Australia's Voluntary AI Safety Standard — the 10 guardrails for organisations

What the Standard is

The Voluntary AI Safety Standard was published by the Department of Industry, Science and Resources on 5 September 2024. It is a voluntary, principles-based standard designed to assist organisations in deploying AI safely and responsibly.

This Standard operates alongside the Government’s proposals for mandatory guardrails for AI in high-risk settings. Consultation on these proposals closed in October 2024.

The Standard’s development considered international frameworks, including the ISO/IEC 42001 AI management system standard and the NIST AI Risk Management Framework.

The 10 guardrails — overview

The Australian Government has developed 10 guardrails to guide organisations in the responsible development and deployment of AI. These guardrails are voluntary and intended to promote safety and public trust in AI systems. They provide a framework for organisations to consider as they integrate AI into their operations.

The first five guardrails focus on foundational elements for responsible AI practices. These include establishing accountability processes, managing risks, protecting AI systems and data, testing and monitoring AI models, and ensuring human oversight. Each guardrail outlines specific actions organisations should undertake.

Further guardrails (G6-G10) address aspects such as transparency, fairness, and ongoing improvement. Collectively, the 10 guardrails aim to support organisations in navigating the complexities of AI and fostering a safe and beneficial AI ecosystem in Australia.

Guardrails 6 to 10

Guardrail 6 requires organisations to inform end-users. This includes providing information about decisions made using AI, interactions with AI systems, and content generated by AI.

Guardrail 7 mandates the establishment of processes. These processes must allow individuals affected by AI systems to challenge the use of those systems or the outcomes they produce. Guardrail 8 focuses on transparency within the AI supply chain, requiring organisations to share information about data, models, and systems with other relevant organisations.

Finally, Guardrail 9 requires organisations to maintain records. These records must be accessible to third parties for the purpose of assessing compliance with the guardrails. Guardrail 10 emphasises stakeholder engagement, requiring organisations to evaluate stakeholder needs and circumstances, with a particular focus on safety, diversity, inclusion, and fairness.

How regulators treat the Standard

The Australia’s Voluntary AI Safety Standard is not directly enforceable. It operates as a set of guidelines for organisations using AI.

The Office of the Australian Information Commissioner (OAIC) and the Australian Competition and Consumer Commission (ACCC) have stated they will consider the Standard’s expectations when evaluating AI-related conduct. This assessment will occur within the context of the Privacy Act and Australian Consumer Law respectively.

Future regulation will include mandatory guardrails for high-risk AI use cases. Organisations demonstrating commitment to the Voluntary AI Safety Standard, including mapping controls against the 10 guardrails and retaining supporting evidence, may be viewed favourably.