Artificial Intelligence and National Security
There's a moment in most technology transitions when the thing stops being a tool and starts being an environment — when it's no longer something you pick up and put down, but something you operate inside of. Nuclear weapons did this to geopolitics in the 1940s. The internet did it to intelligence and commerce in the 1990s. Artificial intelligence is doing it now, and the transition is moving faster than most governments are prepared to handle.
The difference this time is that AI doesn't just change what militaries can do. It changes the speed at which decisions have to be made, the volume of information that has to be processed, and the degree to which human judgment stays in the loop at all. That last part is what makes the national security dimension of AI genuinely different from every prior technology shift.
From Chess Engines to Battlefield Systems
For most of its history, AI was a research curiosity with limited real-world application. Game-playing systems — chess engines, Go programs — were impressive demonstrations, but they operated in closed, rule-bound environments. The intelligence community didn't lose sleep over Deep Blue.
What changed was the convergence of three things: massive datasets, cheap computing power, and neural network architectures that actually work at scale. When those came together in the early 2010s, AI stopped being a narrow tool and became a general-purpose capability. And general-purpose capabilities, historically, end up in weapons systems.
Today's military applications span a wider range than most public discussion acknowledges. Surveillance and reconnaissance are the obvious starting point — AI systems can process satellite imagery at a volume no human analyst team could match, flagging military movements or infrastructure changes in near real-time. The U.S. military's Project Maven, launched in 2017, used computer vision algorithms to analyze drone footage, a job that had previously required analysts to watch hours of video per day. The project generated controversy when Google employees objected to their employer's involvement, but the underlying capability — automated object recognition in military imagery — has continued to develop across multiple contractors and government programs.
Beyond surveillance, AI is reshaping signals intelligence, predictive maintenance for complex weapons systems, logistics optimization, and cyber operations. Each of these is important. None of them is quite as consequential as what's happening to decision-making speed.
The Speed Problem
Modern warfare has always had a pace problem. Information arrives faster than it can be processed, decisions have to be made with incomplete data, and the fog of war is the baseline condition, not the exception. Historically, militaries compensated through doctrine, training, and chain-of-command structures that compressed complexity into manageable decisions.
AI disrupts this in both directions. On one hand, AI systems can analyze sensor data, model probable enemy actions, and surface decision recommendations faster than any human staff process. The U.S. Air Force's ABMS (Advanced Battle Management System) is designed around exactly this idea — an architecture that fuses data from satellites, aircraft, ground sensors, and ships, then surfaces actionable intelligence to commanders faster than current systems allow.
On the other hand, this speed advantage only holds if adversaries aren't doing the same thing. And they are. The practical result is pressure on both sides to reduce human decision latency — which means reducing human involvement. An AI system that recommends an action and waits for human approval is slower than one that acts. In a conflict where response windows are measured in seconds, that latency matters.
This is the mechanism, more than anything else, through which AI creates genuinely new risks in national security. Not because the systems are malicious, but because competitive pressure pushes toward automation, and automation at sufficient speed means humans stop being decision-makers and start being supervisors — and then, when things move fast enough, bystanders. See also: autonomous weapons and the human control question.
Who's Actually Building What
The honest answer is that public information about military AI programs is incomplete almost by definition. What's visible is the tip of a very large iceberg.
The United States has the most visible and extensively documented programs. The Pentagon's Chief Digital and AI Office has been coordinating AI adoption across the services since its predecessor JAIC was established in 2018. DARPA runs dozens of relevant programs at any given time, from autonomous drone swarms to AI-assisted cyber defense. The intelligence community — CIA, NSA, DIA — has its own parallel set of programs that receive almost no public disclosure.
China is the other major actor, and it's worth being specific about what's known versus what's inferred. The Chinese government has publicly committed to AI leadership as a national strategic goal, with a 2017 plan targeting global AI dominance by 2030. The People's Liberation Army has integrated AI into training, logistics, and weapons development, and Chinese companies including Hikvision, SenseTime, and others have built surveillance infrastructure deployed across Xinjiang that represents one of the most extensive AI-powered monitoring systems ever constructed. What remains less clear is how mature China's military AI is for actual combat operations versus administrative and logistical applications.
Russia presents a different profile. Russian military doctrine has emphasized autonomous and remotely operated systems — the Uran-9 ground combat robot, the Poseidon nuclear torpedo, various drone programs — but independent assessments suggest Russian AI capabilities lag behind both the U.S. and China in the underlying research base. What Russia has demonstrated competence in is information warfare: the Internet Research Agency's social media operations during the 2016 U.S. election cycle were an early example of algorithmically assisted disinformation at scale.
Other actors matter too. Israel's military AI programs are extensive and less constrained by the oversight requirements that shape American programs. South Korea has developed autonomous weapons systems for deployment along the DMZ. India is investing heavily in AI for border surveillance. The field is not bipolar.
The Risks That Don't Get Enough Attention
Public discussion of AI and national security tends to cluster around a few dramatic scenarios: killer robots, AI-triggered nuclear war, superintelligent systems that go rogue. These are not obviously impossible, but they crowd out risks that are more immediate and more tractable. For a broader overview, see our piece on AI risks beyond the headlines.
Algorithmic bias in targeting systems. Military AI trained on historical data will reflect historical patterns. If those patterns encode biases — about what vehicles appear in adversary territory, what behaviors indicate threat, what faces appear in watch lists — the systems will perpetuate and accelerate those biases at machine speed. The consequences in a combat environment are not abstract. A 2019 study by MIT Media Lab on facial recognition accuracy disparities across demographic groups remains one of the clearest illustrations of how this plays out technically.
