Australia's Voluntary AI Safety Standard, Plainly

The ten guardrails, compressed

Stripped of policy language, they cluster into four things: know who is accountable, know what could go wrong, be able to show your work, and tell people when AI is involved.

• Accountability - a named owner, an internal policy, and someone whose job it is. Not a committee.
• Risk management - identify and mitigate the specific harms of the specific system, documented, and revisited rather than done once.
• Data governance - know what data trains and feeds the system, its quality, its provenance, and its legal basis.
• Testing and monitoring - evaluate before deployment against defined criteria, then keep monitoring, because models drift and inputs change.
• Human control - a person can intervene, override, or stop the system. This has to be real, not a button nobody has authority to press.
• Informing end users - people are told when AI is materially involved in something that affects them.
• Contestability - a person affected by an AI-influenced outcome has a route to challenge it, and that route reaches a human.
• Supply chain transparency - you know what is in the AI you bought, and you can tell your own customers.
• Record keeping - documentation sufficient for someone else to assess compliance later.
• Stakeholder engagement - consider the people affected, particularly on safety, diversity and fairness.

Which ones cost real engineering effort

Most of the list is organisational and costs meeting time. Four of them are genuine engineering work, and they are the four that are painful to add late.

• Testing and monitoring - needs an evaluation set, a defined pass criterion, and drift monitoring on inputs and outputs. If you have never built this, it is a real project, not a checkbox.
• Human control - an override path has to exist in the data model, not just in the interface. A system that cannot record 'a human overrode this, and why' cannot demonstrate human control.
• Contestability - requires that you can reconstruct why a specific decision came out the way it did, for a specific person, months later. That is a logging and explainability requirement.
• Record keeping - the documentation only exists if it is generated as you go. Nobody successfully writes it retrospectively.

The model choice this pushes you towards

Contestability and human control are much cheaper to satisfy with a model that can explain a specific decision. Gradient boosting on your own tabular data returns per-feature attribution for every prediction, so 'why was I declined' has a real answer. A language model asked to produce the same score gives you a plausible sentence that may not describe the actual computation, which is worse than no explanation because it looks like one.

What is proposed as mandatory, and where

The mandatory proposals target high-risk settings rather than AI generally. The consistent thread in how high risk gets defined internationally is the severity of the impact on a person: decisions about employment, credit, essential services, health, education, and law enforcement tend to land inside. A chatbot that answers questions about your opening hours does not.

A proportionate response for a small business

The failure mode here is a small company building an AI governance programme it does not need, and the opposite failure mode is a small company automating a hiring decision with no logging at all. Both are avoidable.

• Write down what each AI system does, who owns it, and what happens if it is wrong. One page per system is enough to start.
• Decide explicitly whether any of them influence a consequential decision about a person. That answer sets the effort level for everything else.
• For anything that does: log the inputs, the output, and any human override, and make sure you could explain one specific decision a year from now.
• Tell users when AI is materially involved. This is cheap, it is increasingly expected, and it is the thing customers notice you hiding.
• Do not write a policy you will not follow. An unfollowed policy is worse evidence than no policy.