TL;DR:
Epistemic transparency: first piece on this topic; background in science, communication, and geophysical risk, not strategy, or AI S/G/E. Keen for input from people closer to the field. Have I completely misapprehended the dynamic?
For the last seventy-odd years, the dubious logic of nuclear peace has rested on a single, almost paradoxical guarantee: that no one can stop the other side from hitting back. Mutually assured destruction (MAD) doesn’t work because either side wants to be destroyed; it works because both sides know that a first strike would leave enough of the other’s arsenal intact to make retaliation certain. Hidden submarines under the ocean, mobile launchers rolling through forests, or silos scattered across a continent have long been the quiet default insurance policy of the atomic age. As long as an adversary’s forces could not be found in full, they could not be destroyed in full, and so no one had enough incentive to try.
Artificial intelligence threatens this fairly black-and-white logic. Not because it can build a bigger bomb (the physics of fission and fusion haven’t changed) but because it is very good at doing something else entirely: finding things. Things that don’t want to be found. Pattern recognition across satellite imagery, acoustic sensor networks, and signals intelligence is turning what used to be needles in oceans of noise into targets that can be identified and tracked in something close to real time. The stakes are not abstract. They sit underneath the question of whether the world’s nuclear powers can still trust that their retaliatory forces will survive long enough to retaliate, and what they might be tempted to do if they start to doubt it.
Deterrence by MAD depends on one property above all: survivability. It doesn’t matter how many warheads a country has if an adversary can eliminate them all in a single opening move. The nuclear peace has been maintained by the near-impossibility, for most of the atomic age, of achieving this: a “disarming” or counterforce first strike that destroys an opponent’s nuclear forces before they can be used. You make the first move, and it is the last move.
Two things have historically made this impossible: hardening and concealment. Hardening means burying a missile silo under enough concrete and steel that even a nuclear detonation nearby won’t destroy it. Concealment means never letting the enemy know where all your weapons are in the first place: submarines patrolling the deep ocean, road-mobile launchers dispersed into forests and mountains, warheads kept off any map that an adversary’s intelligence services could plausibly compile. Concealment has always been the cheaper of the two and, in many ways, the more powerful guarantee. There’s no need to harden the weapons your adversary can’t find.
As the political scientists Keir Lieber and Daryl Press have argued, this is precisely the balance that new technology has been quietly upending.[1] If concealment fails, i.e., if hidden forces become findable, then the entire architecture of survivability built on mobility and dispersal begins to look far more fragile than anyone had assumed.
The mechanism is not exotic. At its root, it’s a problem of signal and noise, and that happens to be a problem machine learning is unusually good at solving. A road-mobile missile launcher trying to hide from a rival’s intelligence services has, historically, benefited from an overwhelming asymmetry: there is simply too much ocean, too much forest, too much open ground, and too much raw sensor data for any human analyst to sift through in time to matter. It would be something like Where’s Wally (or Waldo), but instead of finding him on one page, you have to find him in all the books ever printed, and when you have found him, he’s already moved. Lieber and Press call the change now underway a revolution in remote sensing; a combination of persistent satellite constellations, underwater sensor networks, and cheap commercial imagery that together produce a volume and continuity of data no analyst pool could ever process by hand, or eye.[1]
AI removes the asymmetry. Pattern-recognition systems can comb through synthetic-aperture radar imagery, acoustic signatures, and electromagnetic emissions continuously and simultaneously, flagging anomalies that a human eye would miss and correlating them across sensor types in ways that used to require weeks of dedicated analysis. Researchers pointed to open-source analysis of North Korean missile bases using little more than commercial satellite photography as an early proof of concept, and noted that a moderately sized constellation of synthetic-aperture radar satellites could plausibly provide near-continuous coverage of most plausible mobile-missile deployment areas.[2] What used to require a lucky break or a well-placed human source increasingly requires only enough compute and enough sensors.
