What changes for warfare with AI?

The shift

The cognitive work that paced war — reading the situation, weighing options, deciding to act, judging when to stop — went from something only human minds could do, at human speed, to something machines can generate and increasingly execute far faster than deliberation. War was gated by human cognition and attention: how fast a person could understand what was happening and choose. That gate is being pulled out of parts of the loop, and where it is, the tempo of conflict is set by machines reacting to machines.

The axioms

• War is paced by human decision speed and attention — force is applied only as fast as commanders can understand a situation and choose to act. Rests on the sensing-to-acting loop being gated by scarce, slow human cognition.
• Deterrence rests on predictable human command — an adversary is deterred because it believes a rational human authority controls whether force is used and can be reasoned with, threatened, and reassured. Rests on the decision to escalate being a legible human act.
• Escalation is paced by humans — the ladder from tension to war is climbed at the speed of deliberation, leaving room to pause, signal, and step back. Rests on the interval between provocation and response being long enough for human intervention.
• Attribution is possible — an attack can be traced to an actor who can then be held responsible, which is what makes deterrence and retaliation coherent. Rests on the cost and difficulty of acting anonymously at scale being high.
• A political authority is accountable for the decision to go to war and to wage it. Rests on accountability being a property only a liable human or state can hold.
• War is bound by the laws of armed conflict, which require moral and legal judgment a human owns. Rests on legal and moral agency being irreducibly human.
• Wars turn on the human will to fight and to stop — resolve, morale, cohesion, and the willingness of populations and soldiers to endure or quit. Rests on will and trust being human and not manufacturable on demand.
• Waging war requires judgment under novel, high-stakes ambiguity — deciding what to do when the situation matches no plan and no precedent. Rests on judgment under genuine novelty being scarce.

Invalid axioms

1. War is paced by human decision speed and attention. The loop from sensing to a recommended — and in narrow cases, executed — response can now run below the speed of human deliberation. The belief was true for all of prior history; it is no longer a floor. The habit-trap: doctrine, command structures, and arms-control instincts all assume there is always time for a human to read the situation and decide, so they invest in the quality of that human decision without noticing the window for it is closing in the fastest engagements.
2. Escalation is paced by humans. When both sides run systems that react to each other faster than people can intervene, the interval that let humans pause, signal, and climb down can collapse. The habit-trap: crisis-management playbooks — hotlines, deliberate signaling, buying time — are built around a tempo that machine-versus-machine interaction can erase, and they are still treated as the safety net when the net may not close fast enough.
3. Attribution is possible because acting anonymously at scale is hard. Generating convincing deniable activity, plausible cover, and confusion at volume is cheap when the cognitive work behind it is cheap. The habit-trap: deterrence and retaliation are still framed around finding and punishing a responsible actor, on the assumption that a serious act leaves a traceable author — an assumption that weakens as the cost of manufacturing ambiguity falls.

Unchanged axioms

1. A political authority is accountable for the decision to go to war and to wage it. Accountability requires someone a population, a court, or a coalition can hold responsible. A model cannot be answerable, and neither delegation to a system nor the speed of an engagement dissolves the political ownership of having chosen war. This becomes more load-bearing as machines generate and execute more of the intermediate steps — the question of who owns the outcome does not get cheaper because the steps did.
2. War is bound by the laws of armed conflict, which require judgment a human owns. Distinction, proportionality, and necessity are legal and moral judgments about specific circumstances, not pattern-matches against prior cases. A system can flag and estimate; it has no standing to accept the responsibility that lawful force requires. The obligation attaches to a human regardless of how much of the process was automated.
3. Wars turn on the human will to fight and to stop. Whether a population endures, whether soldiers hold, whether an adversary's resolve breaks or hardens, whether a war actually ends — these are human relationships and resolve, not outputs. AI can model, message, and try to erode will at scale; it cannot manufacture the will itself, and the decision to stop fighting remains a human and political act that no automation reaches.
4. Waging war requires judgment under novel, high-stakes ambiguity. A conflict that has gone off-script — deception, a first-of-its-kind situation, an adversary doing what no precedent covers — has nothing clean to match against, and this is exactly where confidently-wrong output is most dangerous, because there is no ground truth to check against in the moment and the stakes are civilizational. Deciding what a genuinely new situation means, and what it is worth risking, stays human.

New axioms

1. When the sensing-to-action loop compresses below human control, states must solve for keeping a meaningful human decision in a loop that now moves faster than deliberation. The pressure is structural: if the adversary's loop is faster, staying slow enough to keep a human genuinely in control is a disadvantage — a use-it-or-lose-it pull toward removing the human that no single state created and none can resist unilaterally. This is the load-bearing new problem; how sharply it bites depends on how far autonomous engagement authority is actually delegated in practice, which is moving fast and worth revisiting.
2. When machine systems react to each other faster than humans can intervene, states must solve for flash-escalation — conflict that climbs the ladder in ways no human chose. Two fast loops interacting can escalate on timescales that leave no room to step back. The failure is not either system being wrong; it is the interaction producing an outcome neither side intended and neither can arrest in time — a stability problem created by speed itself, at a level where the stakes may be existential.
3. When the cognitive work behind military mass is cheap, states must solve for the strategic effect of cheap autonomous mass — force that no longer requires the scarce trained people it used to. What made certain capabilities self-limiting was the cost and human investment behind them; as that falls, more actors can field more, and the restraint that scarcity quietly provided thins.
4. When ambiguity is cheap to manufacture, states must solve for deterrence in a world of eroded attribution. Deterrence assumes an act can be traced to an author who can be punished. As deniable action at scale gets cheaper, the link between attack and accountable actor frays, and a deterrence posture built on retaliation-against-a-known-author loses coherence without an obvious replacement.
5. When systems generate and execute the decisions war used to require humans to make, states must solve for the accountability gap — who answers when an automated action is wrong and no human made the specific call. Political and legal accountability was built for deliberate decisions made one at a time; it does not obviously extend to a system acting many times without a human owning each instance, even as the ownership of the outcome remains inescapably human.

Where it breaks

States field systems that compress sensing-to-action below the speed of deliberation (INVALID #1) and interact with an adversary's systems faster than anyone can intervene (INVALID #2) — while political accountability for war and the judgment the laws of armed conflict require still demand a human who genuinely owns each decision (STILL HOLDS #1, #2, #4), and no one has solved for keeping that human meaningfully in a loop moving faster than they can think (NEW #1, #2). The human authority becomes a nominal owner of a tempo no person can actually weigh, which is the opposite of what political accountability for war was for. The advantage and the danger are the same property: the speed that makes the fast side win is the speed that removes the room to stop.

A second collision: deterrence and the fear of retaliation held partly because a serious attack could be attributed to an actor who could be punished (STILL HOLDS relies on this being intact; the axiom is now INVALID #3), and partly because cheap mass and anonymous action were hard. As attribution erodes and autonomous mass gets cheap (NEW #3, #4), the structure that made deterrence stable — a known author, a scarce and costly capability, a predictable human hand on escalation — thins at several points at once, while the instruments states still reach for assume all of it holds. How fast this destabilizes depends on how quickly autonomous authority is delegated and how cheap deniable mass actually becomes — both fast-moving, both worth revisiting.

ai-axioms.vercel.app/axioms/government/warfare-ai-axioms