A famous bit of Nazi cope from the final days of World War II was the idea that a bunch of Wunderwaffen (superweapons) would miraculously turn the tide against the Allies. Of course, even if Germany had produced more jet fighters or had some of its comically large experimental tanks roll off the production lines, it wouldn’t have mattered – especially because the Allies were busy developing atomic bombs. Regardless, the idea of war-winning superweapons outlasted the flailing propaganda apparatus that insisted victory was right around the corner even as the Red Army closed in on the Führerbunker.

From Bond villains to the Death Star (the best Wunderwaffen tend to be space lasers), popular fiction is full of stories featuring a singular weapon that could win a war all on its own or make any form of resistance totally futile. Yet wars in the real world don’t work that way. Regimes with nuclear weapons don’t always get what they want (see Russia and Ukraine). Even general technological superiority doesn’t guarantee a win (see American experiences in Vietnam and Afghanistan). So what gives? And why do so many people think that AI is a game changer?

Let’s start with the AI hype case. Current Pentagon leadership and the broader US military-industrial complex are both aflutter with the prospect of making the US military AI-enabled to boost lethality. The idea behind the buzzwords is that a military that feeds information from a network of sensors on satellites, aircraft, ships, and ground stations into an AI system will be able to identify and shoot targets at a much faster speed than manual coordination. In defense-speak this is called shortening the OODA loop (Observe, Orient, Decide, Act). A system or a unit with a high degree of automation would be able to act and react at superhuman speeds, giving them an advantage on the battlefield.

The US war against Iran and Israel’s bombardment of Gaza have both prominently featured military uses of AI. The Department of Defense(currently identifying as the Department of War) reportedly used Claude to rapidly assemble target lists in Iran in the lead up to the war while simultaneously deeming Anthropic a risk to national security for being too woke. Israel employed an AI system called Lavender to identify Hamas and Islamic Jihad militants, assigning individuals a rating of 1-100 of how likely they were to be an enemy combatant. In both of those cases, the automated systems significantly reduced the review time of target analysis and saw examples of errors and/or negligence that led to strikes on innocent people. Mistakes and war crimes happen with fully manual systems as well. My point here is not to indict the moral failures of military AI, but to evaluate its strategic effectiveness. Even in the callous terms of a commander wholly unconcerned in civilian suffering, neither campaign was a slam dunk despite one side having this supposedly decisive technological advantage.

The reason why is in war the enemy gets a vote. Belief that a super-intelligent AI or a space laser will guarantee victory is to fall victim to the fallacy of the last move. Unlike a game with set rules or an end state, war is an unbounded competition between actors capable of learning and adapting. Edward Luttwak1 highlights what he calls the “paradoxical logic of strategy” wherein the more effective your action or strategy is, the more incentive your enemy has to counter it. Therefore massive gains in combat effectiveness by technology or tactic are generally short-lived because the opposing side will do whatever it can not to lose.

Furthermore, responses to military innovations don’t have to be symmetric. You don’t necessarily need to develop your own super expensive AI system to counter your enemy’s. For AI image recognition systems in particular, there are a lot of looney tunes-esque counter moves. Iran has painted decoy plane silhouettes on tarmacs to make the US and Israel waste munitions. Russia placed tires on top of its planes to confuse image recognition software. In Four Battlegrounds, Paul Scharre has some great examples of US Marines outfoxing an AI-enabled surveillance tower, sneaking past it by hiding under a cardboard box like Solid Snake giggling the whole time. While you can adjust a system to account for any specific tactic, there will always be another one, especially if all it costs is some paint and/or cardboard.

On top of all that, despite the best efforts of any military or security service, no surveillance system is truly ubiquitous or omniscient. There will be gaps in satellite or aerial coverage. Even if your side is lucky enough to have pervasive sensors, the data will still need interpretation. The enemy can present fake targets and hide real ones. Even with accurate targeting information, combatants will face other restraints like limited munitions, travel times for delivery systems, or contested air and ground environments. If the enemy has sophisticated detection capabilities of their own, commanders will have to weigh the advantage gained by firing on targets they’ve found with the risk of exposing their forces to counter fire when they give away missile or artillery launch positions.

Carl von Clausewitz wrote “Everything is very simple in war, but the simplest thing is difficult.” He called all of the confounding factors big and small that make things difficult ‘friction.’ Units get lost in the woods, supplies get delayed, vehicles get stuck in the mud, orders get misunderstood – all of that is friction. And that’s before the enemy starts shooting at you. Militaries with AI models assisting their analysis and decisionmaking will likely be able to make decisions faster. I’m not sure if they’ll necessarily make decisions better. Even if they do, there’s a whole host of problems like dust getting stuck in helicopter air intakes Claude (or any other model) can’t solve for you.

As we’re seeing in real time, an AI enabled force is no guarantee of victory, especially if you don’t have strategic clarity on what exactly you’re trying to achieve through the use of force. While that may be bad news for the current on-again off-again US campaign against Iran, it’s good news for any groups of people who have to resist force applied by an AI-enabled opponent. And sooner or later that might be more of us than you might think.

Footnotes

  1. The guy is a problematic weirdo, but he came up with at least one good idea in the 80s. ↩