Every general counsel who has faced a board question about AI "compliance" knows the feeling. A privacy matter maps to the Privacy Act. A consumer claim maps to the Australian Consumer Law. The AI question maps to everything and nothing. It sits across a voluntary standard, proposed mandatory guardrails, sector-specific regulator guidance, and state-level frameworks, none of which were designed with the others in mind.
We think the government's approach to AI regulation is sound policy. Begin with voluntary principles, build institutional capacity, then move to mandatory obligations for the highest-risk applications. That sequencing lets industry develop capability before regulation bites and lets regulators learn from voluntary adoption before they mandate. It is the kind of measured, iterative approach you want from policymakers dealing with technology they do not fully understand.
The problem is that clients deploying AI systems today need advice today. And the single question that determines what that advice should be, whether their use case counts as "high-risk," is the question the framework deliberately postpones. Sound policy has produced an impossible advisory problem, and we are living inside it.
Australia's AI Ethics Framework, published in 2019, laid out eight principles: fairness, reliability, transparency, accountability, contestability. Aspirational, directional, no obligations attached. The Voluntary AI Safety Standard followed in September 2024: ten guardrails for organisations developing, deploying, or using AI, covering transparency, human oversight, technical testing, risk management, and accountability across the AI lifecycle. The third step, the one practitioners are watching most closely, is the proposed mandatory guardrails for high-risk AI systems following the 2024 proposals paper. Those guardrails would take the voluntary standard into mandatory territory, placing obligations on developers and deployers of systems classified as high-risk.
Consider a client building an AI hiring tool. They need a compliance view today. The voluntary standard's human-oversight guardrail is the first thing that bites: who reviews the model's screening decisions, how, and with what authority to override? The standard is voluntary, but a tool that screens candidates for employment is making decisions about people, and treating human oversight as optional is a position that will not age well.
Then the question underneath everything: will this system be classified as "high-risk" when the mandatory regime arrives? That definition determines whether your client falls within the mandatory regime. Until it is settled, you are advising in the dark on the most consequential classification in the framework.
The framework's answer to the client is: adopt the voluntary standard now, and you will be better positioned when the mandatory regime arrives. That is true. It is also the framework telling the client to build to a standard that may or may not match the final mandatory requirements, because the definition of "high-risk" is still being worked out.
The new framework is only part of the picture. Existing laws apply to AI systems right now, and a practitioner who focuses only on what is coming risks missing what is already in force.
The moment your client's hiring tool ingests CV data, the Privacy Act 1988 (Cth) is in the room. The model is collecting, using, and processing personal information, and the Australian Privacy Principles apply to how that data is collected, used, and disclosed, including for training. Recent amendments have strengthened enforcement and accountability mechanisms under the Act. The voluntary standard tells your client to implement human oversight; the Privacy Act tells them they also need consent or another lawful basis for the data handling that makes the tool work at all.
The next beat is the outcome. When the model screens out candidates in a pattern that tracks a protected attribute, federal and state discrimination legislation engages. The technology does not exempt the user from existing obligations. The legal question is whether the outcome constitutes unlawful discrimination, regardless of whether the algorithm intended it. A hiring tool that systematically disadvantages older applicants, or candidates with particular disabilities, is producing the same harm the discrimination statutes were written to address.
The Australian Consumer Law catches the vendor's representations. A company that markets its AI hiring tool as bias-free, or claims it can identify the best candidates, faces ACCC enforcement if those claims overstate what the system can do. Unfair contract terms provisions can bite on terms embedded in the AI service agreement itself.
And if the client deploying the tool is a regulated entity or a government contractor, the layer thickens further. Financial services licensees using AI remain subject to existing obligations under the Corporations Act 2001 (Cth). A state government agency deploying the tool carries Victoria's public-sector AI governance arrangements, and other states are watching. For clients with public-sector contracts or cross-border operations, the federal instruments are one layer and state frameworks are another.
The gap between voluntary and mandatory is real. The difficulty is the pace at which it is closing and the uncertainty about what will fill it.
A client who adopts the Voluntary AI Safety Standard today is building the systems and processes that mandatory guardrails will likely require of high-risk AI tomorrow. The client who ignores the voluntary standard because it is voluntary may find themselves unprepared when the mandatory regime arrives. Practitioners need to make this case to clients. The voluntary standard is a preview of the mandatory regime. Adopting it now is risk management: building the internal capacity, documentation practices, and oversight mechanisms that legislation will require.
Contract drafting is the other frontier. Agreements involving AI systems need risk allocation that accounts for the regulatory trajectory. A model that produces a discriminatory outcome leaves responsibility between developer and deployer uncertain. A system reclassified as high-risk after deployment means the compliance cost falls wherever the contract allocates it, and the contract was almost certainly drafted before anyone knew the reclassification was coming. Resolving these issues requires a practitioner who can see where the law is heading, and the honest answer is that no one can see it clearly yet, because the definition that would make it clear has not been settled.
To advise on AI compliance, a practitioner needs to hold the full picture at once. The 2019 Ethics Framework, the Voluntary AI Safety Standard, the proposed mandatory guardrails, the Privacy Act and its recent amendments, the Australian Consumer Law, discrimination law, sector-specific guidance from regulators, and state-level frameworks. That is a mapping exercise across the whole of the regulatory terrain, and no single instrument answers it.
The risk of incomplete advice is highest where the framework is newest. Memory might recall the voluntary standard but miss the recent Privacy Act amendments. General-purpose AI introduces its own failure mode: conflating voluntary guardrails with mandatory obligations, or citing the EU AI Act as if it applied in Australia.
A practitioner advising on an AI hiring tool needs to know what the voluntary guardrails say about human oversight, what the Privacy Act requires for personal information used in training, what discrimination law covers in employment decisions, and what sector regulators have said. They need all of it, traced to Australian source, in the form of a working analysis they can advise from.
A general counsel who types that question into Habeas gets back a synthesis drawn from the Australian primary-law corpus: the relevant Privacy Act provisions, the applicable discrimination law, the voluntary guardrails, and any sector-specific guidance, each tied to its source. A general-purpose AI tool asked the same question might retrieve the EU AI Act or a US state law. Habeas returns the Australian instruments, because its corpus is Australian primary law.
The judgment remains the practitioner's. Making the call on what "high-risk" means for a client's specific use case, drafting the contract clauses, advising the board: that work stays. To test Habeas against this regulatory patchwork, book a demo at habeas.ai.
If you want to try for yourself or get in contact, book a demo with us here. We also offer the capacity for self-serve individuals to sign up, and subscribe or register a free trial at app.habeas.ai.
The legal research in this article was conducted and every citation verified using Habeas, the Australian legal AI research platform.
Hero image: Vanessa Gallagher on Pexels
