Russell Buzby

There is now an artificial intelligence component in almost every part of Australia’s defence build. The 2026 National Defence Strategy names AI as a force multiplier and threads it through the capability program from undersea warfare to cyber. The Integrated Investment Plan beneath it commits in the order of A$425 billion across the decade to 2035-36. AI sits inside the targeting and intelligence systems, and in the decision support that feeds the command chain. It has been integrated across nearly every line of the budget. The governance has not spread anywhere near as widely.

The governance does exist. In March 2026 Defence released its Policy Settings for Responsible Use of Artificial Intelligence in Defence, a dedicated framework built on three non-negotiable requirements. Any AI the organisation uses must comply with Australian and international law. A human must remain accountable for it, and that accountability has to be made real through explainability and bias mitigation. And the risk it carries must be managed in proportion, through evaluation suited to the stakes. On paper the framework is sound, and Defence deserves credit for writing a defence-specific policy rather than leaning on whole-of-government rules built for a different risk profile. The difficulty is one of scale. A single document, however well drafted, now has to govern AI built into dozens of programs running at very different tempos and maturity levels.

The Australian Strategic Policy Institute put the problem sharply in late June, arguing that the ADF is integrating AI faster than it can govern it and risks fielding capabilities that outrun accountability and strategic coherence (ASPI Strategist, 2026a). Its proposed correctives are worth dwelling on, because they are concrete rather than aspirational. The first is to build thresholds that deliberately slow AI-assisted decision cycles at the points where escalation is possible, so a human has time to intervene before a machine-speed process commits to something irreversible. Auditable decision support for high-stakes outputs comes next, so that when an AI-shaped recommendation is acted on, there is a record of how it was reached. The third is a clear institutional position on where AI does not belong in Australian defence decision-making at all.

What makes a gap like this so hard to close is the asymmetry between how fast the dependency builds and how slowly it can be unwound. A new measure called AI Reliance, designed to quantify how deep and how reversible an institution’s AI dependency has become, found that United States federal AI use cases rose roughly fivefold in two years while reversibility in the public sector ran far slower and costlier than in private markets (AI Reliance, 2026). Defence sits at the extreme end of that curve. It adopts at the speed of a capability program and would have to disentangle at the speed of a Senate estimates cycle. I have written separately about how the same reversibility problem shows up in Australia’s broader reliance on foreign frontier AI models, and the defence version is that dynamic with higher stakes.

The reason the gap tends to widen rather than close has been modelled directly. The Capability-Friction Dynamics work argues that short-term political pressure to deploy AI quickly crowds out the foundational governance investment that would make the deployment safe, so the rush to field capability actively erodes the capacity to field it responsibly (Capability-Friction Dynamics, 2026). In a defence context that pressure is acute, because every quarter a capability sits ungoverned is a quarter an adversary may be fielding its own. The temptation is to treat governance as the thing you attend to once the capability is in service. The model’s point is that the sequence is self-defeating, because the governance debt compounds while you wait.

The Defence policy asserts human accountability, which is the right principle and also the easiest of the three requirements to assert and the hardest to evidence. Asserting that a human is accountable for an AI-shaped decision means little without an inspectable record of what that human actually saw and could have changed. Recent work on what its authors call a Human in the Log proposes exactly that missing piece: a structured, seven-part oversight record that makes human control over a public-sector AI system inspectable and legally defensible rather than simply claimed (Human in the Log, 2026). That is the kind of implementation detail that turns a governance principle into something an auditor or a court could actually test.

The timing matters because the acquisition model itself is about to change. In late June the Defence Industry Minister, Pat Conroy, signalled that spiral development, the practice of fielding a capability in increments and improving it in service while reducing risk along the way, will become a more prominent part of Australian military procurement (ASPI Strategist, 2026b). For capability that is sensible, and it reflects hard lessons from programs that tried to specify everything up front and delivered late. For governance it raises the bar considerably. A framework that already struggles to keep pace with linear acquisition now has to govern a moving target, because a capability fielded in increments means the system in service next year is not the one that was assessed this year. Spiral development without spiral governance is just a faster way to accumulate ungoverned capability.

None of this is an argument that the answer is simply more policy. A provocative paper released in late June, the Governance Inversion Hypothesis, argues that expanding AI regulation can paradoxically reduce an organisation’s operational control through authority fragmentation and symbolic compliance, as the real expertise drifts out to the external vendors who end up holding it (Governance Inversion Hypothesis, 2026). The warning is well aimed at Defence, where it would be easy to answer the integration problem by generating more frameworks and more sign-off gates, and to mistake that activity for control. The measures that would actually help are narrower and more demanding: the decision-cycle thresholds and auditable records that ASPI described, each of them a capability to be built rather than a document to be published.

This is not the first time the gap between an AI policy and the practice it is meant to direct has shown up in Defence. I have written separately about the March 2026 military AI policy and the way it was drafted for AI as an advisory tool, a system that recommends while a human decides, even as the technology moves toward AI that controls an autonomous loop at a tempo no human can supervise in real time. The integration problem and the implementation problem are two views of one thing. A policy can be sound in principle and still lack the auditable records and the resourcing that would let it govern AI in the wild, across the breadth of programs the investment plan is funding.

A policy exists, and it is a good one. The open question is whether the governance can scale across every program that now contains AI, and whether it can iterate as fast as the capability it is meant to oversee. On current evidence the capability is moving at the speed of the technology and the oversight is moving at the speed of a committee, and the distance between those two speeds is where the risk lives. Closing it is the harder, less visible work: building the auditable records and the decision limits that turn a sound policy into something that holds when an AI system, somewhere in the budget, makes a consequential call faster than anyone can check it.

References

AI Reliance: a reproducible measure of how deep, and how reversible, systematic AI dependency is. (2026). Zenodo. https://doi.org/10.5281/zenodo.20763481

Australian Strategic Policy Institute. (2026a, June 24). The ADF is integrating AI faster than it can govern it. The Strategist.

Australian Strategic Policy Institute. (2026b, June 25). Australia will increasingly use spiral development in defence: Conroy. The Strategist.

The Capability-Friction Dynamics: A macro-architecture for public sector AI. (2026). AMCIS 2026 Proceedings.

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

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

The Governance Inversion Hypothesis: Why more AI regulation may produce less organisational control. (2026). arXiv:2606.26117.

Human in the Log: Public evidence chains for public-sector AI oversight. (2026). Zenodo. https://doi.org/10.5281/zenodo.20759553