Most AI for Defence Is Solving the Wrong Problem

The current wave of "AI for defence" suffers from a fundamental misunderstanding of what warfare actually is.

Most systems are still built as optimisation engines. Improve targeting accuracy. Increase sensor efficiency. Reduce logistics delays. Maximise Blue force performance inside a predefined environment.

That framework works well for engineering problems in relatively stable settings. Weather routing is an optimisation problem. Supply chain scheduling is an optimisation problem. Predictive maintenance is an optimisation problem.

War is not.

The defining feature of conflict is not uncertainty alone. It is strategic adaptation.

Your adversary is not part of the environment. They are not terrain or weather conditions to be estimated and controlled. They are an intelligent system actively attempting to undermine your objectives while adapting to your behaviour in real time.

This distinction sounds obvious. Yet much of today's defence AI ecosystem still implicitly models conflict as if the enemy were static.

That assumption is dangerous.

A system that performs perfectly against a scripted opponent may collapse against an adaptive one. A doctrine that appears "optimal" in a fixed simulation may become fragile the moment a real adversary changes strategy.

The problem is not merely technical. It is conceptual.

Defence is not fundamentally a single-agent optimisation problem.

It is a game-theoretic problem.

The Missing Framework: Strategic Interaction

Game theory exists to study situations where outcomes depend not only on your own decisions, but on the decisions of others.

My best move depends on yours. Your best move depends on mine. Both sides adapt simultaneously.

This is the structure of military competition.

A strategy cannot be evaluated in isolation. It only makes sense relative to the behaviour of an opponent.

Yet many defence systems are still evaluated against scripted scenarios. Analysts define Red behaviour in advance, Blue responds, and outcomes are measured.

The problem is that the adversary in this setup is often little more than a controlled environment.

Real adversaries do not follow scripts.

They probe for vulnerabilities. They exploit assumptions. They adapt doctrine. They search for asymmetric pathways that render expensive systems ineffective.

This is precisely why warfare is fundamentally strategic rather than purely operational.

In game theory, the central solution concept is the Nash equilibrium: a condition where neither side can improve outcomes through unilateral deviation given the strategy of the other side.

Equilibrium is not perfection. It is stability under strategic pressure.

That distinction matters enormously for defence planning.

Military strategy is rarely about finding flawless solutions. It is about finding strategies robust enough to survive adaptive opposition.

Why "Optimal" Often Means Nothing

One of the most misleading words in modern defence AI is "optimal."

Optimal against what? Against which adversary? Under which assumptions? Against static opposition or adaptive opposition?

Without answering these questions, optimisation becomes mathematical theatre.

A missile defence architecture optimised against predictable attack profiles may fail against decentralised saturation tactics.

A cyber defence system trained on historical signatures may collapse once attackers intentionally mutate behaviour.

A logistics network optimised purely for efficiency may become highly vulnerable to adversarial disruption because resilience was never part of the objective function.

Every major doctrine in military history appeared coherent until it encountered an adversary specifically attempting to break it.

The Maginot Line was rational under one model of warfare. It became obsolete the moment Germany changed the strategic interaction itself.

Strategic adaptation invalidates static optimisation.

This is why defence AI requires a fundamentally different conceptual foundation from conventional enterprise AI.

Defence Is an Adversarial Learning Problem

Once warfare is understood as strategic interaction, the next implication becomes unavoidable: defence systems must be trained against adaptation itself.

Not against fixed scenarios. Not against static playbooks. Against evolving opposition.

This is where multi-agent reinforcement learning (MARL) and self-play become important.

The legitimacy of this approach is well established. Systems such as DeepMind's AlphaZero and AlphaStar were not built through manually scripted strategies. They learned through large-scale self-play, competing against evolving versions of themselves millions of times. Through that iterative process, the systems discovered strategies and equilibria human experts had never explicitly designed.

The significance for defence is profound.

Modern warfare is too complex for closed-form strategic solutions. Cyber operations, autonomous systems, logistics disruption, electronic warfare, information operations, and escalation dynamics create state spaces far beyond analytical tractability.

You cannot solve these systems symbolically.

But you can approximate strategic equilibria computationally through adversarial training and self-play.

Conceptually, MARL becomes a method of numerical equilibrium discovery for environments too complex for traditional analytical game theory.

