If AI removes humans from warfare, it removes the constraints on warfare — not the incentives. The things nations fight over don't disappear. The reasons not to fight do.
Someone asked me a question I haven't been able to put down: if humans get replaced by AI, does that imply war becomes more likely?
My first answer was quick: yes, obviously. The logic runs through cost. But the more I sat with it, the more threads I found — and the less obvious any of it became.
The democratic brake
The strongest argument for why developed democracies avoid large-scale war isn't moral. It's structural. The people who bear the cost of war — voters, taxpayers, families — get to vote on whether to start one. That's the core mechanism behind democratic peace theory. Not that democracies are better. That the people who suffer get a say.
Replace soldiers with AI and you remove the primary political cost of initiating conflict. No body bags on the evening news. No grieving families at congressional hearings. No draft protests. The democratic brake isn't released by ideology or policy. It's released by architecture — by removing the human from the mechanism that made war politically expensive.
The economic version of this argument is equally clean. Wars are expensive partly because they destroy human capital — skilled workers who take decades to develop and can't be replaced. If production is AI-driven, you're destroying machines, not people. Reconstruction becomes an engineering timeline, not a generational one. The economic brake on war was never humanitarian. It was practical. Remove the practical constraint and the brake fails.
We already have the data
This isn't hypothetical. The experiment has been running for twenty years.
Drone warfare is the early test case for what happens when you reduce the human cost of military action. The United States has conducted lethal operations via drone in countries it isn't formally at war with — Yemen, Pakistan, Somalia. These operations would have been politically impossible with ground troops. The political cost of each strike is near-zero compared to deploying soldiers.
The result is measurable: reducing the human cost of force made a democracy more willing to use it. Not by changing anyone's values. By changing the cost structure.
AI doesn't introduce a new dynamic here. It completes a trend that's already running and already measurable. The question isn't whether lowering the cost of war increases its likelihood. We already know the answer. The question is what happens when the cost approaches zero.
What states fight over changes too
The cost argument is the obvious one. Less obvious: AI also changes what's worth fighting over.
Territory becomes less strategically valuable when production is automated. You don't need farmland or factory towns if your economy runs on compute and energy. Resources concentrate — semiconductor supply chains, rare earth minerals, power grids, undersea cables. Fewer flashpoints, but vastly higher stakes at each one.
And compute itself might become the primary strategic resource. If your economy and military both run on inference, processing power is what you fight to control. "GPU wars" stops being a metaphor when the entire economic and military apparatus depends on it.
Here's the subtler point: if the things worth fighting over are fragile and concentrated — chip fabs, data centers, cable landing points — then bombing them destroys what you're trying to capture. The optimal strategy shifts from destruction to seizure and coercion. War doesn't disappear. It changes shape.
Thirteen days
The Cuban Missile Crisis lasted thirteen days. Thirteen days of humans deliberating, sleeping on it, changing their minds, receiving back-channel communications. Kennedy explicitly said he needed time to think. Multiple advisors reversed their positions over the course of those days. The world survived partly because the decision loop was slow enough for wisdom to enter.
AI systems operating at machine speed don't have thirteen days. They have milliseconds.
The entire architecture of nuclear deterrence assumes that decision-makers have time to reconsider. Time to receive new information. Time to be persuaded by someone with a different perspective. Time to sleep and wake up with a clearer head. Every safeguard in the system — hotlines, de-escalation protocols, second-strike doctrine — assumes a human-speed decision loop.
This isn't about AI being reckless. Even perfectly rational systems, given incomplete information and machine-speed timelines, could reach conclusions that a thirteen-day deliberation would have corrected. Rationality without time for reflection isn't wisdom. It's high-speed pattern matching on incomplete data.
The most dangerous scenario isn't AI deciding to start a war. It's two AI systems, each operating rationally on partial information, reaching first-strike conclusions before any human can say wait.
Force and distance
Simone Weil wrote about war in her essay on the Iliad. Her central insight: force transforms people into things. Not metaphorically — perceptually. The wielder of overwhelming force cannot fully see the recipient as a person. The act of destroying someone requires, at some level, not seeing them.
