Russell Buzby

On 27 July the agency that writes Australia’s whole-of-government digital policy explained to a Senate committee how it had been getting that policy to the people meant to follow it. Staff had been guessing email addresses. When a guess was wrong and the message bounced, nobody was reading the alerts that said so. One senator called the Digital Transformation Agency’s handling of IT policy dangerously incompetent (The Mandarin, 2026a).

Policy distribution is dull work, which is part of why the incident deserves more attention than a news cycle gave it. The DTA sits at the centre of Commonwealth digital administration. It negotiates the software arrangements the public service runs on, maintains the AI transparency register agencies publish to, and issues standards agencies are told to meet. An organisation that cannot reliably reach its own audience is poorly placed to assure anything about how that audience uses artificial intelligence.

The following day produced a second finding. InnovationAus reported that politicians and their staff fall outside the government’s central AI governance framework (InnovationAus, 2026a). The people legislating mandatory standards for everyone else are not covered by the rules they administer to the public service. Ministerial offices draft policy and work constituent casework at volume, which makes them the kind of decision environment the framework was written for.

The same day, The Mandarin published a warning from public sector leaders that AI ambition is running ahead of organisational readiness, and that rapid deployment amplifies failure where people and culture have not been brought along (The Mandarin, 2026b). All three items arrived within a fortnight of the Prime Minister’s 15 July keynote, which committed Australia to Australian Standards for AI applying across the economy with mandatory obligations for the first time, legislation promised by early 2027, and an Office of AI stood up inside the Department of the Prime Minister and Cabinet (Gilbert + Tobin, 2026).

The direction of that policy deserves credit. Australia spent three years on voluntary frameworks and self-attestation, and the record of what that produced is documented well enough now to close the argument, which is the argument of The Transparency Illusion. Mandatory standards are the right correction. The difficulty is that this correction assumes an administrative capability the same fortnight put in doubt, and close observers have read the wider policy package (The Conversation, 2026) as mostly older initiatives brought forward after long periods of neglect.

Research published on 30 July puts a technical floor under the doubt. A study of general-purpose systems in policing argues that public-service AI governance frameworks were built for narrow tools with a single purpose and cannot simply be carried across to large language models (Why Public Service AI Governance Frameworks Risk Failing in the Age of General-Purpose AI, 2026). Accuracy and bias were tractable properties, and explainability was at least approachable, when a system did one job on one dataset. The paper’s more uncomfortable finding concerns the standard mitigations: expert evaluation and human-in-the-loop oversight both presuppose measurability and bounded outputs, and general-purpose systems remove both. Full compliance with such a framework can therefore assure very little about the system it was applied to.

State politics is moving the other way at speed. The NSW Opposition has proposed mandating AI productivity tools across state government from the top down, backed by agency-level investment powers and annual public service productivity targets (InnovationAus, 2026b). Set aside whether the productivity case holds. A target-driven adoption mandate is a governance instrument with the governance taken out. It creates an obligation to deploy without a matching obligation to show the deployment works, and productivity targets attached to tool usage tend to be met by usage.

What connects the federal items is a single failure mode, and it has nothing to do with technology. Each is a gap between what government asks of others and what it demonstrates itself. The politician exemption is the clearest of the three, because exemptions communicate. When the rule-maker sits outside the rule, everyone reading the rule learns what kind of instrument it is: something government does to other people. That reading will follow the mandatory standards into every boardroom and agency executive that has to implement them, and it will be difficult to argue with.

The consequences run past the civilian public service. Australia funds artificial intelligence across nearly every line of the defence budget while governing it through a single policy, a mismatch I traced in Funded Across the Budget, Governed by One Policy. Defence assurance does not sit apart from the Commonwealth’s: it draws on the same procurement panels and the same shallow pool of people who can genuinely evaluate a model. When ASPI observed in mid-July that Australia’s plan for sovereign AI capability remains poorly defined even as the Prime Minister elevates AI to a national priority (ASPI, 2026), the gap it described was this one.

The remedies are unglamorous and mostly not about AI at all. Apply the framework to ministerial offices on the same terms agencies already carry. Fund the DTA to do policy distribution properly and publish whether agencies actually received what was sent, which is a two-line metric and a genuine test of whether the centre functions. Draft the mandatory standards against a general-purpose system before legislating them, because a standard written for narrow tools will wave through a chatbot that ought to fail it. Not one of those steps requires the national cabinet process the Prime Minister has set in train.

There is a version of the next eighteen months where Australia legislates a serious standard and cannot administer it, and that version looks a great deal like the last three years with penalties attached. Avoiding it means treating internal administration as part of the governance system instead of the overhead around it. Administration is also where the real constraint sits, and the fortnight just past was an unusually honest account of how much of it Australia has.

Australia is about to write a mandatory AI standard for the whole economy. On the current settings it will not cover the people writing it, and that is the detail every reader will notice first.

References

ASPI (2026, 17 July). As Australia prepares for the AI future, its plan for sovereign capability is unclear. The Strategist. https://www.aspistrategist.org.au/as-australia-prepares-for-the-ai-future-its-plan-for-sovereign-capability-is-unclear/

Gilbert + Tobin (2026). Australian Government announces mandatory AI standards for large-scale data centres and new Office of AI. https://www.gtlaw.com.au/insights/australian-government-announces-mandatory-ai-standards-for-large-scale-data-centres-and-new-office-of-ai

InnovationAus (2026a, 28 July). Politicians skirt bounds of government AI policy. https://innovationaus.com/pollies-skirt-bounds-of-government-ai-policy

InnovationAus (2026b, 27 July). NSW Coalition pledges to protect jobs in AI productivity push. https://innovationaus.com/nsw-coalition-pledges-to-protect-jobs-in-ai-productivity-push

The Conversation (2026, 22 July). Australia just outlined more ‘AI safety’ priorities, but is the plan actually coherent? https://theconversation.com/australia-just-outlined-more-ai-safety-priorities-but-is-the-plan-actually-coherent-287882

The Mandarin (2026a, 27 July). DTA’s IT policy management ‘dangerously incompetent’, claims senator. https://themandarin.com.au/317040-dtas-it-policy-management-dangerously-incompetent-claims-senator

The Mandarin (2026b, 28 July). Why AI ambition is outpacing readiness, and what public sector leaders can do about it. https://themandarin.com.au/316986-why-ai-ambition-is-outpacing-readiness-and-what-public-sector-leaders-can-do-about-it

The New Daily (2026, 15 July). PM outlines standards for AI, data centres in major speech. https://www.thenewdaily.com.au/life/tech/2026/07/15/pm-standards-ai-data-centres

Why Public Service AI Governance Frameworks Risk Failing in the Age of General-Purpose AI: Lessons from Policing (2026). arXiv:2607.25648. https://arxiv.org/abs/2607.25648