3The Australian Government AI mandate landscape

An agency cannot govern AI well without a clear view of what it is obliged to do. The difficulty is that the obligations are spread across more than a dozen instruments, owned by different parts of the Commonwealth, and several of the most important were issued or rewritten in 2025. This section provides a single, current map. It is organised into four families, with a cross-cutting base of general public-law and delivery obligations, and it closes by explaining how this framework turns many instruments into one operating response.

Ethical bedrock

• Australia’s AI Ethics Principles (eight)

AI policy and assurance

• Policy for the responsible use of AI in government (v2.0)
• AI technical standard and its agentic addendum
• National framework for the assurance of AI in government
• Guidance for AI Adoption (AI6)
• National AI Plan; APS AI Plan

Security and sovereignty

• Protective Security Policy Framework
• Information Security Manual (AI controls) and Essential Eight
• IRAP; Hosting Certification Framework
• ASD AI security guidance

Privacy and records

• Privacy Act and the APPs (ADM transparency from December 2026)
• OAIC guidance on AI
• Archives Act

Cross-cutting base

PGPA Act duties; Commonwealth Procurement Rules; administrative law and the Automated Decision Making Better Practice Guide; scrutiny by the ANAO, the Ombudsman and the Parliament.

3.1The ethical bedrock: Australia’s AI Ethics Principles

Beneath every Commonwealth AI instrument sit Australia’s eight AI Ethics Principles, maintained by the Department of Industry, Science and Resources. They are the shared language of assurance, and every other instrument maps back to them. They are: human, societal and environmental wellbeing; human-centred values; fairness; privacy protection and security; reliability and safety; transparency and explainability; contestability; and accountability. The principles are not a compliance checklist but a set of outcomes to design toward. The assurance framework in Section 9 shows how to evidence each one.

3.2The Policy for the responsible use of AI in government (v2.0)

The DTA’s Policy for the responsible use of AI in government is the binding instrument for non-corporate Commonwealth entities. Version 1.1 took effect on 1 September 2024; version 2.0, effective 15 December 2025, substantially strengthened it. The policy now turns on a small number of mandatory obligations that this framework is built to discharge.

The policy is supported by an AI Impact Assessment tool and new guidance on AI procurement in government, both released alongside version 2.0. The framework adopts these directly rather than reinventing them: the assurance lifecycle in Section 10 is built around the impact assessment, and Section 13 shows how to apply the procurement guidance in day-to-day buying. Note the phasing: version 2.0 took effect on 15 December 2025; the strategic position on AI adoption, communicated to staff, was due by 15 June 2026; use-case owners, the register, responsible AI practices (including AI incident processes), staff training and the screening of new use cases are due by 15 December 2026; and existing use cases not yet assessed must be brought into line by 30 April 2027. An agency adopting this framework meets each deadline as a by-product.

3.3The AI technical standard

Published by the DTA on 31 July 2025, the AI technical standard translates the policy and the ethics principles into concrete practices across an AI system’s life. It structures that life into three phases, Discover, Operate and Retire, and marks the criteria within its 42 statements as either Required, where they are driven by legislation, or Recommended, where they are good practice. Policy v2.0 strongly recommends the standard rather than mandating it. The standard is the natural backbone for an agency’s assurance lifecycle, and Section 10 maps this framework’s gates directly onto its three phases. Treating the Required items as non-negotiable and the Recommended items as the default, departing only with a documented reason, is a simple and defensible way to operationalise it. In June 2026 the DTA extended the standard with an agentic AI addendum for higher-autonomy systems under human oversight; its concerns, bounded autonomy, runtime control and human override, are exactly the disciplines Sections 8 and 9 of this framework build.

3.4The National framework for the assurance of AI in government

Agreed by the Data and Digital Ministers Meeting in June 2024, the National framework for the assurance of AI in government gives all Australian governments a common approach to assuring AI against the eight ethics principles. It defines assurance as evidence-based confidence that an AI system is safe, lawful and fit for purpose, and it takes a risk-based approach, with heavier requirements for higher-risk uses. It sets five assurance cornerstones, which this framework adopts as the foundations of its governance force.

3.5The National AI Centre’s Guidance for AI Adoption (AI6)

In October 2025 the National AI Centre published its Guidance for AI Adoption, the first update to the 2024 Voluntary AI Safety Standard, which condenses the earlier ten guardrails into six essential practices (this framework calls them AI6 for short), and comes in two parts: Foundations, for organisations beginning their AI journey, and Implementation Practices, for technical and governance professionals. Although issued as guidance for organisations generally rather than as a binding government policy, the six practices are an excellent, plain-language operating checklist, and they run as a thread through the assurance lifecycle in Section 10. They are: decide who is accountable; understand impacts and plan accordingly; measure and manage risks; share essential information; test and monitor; and maintain human control.

