What changes for the military and defense with AI?

The shift

Synthesizing intelligence from scattered sources, analyzing targets, planning operations, and running logistics — the cognitive work of preparing a military decision — goes from scarce, slow staff-hours to near-instant and abundant. The constraint on military action was rarely knowing that force was possible; it was the time and trained people needed to assemble the picture, and increasingly the loop from sensing to acting is being compressed by autonomous systems that don't wait for a human to read the picture at all.

The axioms

• Intelligence synthesis is bottlenecked by analysts who can fuse sensor feeds, signals, and reports into a usable picture. Rests on fusion-of-large-inputs being scarce, expensive labor.
• Targeting analysis — identifying and prioritizing what to strike — requires trained specialists working slowly enough to be deliberate. Rests on target-development synthesis being scarce and time-costly.
• Operational planning and wargaming are gated by the number of staff officers who can generate and compare courses of action. Rests on planning and option-generation being scarce expert labor.
• Logistics — moving the right materiel to the right place — is bottlenecked by planners tracking a combinatorial problem by hand. Rests on large-scale optimization being scarce.
• A commander must be accountable for the decision to use lethal force. Rests on accountability being a property only a liable human can hold.
• The use of force is bound by the laws of armed conflict, which require moral and legal judgment a human owns. Rests on legal/moral agency being irreducibly human.
• Fighting requires judgment under fog-of-war novelty — deciding when the situation on the ground doesn't match any plan. Rests on judgment under genuine novelty being scarce.
• Wars turn on human will, morale, cohesion, and trust — soldiers' and populations' willingness to keep fighting or stop. Rests on will and trust being human and not manufacturable on demand.
• Physical force — holding ground, moving, striking, defending — requires assets and people in the physical world. Rests on physical action being irreducible.

Invalid axioms

1. Intelligence synthesis is bottlenecked by analysts who can fuse feeds into a picture. Fusing satellite imagery, signals, open-source, and field reports into a coherent read is synthesis-of-large-inputs at scale — exactly what AI does cheaply and fast. The habit-trap: staffing intelligence shops sized for a world where every assessment took a human days to assemble, and treating analyst headcount as the throughput limit rather than the review-and-verification step that now is.
2. Targeting analysis requires specialists working slowly enough to be deliberate. First-pass identification, correlation, and prioritization can now be generated in seconds. The habit-trap: assuming the deliberation was in the analysis time itself, when the load-bearing act — confirming the target is what the model says and lawful to strike — was never the speed of the synthesis and doesn't get faster with it.
3. Operational planning and wargaming are gated by staff officers generating courses of action. Generating, comparing, and stress-testing plans against thousands of scenarios is option-generation at near-zero marginal cost. The habit-trap: sizing planning staffs and timelines around hand-built courses of action when the scarce step has moved to choosing among options and owning the choice.
4. Logistics is bottlenecked by planners tracking a combinatorial problem by hand. Routing, sequencing, and demand-forecasting across a supply network is optimization AI handles well. The habit-trap: budgeting logistics as a headcount problem rather than a verification-and-resilience problem — the plan is cheap; knowing it survives contact and disruption is not.

Unchanged axioms

1. A commander must be accountable for the decision to use lethal force. Accountability requires someone who can be court-martialed, prosecuted, or relieved. A model cannot be liable, so every AI-assisted targeting recommendation or engagement still needs a human who owns the outcome — this becomes more load-bearing, not less, as AI-generated targeting options multiply faster than any one commander can genuinely weigh them.
2. The use of force 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. AI can flag and estimate; it has no standing to accept the moral and legal responsibility that lawful force requires.
3. Fighting requires judgment under fog-of-war novelty. AI matches against precedent; a battlefield that has gone off-script — deception, a situation no one has faced — has no clean precedent to match, and this is exactly where confidently-wrong output is most dangerous because there's no ground truth to check against in the moment.
4. Wars turn on human will, morale, cohesion, and trust. Whether soldiers hold, whether a population endures, whether allies stay bound to a commitment — these are human relationships and resolve, not outputs. AI can model and message around them; it cannot manufacture the will itself.
5. Physical force requires assets and people in the physical world. Holding terrain, moving materiel, defending a position — none of this is token generation. AI can decide where to send force faster; the force still has to physically be there and act. (This one is moving: uncrewed and autonomous systems shift who is physically present, though not the requirement for physical presence itself — flagged as fast-moving below.)

New axioms

1. When AI compresses the loop from sensing to recommended action below human reaction time, forces 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 to keep a human genuinely in control becomes a disadvantage — a use-it-or-lose-it pull toward removing the human that no single commander created and none can resist alone.
2. When engagements can be executed by autonomous systems, forces must solve for who is accountable when an autonomous action is wrong — and how that accountability survives when no human made the specific call. Individual sign-off was built for one deliberate decision at a time; it doesn't obviously extend to a system acting many times without a human in each instance.
3. When AI presents fluent, confident targeting and intelligence recommendations, forces must solve for automation bias — operators deferring to the machine's read precisely when time is short and the output looks authoritative. The failure isn't the model being wrong occasionally; it's the human review becoming a rubber stamp exactly when verification matters most.
4. When both sides run AI-driven decision systems, forces must solve for escalation dynamics that emerge from machines reacting to each other faster than humans can intervene. Two fast loops interacting can escalate in ways neither side chose, on timescales that leave no room to step back — a stability problem created by speed itself.
5. When AI does the synthesis, planning, and targeting analysis that once trained a commander's judgment, forces must solve for deskilling — how officers develop the judgment to overrule the machine when the reps that built that judgment have been automated away. The scarce thing (judgment under novelty) is exactly what atrophies when the practice that produced it is handed to AI.

Where it breaks

Forces field AI that compresses sensing-to-recommendation below the speed of human deliberation (INVALID #1, #2) — while accountable command over lethal force still requires a human to genuinely own each decision (STILL HOLDS #1) and no one has solved for keeping that human meaningfully in a loop moving faster than they can think (NEW #1, #3). The human sign-off becomes a rubber stamp on a tempo no reviewer can actually weigh, which is the opposite of what command accountability was for — and automation bias makes the stamp feel earned.

A second collision: the same speed that makes AI-assisted planning and targeting an advantage (INVALID #2, #3) creates escalation dynamics when the adversary runs it too (NEW #4), so the pressure to compress the loop further — to not be the slower side — pushes directly against the deliberation that lawful, accountable, judgment-bound force requires (STILL HOLDS #1, #2, #3). The advantage and the danger are the same property. How fast this bites depends on how quickly autonomous engagement authority is actually delegated in practice — a fast-moving call worth revisiting.

ai-axioms.vercel.app/axioms/government/military-defense-ai-axioms