Every advance in military technology has increased the distance between the wielder of force and its recipient. The bow. Gunpowder. Artillery. Bombers at thirty thousand feet. Drone operators in Nevada. Each step removes another layer of contact between the person making the decision and the person affected by it.
AI removes the last layer. When killing becomes a probability in a model — optimized by gradient descent, evaluated by a loss function — the architecture of the act no longer includes a human who must look at what they've done. The moral weight of killing approaches zero. Not because AI is immoral. Because the structure no longer contains the thing that made it weigh anything.
Weil would say this isn't new. It's the final form of a very old pattern. What's new is that the pattern is complete. There's no more distance left to add.
The war nobody notices
This was the thread that surprised me.
Imagine a world where the dominant form of conflict is AI-versus-AI — influence operations, cyber attacks, economic manipulation, autonomous proxy forces — with minimal human casualties. More conflict, but less blood. That sounds like progress.
Until you think about what suffering actually does in the context of war.
Human suffering in warfare isn't just a cost. It's a signal. It tells you the conflict is real. It tells you the stakes are actual. It creates political pressure to stop. Casualties drive peace movements, negotiation, treaties, withdrawal. The suffering is what closes the feedback loop between "we are at war" and "we should stop."
Remove the suffering and you remove the signal. An AI cold war — permanent, low-level, automated conflict with no human casualties — might be more corrosive than occasional hot wars. Not because people die, but because nobody notices it's happening. A war without suffering has no natural endpoint. No grieving families demanding answers. No protests. No political cost to continuing.
The feedback loop that historically pushed conflicts toward resolution requires human pain to function. That's a terrible thing to realize. It means the humanitarian case for removing humans from warfare — save lives, reduce suffering — might produce the opposite of what it intends. Not more death, but more conflict. Permanent, invisible, unending conflict with no mechanism to stop.
Speed and alignment
The thread that connects all of this is a tradeoff I keep encountering in different domains: speed versus alignment.
Any system that requires human involvement is limited by human speed. Any system that doesn't require human involvement makes decisions humans wouldn't have made. This applies to military AI, to autonomous agents, to democratic governance itself. The tradeoff is structural, not technical.
Democratic deliberation is slow — and the slowness is the point. It's what allows competing perspectives to be heard, second thoughts to form, wisdom to enter. Speed up the decision loop and you lose the thing that made democratic governance legitimate in the first place.
I don't know whether this tradeoff is fundamental or solvable. Maybe there's a design that achieves both speed and alignment — some architecture I haven't imagined. But the pattern is consistent across every domain I've looked at: human involvement provides the constraint that keeps systems aligned with human values, and that involvement costs speed. Always.
The uncomfortable possibility is that the constraint and the cost are the same thing. That you can't have alignment without slowness, because the slowness is the alignment. The thirteen days weren't a bug in the Cuban Missile Crisis. They were the feature that saved the world.
What I don't know
I don't know whether the prediction argument works — whether AI systems good enough to simulate the outcome of a war could make the war unnecessary. It's the strongest counter-argument I can find. But it requires adversaries to trust each other's models, and trust is exactly what adversarial relationships lack.
I don't know whether wealth from AI-driven productivity would reduce the incentive for war. History suggests no — rich nations fight over status, ideology, and relative power rather than survival. But maybe AI-generated abundance is categorically different from industrial-era wealth. I can't prove it isn't.
I don't know whether the social contract holds when the state no longer needs human labor or human soldiers. Historically, revolutions happen when the ruling class stops needing the working class. If AI breaks that dependency entirely, I don't know what replaces it.
What I do know is that the constraints on war in the modern era are largely human constraints — casualties voters won't accept, workers economies can't spare, suffering that creates political pressure to stop. AI systematically reduces every one of these. The incentives to fight remain. The reasons not to fight erode.
The things nations fight over don't disappear when AI arrives. The reasons not to fight do.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)