3.6The National AI Plan and the Australian AI Safety Institute

On 2 December 2025 the Government released the National AI Plan, built on three pillars: capture the opportunity, spread the benefits, and keep Australians safe. The plan builds regulation on existing laws and sector regulators, with targeted new laws only where needed, rather than a single AI Act. It is supported by voluntary guidance and a new Australian AI Safety Institute, announced in November 2025 and operational from early 2026, which tests and monitors advanced AI, coordinates with regulators such as the Office of the Australian Information Commissioner, and connects Australia to the international network of AI safety institutes. For the public service specifically, the companion AI Plan for the Australian Public Service, released in November 2025 around three pillars of trust, people and tools, drives adoption from the centre: a Chief AI Officer in every agency by July 2026, mandatory foundational AI training for staff, and GovAI, a Finance-operated, government-hosted AI platform whose assistant, GovAI Chat, is being rolled out across the APS through 2026. The practical implication for agencies is important: there is no single AI law to comply with, so the burden falls on each agency to assemble its own coherent governance from the instruments above. That is precisely what this framework does.

3.7The security, privacy and records spine

AI does not displace the Commonwealth’s existing obligations; it intensifies them. Four bodies of rule matter especially for any AI that touches government data or runs on government systems.

The spine has grown AI-specific teeth. The Information Security Manual has carried controls for AI application development since June 2024. It expanded them substantially in December 2025, including a requirement for an AI usage policy; again in June 2026, covering AI application hardening, AI-assisted detection and testing, and the deletion of chat data; and in September 2026 it added an AI agent register, a unique identity for each agent, and controls on agent permissions. ASD has published a steady run of AI security guidance that an agency’s security team should have at hand: Engaging with Artificial Intelligence (2024), Deploying AI Systems Securely (2024), AI Data Security (2025) and, most relevant to this framework, Careful Adoption of Agentic AI Services (May 2026, written with international partners). The OAIC’s twin 2024 guidance notes, on privacy and commercially available AI products and on developing and training generative AI models, do the same for privacy. One transition belongs on the forward radar: in June 2026 ASD consulted on a proposed evolution of the Essential Eight into an Essentials series, beginning with Essentials for enterprise IT. The Essential Eight remains in force; agencies should track the change, not wait for it.

Two timing points deserve emphasis. First, the Privacy and Other Legislation Amendment Act 2024 introduces a new transparency obligation for automated decision-making: from 10 December 2026, APP entities must explain in their privacy policies the automated decisions, using their personal information, that significantly affect people’s rights or interests. Any AI that contributes to such decisions must be designed for that transparency now, not retrofitted later. Second, under the Archives Act, the records an AI system creates, including the decisions it makes and the basis for them, are Commonwealth records that must be kept; provenance is not optional. Both points shape the design and governance forces later in this framework.

3.8The scrutiny landscape, and the reform horizon

Obligations are one half of the mandate; scrutiny is the other, and it is no longer hypothetical. The Australian National Audit Office has moved from warning about AI to auditing it: its audit of the ATO’s governance of artificial intelligence (Report No. 26 of 2024-25) found the arrangements only partly effective and set out, in its criteria, what auditors now expect any agency to show, an AI strategy with implementation arrangements, enterprise-wide accountabilities, AI-specific policies and risk management, ethics assessment of models, controlled deployment, and structured monitoring of what is in production. Its audit of AI use at IP Australia (Report No. 43 of 2025-26) found a largely effective agency and still recommended stronger AI cyber-security governance and risk-based monitoring. The pattern is set: AI governance is now a standing subject of performance audit, and the difference between the two findings is essentially whether a framework like this one is in place and operating.

The Parliament and the integrity bodies have joined in. The Joint Committee of Public Accounts and Audit’s report on public-sector AI, pointedly titled Proceed with Caution (February 2025), called for robust governance frameworks in every agency. The Senate Select Committee on Adopting Artificial Intelligence reported in late 2024, and the Government’s April 2026 response leant on the National AI Plan settings and the Privacy Act’s automated-decision-making transparency reforms. The Commonwealth Ombudsman, the OAIC and the Attorney-General’s Department jointly reissued the Automated Decision Making Better Practice Guide in March 2025, the plainest available statement of what lawful, fair automated government decision-making requires: clear legal authority, transparency to the public, statements of reasons, accessible review paths and ongoing assurance. And the Attorney-General’s Department is developing the whole-of-government legal framework for automated decision-making that the Robodebt Royal Commission recommended; no legislation has been introduced as at July 2026, but its direction is visible in the better-practice guide. An agency that builds to this framework now will meet that legislation as confirmation, not as remediation. Two narrower 2026 instruments complete the picture: the APSC’s principles for the use of AI in recruitment, which agencies were to implement by 1 June 2026, which keep selection decisions with humans, and the OAIC’s resources on the December 2026 transparency obligation, published in September 2026.

3.9From many instruments to one framework

Read individually, these instruments can feel like a thicket. Read together, they describe a single, coherent expectation: that an agency will know where it is using AI, will have decided who is accountable, will assess and manage the risk of each use, will build and run AI on a secure and sovereign foundation, will be transparent and keep people able to contest decisions, and will assure all of this over the life of each system. Those are not a dozen separate programs. They are five domains of one system. The next section introduces the model that binds them.