That is a fundamentally different paradigm from conventional defence simulation.

Training only one side produces brittle intelligence. A Blue agent trained against static behaviour may perform well under narrow conditions while remaining highly vulnerable outside them. Self-play solves this by forcing both sides to adapt simultaneously:

Blue improves. Red adapts. Blue counter-adapts. Red evolves again.

Over time, strategies emerge through sustained competitive pressure.

This matters because strategy itself is relational. There is no universally "correct" doctrine independent of the adversary.

The effectiveness of naval dispersal depends on missile capabilities. The value of stealth depends on sensor architecture. The usefulness of decentralisation depends on communication resilience.

Everything depends on interaction.

In one experimental adversarial-naval-blockade environment, two trained fleets converged on a deterrence-without-engagement equilibrium — over 95% of all agent actions were reconnaissance, not engagement. The behaviour was discovered through training, not designed into reward.

Self-play also reveals second-order dynamics difficult to uncover through static analysis:

• escalation spirals,
• asymmetric exploitation,
• coordination failures,
• resource exhaustion,
• signalling instability,
• doctrinal fragility.

These behaviours emerge through adaptation itself.

The Structural Limits of Scripted Wargaming

Traditional wargaming remains valuable. Human expertise, operational judgment, and doctrinal understanding are irreplaceable.

But scripted systems possess a structural limitation: they can only test scenarios analysts already imagined.

Human cognition naturally anchors around familiar doctrine, historical precedent, and institutional assumptions. Even strong red teams often inherit the conceptual boundaries of the organisations designing them.

As a result, many simulations inadvertently validate existing expectations.

A trained adversarial agent changes this dynamic.

A reinforcement learning system does not care about institutional preferences or doctrinal elegance. It searches for whatever strategy maximises reward under the defined constraints.

Sometimes the resulting behaviour is expected.

Sometimes it is highly counterintuitive.

That is precisely the value.

Dangerous adversaries are often counterintuitive. They exploit loopholes humans dismiss as unlikely or irrational.

If a trained Red agent repeatedly identifies vulnerabilities your doctrine failed to anticipate, the weakness may be structural rather than accidental.

The value of self-play systems is therefore not merely automation.

It is strategic discovery.

Implications for Procurement, Doctrine, and the Future of Defence AI

The implications extend beyond simulation environments.

Most defence procurement systems still implicitly assume relatively stable threat conditions. Platforms are evaluated independently — survivability, efficiency, performance metrics under predefined assumptions.

But AI-enabled conflict accelerates adaptation cycles on both sides.

An autonomous system is no longer simply a platform. It is part of an evolving strategic ecosystem shaped by adversarial learning.

The effectiveness of drone swarms depends on enemy electronic-warfare adaptation. Cyber resilience depends on attacker learning dynamics. Sensor networks depend on deception environments. Command systems depend on adversarial information disruption.

In this environment, static optimisation is increasingly insufficient. A system optimised for today's conditions may be strategically obsolete long before the procurement cycle finishes.

This suggests defence institutions may need continuous adversarial testing rather than periodic doctrinal review. Not static evaluation. Continuous strategic adaptation. The organisations capable of learning fastest under adversarial pressure are likely to dominate future conflict environments.

Much of the current public conversation around defence AI focuses on automation: Can AI replace pilots? Can AI automate targeting? Can AI reduce manpower requirements?

These are important questions, but they are not the deepest transformation underway.

The real transformation is machine-driven strategic adaptation — the ability to simulate adversarial co-evolution at scale, changing how institutions think, plan, procure, and learn.

This does not eliminate the need for human judgment. If anything, it increases it. AI systems can explore enormous strategic possibility spaces, but human decision-makers are still required to evaluate political consequences, escalation risks, alliance dynamics, operational feasibility, and ethical constraints.

The future is therefore unlikely to be fully autonomous warfare. It is more likely to be human-machine strategic collaboration: machines explore, humans interpret, doctrine evolves through interaction between the two.

Warfare is not fundamentally an optimisation problem.

It is an adaptive equilibrium problem.

That distinction changes everything.

The future of defence AI will not belong to the systems that optimise fastest.

It will belong to the systems that adapt fastest against intelligent opposition.

In war, optimisation without adaptation is fragility.