Russell Buzby

A Ukrainian drone operator using a $5,000 commercial quadcopter neutralised a Russian motorised company on the Pokrovsk axis last month inside an afternoon. The unit cost ratio between the quadcopter and the company it engaged was around six orders of magnitude, and the decision loop from drone sensor to weapon release was measured in seconds. While that drone engagement was unfolding, the Australian Department of Defence was working through committee comments on its March 2026 “Policy Settings for Responsible Use of Artificial Intelligence in Defence” (Department of Defence, 2026). The Pokrovsk engagement and the Defence policy document, taken together, illustrate how widely the operational and governance worlds of military AI have now diverged.

The Defence policy, on its own terms, is a credible piece of work. It commits the ADF to four overarching requirements when AI is used in Defence settings: legal compliance with Australian and international law, individual accountability for decisions enabled by AI systems, explainability of how those systems reach their outputs, and a baseline of operational reliability. These principles are sound, and would benefit any equivalent civilian agency held to the same standard. The Defence AI Centre, established in 2024 as the institutional home for the framework, is the obvious place where its application across services and acquisition programs would be coordinated.

The Australian Strategic Policy Institute has now made the case that the framework, as written, cannot keep up with the kind of warfare it is meant to govern. The argument from the Strategist piece is that the policy’s requirements were drafted with AI as an advisory or decision-support tool in mind, not as the controller of an autonomous loop running at machine tempo (Australian Strategic Policy Institute, 2026). Modern electronic warfare and cognitive EW already operate at speeds where requiring an explainable and individually attributable decision becomes either a fiction or a procurement-stopper. The question raised by the ASPI piece is whether the policy will be implemented in a way that excludes Australia from the class of capabilities currently deciding outcomes overseas, or whether the implementation will carve out exceptions in practice that the document does not name in writing.

That reading carries more weight when paired with two empirical findings from the same fortnight. The ARMOR 2025 benchmark, published in early May, tested 21 commercial large language models against scenarios drawn from the Law of War and Rules of Engagement, and found none of them ready for legally consequential decisions in a military context without significant additional safeguards (ARMOR 2025 benchmark, 2026). The MEBL framework, working from a different angle, evaluated LLMs across a range of military-specific axes and concluded that model size on its own is a weak predictor of military utility (MEBL framework, 2026). Read together, those two findings close off the easy escape route. A defence procurement officer cannot simply buy the largest commercial LLM available and assume it will pass an explainability or reliability test in an operational setting. The off-the-shelf maturity that the policy framework appears to assume does not yet exist.

The implementation gap here is therefore not the standard public-sector pattern I have written about separately in the context of the earlier Australian military AI policy. It is more specific. The framework requires legal compliance and accountable individual decision-making. Current commercial AI maturity does not reliably produce those properties at the speed required for warfighting. Suppliers who could plausibly meet the requirements at warfighting speed are small in number, and the sovereign Australian capability to evaluate their claims is even smaller. On top of all that, the actual war the policy is presumably written to anticipate is being fought now by adversaries whose systems would not survive the policy’s filters in any honest reading.

A useful comparison point sits in Taipei. Taiwan’s All-Out Defence Mobilisation Agency was highlighted last week in Lowy Institute analysis for a structural reason that maps onto the Australian situation (Lowy Institute, 2026). The agency is small. Its institutional authority comes from its proximity to cabinet-level decision-making rather than from its budget or staff numbers. It is the body in the room when the relevant decisions are made. Australia’s Defence AI Centre, by contrast, sits some distance from the National Security Committee and from the operational chain that would have to execute against an AI-enabled adversary. If the location of the governance body matters more than its formal mandate, as the Taiwan example suggests, then Australia’s current arrangement is not where it needs to be.

When this argument is made, the easy translation is into a case for less governance. That translation misses the point. Responsible use in Defence is non-negotiable; getting it wrong has a permanent cost in both legal and human terms. What the current pressure exposes is that the framework was written at one tempo and the threat environment is operating at another, and the gap will be closed by one of two routes. Either the policy gets recalibrated for the speed and autonomy that modern EW demands, or the implementation will be characterised by exceptions and informal accommodations that hollow the framework out from inside without ever being acknowledged in writing. The second outcome is worse. It produces capability that operates outside its own governance regime, and a public account of that fact which no one can credibly defend under scrutiny.

The 2026 National Defence Strategy commits Australia to deterrence by denial and to accelerated acquisition of autonomous systems (Department of Defence, 2026). That commitment requires a governance framework that can keep pace with the procurement and deployment cycles the strategy is asking for. The current AI policy was drafted before the implications of the Ukrainian war for autonomy and tempo were fully visible. Those implications are now visible. A public, deliberate revision of the policy is the obvious next move, one that names the operational context honestly. AI in Defence is no longer purely an advisory category, and a framework written for that case will not govern the rest of the workload.

For consultants and policy professionals working in or around Defence, the practical observation is that this is the kind of gap that would normally close through exceptions and informal practice. Resisting that pattern, and pushing for a written reframing instead, is the higher-quality contribution to make. Frameworks that exist on paper but operate through informal exceptions are the structures that fail hardest under pressure, and the failure mode is not contained to the original framework: it spreads to any related governance system that has to interact with it.

A policy written for the wrong tempo will not be repaired by good intentions. Correction comes from rewriting it with the actual operational context in view. The longer that revision is delayed, the more of Australia’s AI-enabled defence capability will end up living outside the governance framework meant to discipline it. Responsible use that cannot keep up with the war it is meant to govern lapses, in operational settings, into the appearance of responsibility, and that appearance lasts only as long as the first operational decision that puts it under pressure.

References

Australian Strategic Policy Institute. (2026). Australian Defence AI policy risks writing modern EW out of the force. The Strategist.

ARMOR 2025 benchmark. (2026). Testing of 21 commercial LLMs against Law of War and Rules of Engagement scenarios. arXiv preprint.

Department of Defence. (2026, March). Policy Settings for Responsible Use of Artificial Intelligence in Defence. Commonwealth of Australia.

Department of Defence. (2026). 2026 National Defence Strategy and Integrated Investment Program. Commonwealth of Australia.

Lowy Institute. (2026). Taiwan’s mobilisation model holds lessons for Australia. The Interpreter.

MEBL framework. (2026). Military Evaluation Benchmark for Large Language Models. OpenAlex.