Adversarial attacks on AI systems. Neural networks can be fooled in ways that exploit their architecture rather than their knowledge. Adversarial examples — inputs specifically crafted to cause misclassification — have been demonstrated in image recognition, speech processing, and other domains. An adversary who understands how a targeting system works may be able to manipulate its inputs rather than defeat it directly. This is a known vulnerability with no fully satisfying defense.
Escalation dynamics in cyber conflict. AI-assisted cyber operations are already deployed. They can identify vulnerabilities, craft attacks, and adapt to defenses faster than human operators. The risk isn't just that they're more effective — it's that they compress the time between initial conflict and serious escalation in ways that existing crisis management frameworks weren't designed to handle. Related: how AI is changing the cybersecurity landscape.
The accountability gap. When an autonomous system makes a decision that results in civilian casualties, it's genuinely unclear who is responsible. The operator who deployed it? The programmer who designed it? The commander who approved its use? This isn't a philosophical puzzle — it's a practical problem for international humanitarian law that remains unresolved.
Ethics Without Answers
The ethical arguments around autonomous weapons have been running for about a decade, and they haven't converged on much.
The Human Rights Watch / Harvard Law School "Campaign to Stop Killer Robots" has pushed for a preemptive ban on fully autonomous weapons since 2013. Their argument rests on something like a dignity claim: there is something wrong, independent of consequences, with delegating life-and-death decisions to a machine. Military necessity can justify many things; it cannot justify removing human moral agency from lethal force.
The counterarguments are uncomfortable but not obviously wrong. Human soldiers make poor decisions under stress, exhaustion, and fear. They commit atrocities. An AI system, the argument goes, might in principle be more consistent in applying rules of engagement than a human soldier in a firefight. This doesn't prove autonomous weapons are good — it proves the ethical landscape is more complicated than either side of the debate usually admits. For a deeper dive into this tension, see the ethics of machine decision-making in conflict.
What's harder to argue around is surveillance. The Xinjiang surveillance infrastructure has documented what large-scale AI-assisted population monitoring looks like in practice: continuous tracking, behavioral prediction, preemptive detention based on algorithmic scores. The technology exists, it works, and authoritarian governments have shown no reluctance to deploy it. The question of whether democracies will constrain their own use of similar technology — or whether competitive pressure will erode those constraints — is open.
The Governance Gap
International law governing armed conflict — the Geneva Conventions, the laws of war — was designed for weapons that humans operate. It assumes a commander who can be held accountable, a targeting process with human judgment, a distinction between combatants and civilians that someone is responsible for maintaining.
Autonomous weapons don't map cleanly onto any of this. The UN Convention on Certain Conventional Weapons has held discussions on lethal autonomous weapons systems since 2014 without producing binding rules. A handful of countries — Austria, Chile, New Zealand among them — have called for a preemptive ban. Major military powers have not.
The dynamic resembles early nuclear arms control — except that AI systems can be developed faster, deployed more widely, and are harder to verify than nuclear warheads. There's no equivalent of satellite imagery for confirming that a country has or hasn't deployed autonomous targeting systems.
What governance frameworks exist tend to be national rather than international. The U.S. Department of Defense AI ethics principles, published in 2020, require "appropriate levels of human judgment" in lethal force decisions — language that is deliberately vague. The EU's AI Act creates categories of high-risk AI, but it's explicitly scoped to civilian applications and doesn't touch military systems. There is, at present, no binding international framework governing military AI. See also: a timeline of AI governance milestones.
The Next Decade
Predicting AI development specifically is hard enough. Predicting how it intersects with geopolitics, domestic politics, arms control, and conflict is harder still. But some things seem reasonably likely.
The competition between the United States and China will continue to drive military AI investment by both sides and pull allied nations along. The pressure toward autonomy — more systems making more decisions without human review — will increase as the strategic logic of speed becomes harder to argue with. Disinformation capabilities will grow faster than the ability to detect or counter them.
Less certain but worth watching: whether any major conflict in the next decade provides a real-world test of AI-enabled military operations at scale, and what lessons get drawn from it. The Russia-Ukraine war has already demonstrated drone autonomy and AI-assisted targeting in ways that defense establishments worldwide are analyzing carefully.
The governance trajectory is discouraging. The gap between technological capability and regulatory framework tends to widen before it narrows, and it's not obvious what shock — short of a serious incident involving autonomous weapons — would accelerate international agreement.
Asset or Threat?
The framing in most policy discussions sets these up as alternatives: AI will either make us safer or it will be our undoing. The honest answer is that it will almost certainly be both, in different contexts and at different times, and the balance will depend on choices that haven't been made yet.
What AI definitely does is accelerate. It makes surveillance more thorough, attacks more targeted, defenses more responsive, disinformation more scalable. Whether that acceleration benefits stability or undermines it depends on who's deploying it, against whom, under what constraints, and with what accountability.
The technology itself doesn't resolve any of those questions. A military AI system is not inherently stabilizing or destabilizing — it depends on the doctrine it operates under, the verification mechanisms that surround it, the legal frameworks that govern its use, and the political relationships between the actors involved.
What concerns me, more than any specific application, is the governance lag. The systems are being built faster than the rules. The competitive pressures are real, and they push toward less human oversight, not more. And the international community has shown, so far, a remarkable capacity to discuss these problems without resolving them.
Comments
Post a Comment