The Future of AI in Military Strategy: A New Era of Defense

Military organisations have reasons to examine AI: information is abundant, resources are constrained and decisions can be time-sensitive. Those pressures also make unearned confidence dangerous. A system that helps organise information may be useful without being qualified to decide what action should follow.

The strategic question is therefore more precise than whether AI will transform warfare. Which function does a proposed system serve, what evidence supports it, and who remains responsible when its recommendation is wrong? Those questions must survive the pressure to describe every technical advance as a decisive advantage.

Support functions and force decisions are different

Potential uses include logistics planning, training exercises and organising information for analysts. These possibilities should be assessed separately from systems involved in selecting or engaging targets. Evidence of utility in a support function is not evidence of safe performance in a force-related decision.

AI and autonomy also describe different properties. A system can automate a predefined task without relying on modern machine learning; an AI component can provide advice without having authority to act. Procurement discussions should describe the actual function and permissions rather than infer them from a label.

UNIDIR’s analysis of AI and international security discusses risks and confidence-building possibilities. It provides a basis for considering how technical development can affect strategic relationships, rather than a guarantee that a particular deployment reduces risk.¹

Test the decision, including its uncertainty

My recommendation is to evaluate a proposed system against the decision it is meant to support. What information can it miss? How does it behave when conditions differ from training or testing? Can a reviewer identify an uncertain inference? Does the interface make alternative explanations easy to examine?

A good average can conceal consequential failures. Planners should examine disagreement, ambiguous cases and the costs of false positives and false negatives. If a system is useful only under narrow conditions, those boundaries belong in the operational description.

Human involvement should also be described concretely. A person asked to approve a suggestion needs time, relevant information and authority to refuse. The presence of an approval button alone does not establish meaningful judgment. These are proposed evaluation criteria, not claims that any named force has already met them.

Speed can create pressure as well as advantage

Faster information processing does not automatically improve strategic outcomes. In a crisis, a confident but poorly supported recommendation could narrow the time available to question an interpretation. Institutions should investigate that possibility rather than assume that more speed always means more control.

I would include exercises in which information is incomplete, disputed or deliberately misleading. Observe whether participants recognise uncertainty and whether the workflow permits escalation. The purpose is to test judgment under pressure, not to claim that a simulation proves safety in every real confrontation.

Technology does not erase existing responsibilities

The ICRC’s discussion of proportionality in cyber operations explains that anticipated incidental civilian harm must be assessed against the concrete and direct military advantage expected in relevant attacks during armed conflict. Novel computing does not make such legal principles disappear.²

That source concerns a defined legal context. It should not be stretched into a claim that every intelligence application is a weapon or that all uses of AI share identical legal requirements. Appropriate legal review needs the facts of the proposed system and use.

Confidence without false reassurance

For cross-border discussion, I favour concrete descriptions of system roles, limits and incident concerns over sweeping claims of benign intent. Where appropriate, communication about testing and human authority may reduce ambiguity. Such proposals still require political agreement and verification; they are not substitutes for either.

Claims that AI will make nuclear arsenals obsolete or neutralise every threat before launch would require operational evidence. Without it, planners should concentrate on decisions that can actually be investigated, while treating more ambitious possibilities as scenarios rather than assurances.

AI’s contribution to military strategy should be earned through evidence and constrained by responsibility. The most valuable capability may sometimes be helping a person understand uncertainty, rather than allowing an institution to act faster than it can explain.

Sources and notes

1. Ioana Puscas, AI and International Security: Understanding the Risks and Paving the Path for Confidence-Building Measures, UNIDIR (October 2023). Research analysis and proposals, not a binding international agreement. ↩
2. ICRC, Cyber Operations During Armed Conflict: The Principle of Proportionality (March 2023). Legal guidance for the specified context, not certification of an AI system. ↩