of where the consensus landed, a marker in the ground that the next administration will have to reckon with or explicitly undo.

That's not science fiction. That's just how this works. The pace is the thing that reads like fiction.

It all happened so fast.