In March this year the Digital Transformation Agency announced a milestone that deserves more attention than it received: every one of the 94 agencies subject to the Australian Government’s AI transparency standard had published its transparency statement, with another 20 publishing voluntarily, all discoverable through a new central register (Digital Transformation Agency, 2026). The contrast with the recent past was sharp. In February, reporting had described researchers hunting for statements that were missing, buried or vague, and finding some 30 entities with nothing findable at all (Johnson, 2026). Within weeks, the findability problem was fixed. On the compliance ledger, Australian government AI transparency is now a solved problem.

A study released in June explains why that sentence should make nobody comfortable. Researchers analysed 92 of those Australian Government AI transparency statements against the needs of different stakeholder groups and found a systematic skew (AI transparency, 2026). Agencies broadly satisfy the disclosure mandate. The criteria that matter most to high-risk, low-control stakeholders, the people most exposed to government AI decisions with the least power over them, are the criteria least often and least substantively addressed. The authors give the pattern a name, the Transparency Illusion: “a condition in which transparency appears satisfied through compliant artefacts yet remains unevenly calibrated to stakeholders bearing the greatest exposure.”

It helps to make the stakeholder abstraction concrete. A departmental compliance officer reading a transparency statement wants to know that a policy exists and an accountable owner is named, and the statements serve that reader well. A welfare recipient whose payment was flagged by an automated system, or a visa applicant whose case was triaged by a model, wants to know something different: what the system actually does, what data it saw, how often it is wrong, and what to do when it is. Those are the disclosures the study found thin. The people with the least ability to walk away get the least usable information, and they cannot walk away, which is the point. A companion study of what its authors call the opt-out paradox found that in compulsory, AI-enabled government services, citizens who distrust the system do not exit, because they cannot. Their distrust shows up instead as guarded, hedged, less authentic engagement, a degradation in the quality of the relationship that no adoption metric will ever register (Compelled but not cooperative, 2026).

None of this is an argument that the policy framework is misconceived. The DTA’s updated Policy for the Responsible Use of AI in Government took effect in December 2025 and its requirements are sensible (Digital Transformation Agency, 2025). Agencies owed a strategic position on AI adoption by mid-June this year. Within twelve months they owe an internal register of in-scope AI use cases shared with the DTA, an accountable owner for each, and responsible-AI training for staff. Australia’s architecture is considered, and it compares well internationally. A preprint analysis circulating in July describes the United Arab Emirates deploying AI that makes autonomous work permit decisions for every applicant, with no published audit or bias-testing safeguards identified and no stated appeal path (The UAE created a Federal AI Authority, 2026). Whatever Australia’s gaps, they are gaps inside a real accountability structure. That is worth acknowledging before the critique, because the critique is about where real accountability structures go to die: the space between what a mandate requires and what its purpose was.

The international research suggests the drift is structural rather than an Australian quirk. A study of a Swedish government agency found public sector AI communications reproducing the private sector’s “machinewashing” playbook, framing AI through narratives of collective benefit while the political choices and trade-offs inside the systems stay undescribed (Public sector machinewashing, 2026). The vagueness has deep roots. An analysis of 91 highly cited public administration AI papers found 55 per cent never adequately specify the AI system they discuss, and 41 per cent draw conclusions broader than the system studied supports (A technical typology, 2026). When the research literature itself blurs what an AI system is, agencies writing statements from that literature inherit the blur. One corrective on offer is vocabulary: a framework proposed in July distinguishes six functions of AI-enabled digital government: information; service delivery; decisions; oversight and audit; participation; and policy support (Concepts of AI-enabled digital government, 2026). A statement forced to declare which of those a system performs would already be harder to write vaguely.

I have written separately about the readiness gap in public sector AI governance and about the Chief AI Officer mandate that created accountable owners faster than it created their capability, and this is the same story at the disclosure layer. I write separately about a Canadian study that found AI consultation summaries silencing dissenting submissions while passing every standard quality check. The rhyme between the two findings is exact. In both cases the artefact satisfies its formal test (an accurate summary, a published statement) while the failure sits with the people the artefact was nominally for. Compliance is measured at the centre; the harm accrues at the edge.

The practical shift is to treat transparency as a design property of the system rather than a reporting obligation about the system. The test of a statement is behavioural: could the person affected by this system, reading this document, understand what happened to them and mount a contest? Could a journalist reconstruct the decision chain? Could an auditor find the failure modes without a briefing? Statements written to that test look different. They name error rates, review pathways and the boundary where the machine stops and a human answers. There is even a hard institutional incentive buried in the research: a six-country survey of 1,247 adults found that disclosure practice is the strongest predictor of public trust in AI-mediated digital media (Ethical AI communication and public trust, 2026). Substance, disclosed plainly, is not a compliance cost. It is the trust budget.

Australia has done the hard first mile. The register exists, the statements exist, the deadlines are being met, and agencies deserve the credit for that, because plenty of governments have neither. The remaining mile is the one the illusion hides. Ninety-four published statements answer the auditor’s question in full. The citizen’s question (what is this system doing to me, and what can I do about it?) is still mostly unanswered, and no volume of compliant artefacts will answer it by accumulation. Transparency earns its name at the moment the person with the most to lose can use it. Until then it is documentation, and documentation was never the point.

References