The industry problem / opportunity
The ecosystem of governance guidance in relation to AI remains difficult to navigate and translate, particularly in healthcare. When organisations do pursue good governance practice, they struggle to describe an AI system and its deployment context. The aim of this project was to understand whether a standardised approach to classifying AI systems and deployments in the Australian healthcare sector could provide value. The Governments expressed interest in a tool that offered potential for dealing with AI solutions in health care, given that many such solutions were being classed as high-risk software as a medical device.
The starting point was localisation of the existing 2022 OECD AI Classification Framework (AICF) into the Australian health context and development of an easy to use and intuitive webbased tool which implements the AICF. The hypothesis was that a web-based tool could initially be used by researchers and a small group of stakeholders working within the health sector in Australia to evaluate its suitability to the local context for differentiating between different AI solutions.
The solution and outcomes
The localisation of the OECD AI Classification Framework for the Australian health sector project employed three phases of activity:
- Assessing and enhancing the alignment of the AICF with Australian AI standards and principles;
- Localisation of the framework to the health sector; and
- Moving the static AICF instrument into an interactive web-based tool to enable uptake and use of the classification system.
Because the OECD AICF predated development of the Australian AI guidelines including Australia’s Voluntary AI Safety Standards, it was important to ground the AICF to the Australian context. The project considered the breadth of AI regulatory standards and selected a subset of key reference materials to guide the alignment of the AICF to the Australian context.
The mappings between the AICF and the respective Australian guidelines informed revisions to the questions in the AICF. These were then revised using feedback from small focus group participants (largely developers and deployers of AI systems) and fed into the prototype tool in three ways. Firstly, by providing contextual information alongside the AICF questions to explain why the question is important and relevant in Australia; secondly, by providing dynamic classification based on user responses to the modified AICF question set; and, thirdly, by producing a dynamic final report providing contextual information for each relevant potential impact area and for each relevant action area identified in the classification process.
The impact
This short, 8-month project has demonstrated there is value in being able to classify different AI solutions for use in healthcare in Australia as an enabler of good governance. Further work is required to validate the contextual mapping and revisions made to the original OECD AICF questions. The prototype web-based tool provides a practical way in which this can be done.
The insight
Professor Adam Berry, project lead and Deputy Director, Human Technology Institute, UTS said, “The project established a foundation for future phases, with clear directions for further user testing, question refinement, and potential expansion into governance or inventory support tools.”
What’s next
The project has stimulated interest in the localisation of the AICF specifically to support the development of organisational AI inventories, public AI inventories, or both, in the Australian health context. Registers of AI systems have been required in other jurisdictions – such as the UK, where the Government Digital Service has developed an algorithmic transparency recording standard, and a public register. The concept of an organisational AI inventory also has support within the Australian Voluntary AI Safety Standard.
More generally, the widespread uptake and use of standardised inventory frameworks could support good governance, potentially enable improved transparency and support clearer communication about the types of AI being adopted, where it is being used, and how it is designed and operated. Two key questions, then, emerge:
- whether a localised form of the AICF would receive sufficient uptake to become (at least the de facto) standard for Australian health AI inventories, and,
- if so, whether organisations would be willing to embrace something as comprehensive as the AICF to (potentially broadly) communicate about AI systems they are developing and/or deploying.
The DHCRC is pursuing both these options.