This is the counterforce logic in its rudest form. If a state can locate an adversary’s mobile launchers and submarines with reasonable confidence, and if its own missiles are accurate enough to destroy them once found, then the theoretical possibility of a disarming first strike stops being theoretical. It becomes a live strategic option, one intelligence agencies are obligated to plan for, and one adversaries have to assume the other side is planning for, whether or not it’s ever intended to be used. Not every analyst agrees that the shift is as decisive as Lieber and Press suggest. Critics note that hiders have historically adapted faster than hunters expect, and that no sensor advance has yet proven so complete that it forecloses concealment entirely.[3] Dr. Jingjie He talked about it recently at UNIDIR, citing studies that can make a car disappear using an adversarial patch on the roof, or even driving it into a box demarcated by the same patch.[9] The Adversarial Patch has been around for almost a decade, in fact, making headlines when researchers turned a banana into a toaster.[10] So, although the direction of travel, and the anxiety it generates in planning rooms, is entirely justified, the tug-of-war between detection and concealment continues.
The real danger cuts in two directions, and that’s what seems to make it genuinely destabilising rather than simply a shift in who has the possible advantage.
For the side gaining the tracking capability, the temptation is obvious even if it never intends to act on it. A state that believes it can locate and destroy most of a rival’s retaliatory force has, for the first time, something resembling a plausible disarming option, not necessarily one it plans to use, but one that could change the shape of every crisis calculation that follows. Deterrence theorists have long worried less about deliberate first strikes in peacetime than about what happens when that option exists on the table during an acute crisis, when misperception and time pressure are already running high.[4]
For the side whose forces are becoming findable, the pressure runs the other way, and it’s arguably more dangerous. It might incite a classic “use them or lose them” dynamic: if a country believes its nuclear forces might be located and destroyed before they can be launched, the rational response under extreme pressure is simply to launch them first, or at least to adopt launch postures, such as delegating authority further down the chain of command, or shortening the fuse on retaliation. Clearly accidental or hair-trigger decisions leading to war become more likely.[5, 11, 12] Neither side has to want war for this dynamic to raise its probability. Crisis instability, in the language of strategic studies, is not about intentions; it’s about what each side believes the other might be forced to do under pressure, and AI-enabled ISR raises the stakes of that belief on both sides simultaneously.
There’s a second, compounding layer to this, and it has less to do with finding targets than with what happens once something has been found. AI doesn’t just improve detection; it compresses the time available to think about what to do with what’s been detected. Human thinking. Systems designed to fuse sensor data and flag threats in real time inherently push decision-making toward faster tempos, and faster tempos leave less room for the kind of deliberate, skeptical judgment that has, on more than one occasion during the Cold War, prevented a false alarm from becoming a launch.[6]
Layered onto compressed timelines is the well-documented tendency researchers call automation bias: the tendency of stressed decision-makers under time pressure to defer to a machine’s assessment rather than interrogate it. A 2018 RAND publication on this dynamic warned that as AI systems demonstrate ever more analytical sophistication, human commanders are likely to start treating their recommendations as equal to or superior to those of human advisors, even in circumstances where the system’s confidence is unearned.[7] Military and political leaders have repeatedly pledged that a human will always remain in the loop on any nuclear launch decision. Those pledges are almost certainly sincere. They are also, as things stand, largely unverifiable; there is no external mechanism by which one state can confirm that another’s human-in-the-loop is meaningfully deliberating, rather than rubber-stamping a machine’s threat assessment thirty seconds before end-of-flight.[8] The paradigm then becomes human-on-the-loop.
None of this means nuclear war is more likely tomorrow than it was yesterday, and it would be a mistake to treat this as an argument that catastrophe is now inevitable. Deterrence has proven more adaptable than pessimists have historically given it credit for, and concealment techniques have a long history of evolving faster than the sensors built to defeat them. But, I think, it would be an equally serious mistake to treat this as a manageable technical footnote. The specific thing that has kept the nuclear peace, the near-impossibility of finding what an adversary wants hidden, is being eroded by a technology whose whole purpose is finding things. And no one has built the rules to compensate for that.
What might those rules look like? Probably some combination of things that are individually modest and collectively hard: verifiable commitments to human control over launch decisions, rather than merely rhetorical ones; mutual restraint on the specific application of AI to counterforce-relevant ISR, even where restraint elsewhere in AI development is unrealistic; and transparency measures that let rival states confirm, without compromising their own sensors, that the other side isn’t quietly closing the detection gap faster than anyone has agreed is safe. None of that solves the underlying problem. But the question worth mulling over isn’t whether AI will change nuclear strategy; it already has. It’s about whether anyone will build the guardrails before the technology fully arrives, rather than panicking to rush it through after the fact.
