Introduction to the standard

The Voluntary AI Safety Standard gives practical guidance to all Australian organisations on how to safely and responsibly use and innovate with artificial intelligence (AI). Through the Safe and Responsible AI agenda, the Australian Government is acting to ensure that the development and deployment of AI systems in Australia in legitimate but high-risk settings is safe and can be relied on, while ensuring the use of AI in low-risk settings can continue to flourish largely unimpeded.

In 2023, the government underwent consultation through its discussion paper on ‘Safe and Responsible AI in Australia’. In the Interim Response areas of government action were outlined, including: 

  • delivering regulatory clarity and certainty
  • supporting and promoting best practice for safety
  • ensuring government is an exemplar in the use of AI
  • engaging internationally on how to govern AI.

The response also recognised the need to consider building AI capability in Australia. 

To support and promote best practice, an immediate action was to work in close consultation with industry to develop a Voluntary AI Safety Standard. This standard complements the broader Safe and Responsible AI agenda, including developing options on mandatory guardrails for those developing and deploying AI in Australia in high-risk settings. 

While there are examples of good practice through Australia, approaches are inconsistent. This is causing confusion for organisations and making it difficult for them to understand what they need to do to develop and use AI in a safe and responsible way. The standard establishes a consistent practice for organisations. It also sets expectations for what future legislation may look like as the government considers its options on mandatory guardrails. 

The standard consists of 10 voluntary guardrails that apply to all organisations throughout the AI supply chain. They include testing, transparency and accountability requirements across the supply chain. They also explain what developers and deployers of AI systems must do to comply with the guardrails. The guardrails help organisations to benefit from AI while mitigating and managing the risks that AI may pose to organisations, people and groups.

The 10 guardrails

  1. Establish, implement, and publish an accountability process including governance, internal capability and a strategy for regulatory compliance. 
  2. Establish and implement a risk management process to identify and mitigate risks. 
  3. Protect AI systems, and implement data governance measures to manage data quality and provenance.
  4. Test AI models and systems to evaluate model performance and monitor the system once deployed.
  5. Enable human control or intervention in an AI system to achieve meaningful human oversight.
  6. Inform end-users regarding AI-enabled decisions, interactions with AI and AI-generated content.
  7. Establish processes for people impacted by AI systems to challenge use or outcomes.
  8. Be transparent with other organisations across the AI supply chain about data, models and systems to help them effectively address risks.
  9. Keep and maintain records to allow third parties to assess compliance with guardrails. 
  10. Engage your stakeholders and evaluate their needs and circumstances, with a focus on safety, diversity, inclusion and fairness.

The first 9 voluntary guardrails have been aligned closely with proposed mandatory guardrails, with the exception of the 10th voluntary guardrail, which emphasises the importance of ongoing engagement with stakeholders to evaluate their needs and circumstances. Conformity assessments, proposed in the 10th mandatory guardrail, are being prepared for in the Voluntary AI Safety Standard through several voluntary steps organisations can be taking now to improve their record keeping, transparency and testing approaches.

An infographic displaying the 10 Guardrails apply to all organisations and all stages of the AI lifecycle; and the role of developers across the Design, Data, Train, Build and Test steps and Deployers across the Integrate, deploy and monitor stages.

Figure 1: Application of the guardrails

An AI deployer is an individual or organisation that supplies or uses an AI system to provide a product or service. Deployment can be internal to the business or external. When deployment is external it can impact others, such as customers or other people, who are not deployers of the system.

While the first version of the standard applies to both AI deployers and AI developers, it focuses on providing guidance at the organisational and system level for AI deployers. This reflects feedback received while we developed the standard. We heard that deployers, which are the majority of businesses in the Australian ecosystem who are using AI, had the greatest need for guidance on how to adopt best practice. 

Focusing on deployers also supports them to work with developers on the practices needed to support the safe and responsible use of AI across the supply chain. We will include the additional, more complex guidance for AI developers in the next version of the standard. 

To aid deployers of AI systems, the 10 guardrails include procurement guidance. This will ensure AI suppliers and developers are aligning to the guardrails through contractual agreements. 

A human-centred standard

This standard adopts a human-centred approach to AI development and deployment. This is in line with Australia’s AI Ethics Principles and Australia’s commitment to international declarations such as the Bletchley Declaration. A human-centred approach helps make sure technologies are fit‑for‑purpose while serving humans, respecting individual rights and protecting marginalised groups.

In the context of safe and responsible AI system usage, a human-centred approach means:

  • Protecting people. The standard is designed to help leaders and business owners identify, prevent, minimise and remedy a wide range of harms and AI-related risks relevant to their organisation. This is in line with the government’s Interim Response. However, its main purpose is to protect the safety of people and their rights. A human-centred approach to AI upholds Australia’s responsibility to human rights protections. These protections are enshrined in a range of federal and state and territory instruments, the Australian Constitution and the common law.

Read more about the specific characteristics of AI systems that can amplify existing risks and create new harms for people, organisations, groups or society.

  • Upholding diversity, inclusion and fairness. The standard is designed to help organisations ensure AI systems serve all people in Australia, regardless of racial background, gender, age, disability status or other attribute. 
  • Prioritising people through human-centred design. Human-centred design is an approach to technology design, development and deployment that recognises and balances human goals, relationships and social contexts with the capabilities and limitations of technical systems (Gasson 2023). The standard offers practical ways to prioritise the needs of humans in the use of AI systems.
  • Deploying trustworthy AI systems to support social licence. To unlock the greatest possible value from AI, an organisation deploying it must have social licence for its use. This social licence is based on stakeholders believing in the trustworthiness of the AI system. It is only by earning and maintaining the trust of stakeholders that an organisation can be confident it possesses the social licence needed to deploy AI systems. 

Bias

This standard defines bias as the ‘systematic difference in the treatment of certain objects, people or groups in comparison to others’. It can be the basis for unfairness, defined as ‘unjustified differential treatment that preferentially benefits certain groups more than others’.

For some use cases, such as healthcare, accounting for gender differences can be essential to understand the risk factors or treatment appropriate for an individual or group. This justifies a differential treatment (Cirillo et al. 2020). 

Bias becomes problematic or ‘unwanted’ when it results in unfavourable treatment for people or groups. This unfair disadvantage then becomes unlawful discrimination if that treatment is a result of a ‘protected attribute’:

  • age
  • disability
  • race, including colour, national or ethnic origin or immigrant status
  • sex, pregnancy, marital or relationship status, family responsibilities or breastfeeding
  • sexual orientation, gender identity or intersex status. 

An internationally consistent standard

Recognising that Australia is an open, trading economy, the standard’s recommended processes and practices are consistent with current international standards and best practice. This supports Australian organisations who operate internationally by aligning Australian practices with other jurisdictions’ expectations. It also aims to avoid creating barriers to international organisations operating in Australia compared to other markets. 

The standard draws on and is aligned with a range of international standards. Most important is the leading international standard on AI management systems, AS ISO/IEC 42001:2023, and the US standard on AI risk management, NIST AI RMF 1.0. Each requirement in the guardrails is aligned with relevant international and local standards or practices. 

Future versions will reflect changes in the international landscape.