Towards Risk-Aware Artificial Intelligence and Machine Learning Systems: An Overview
A structured taxonomy of AI/ML prediction risks offers APS risk and assurance teams a reference for categorising failure modes in high-stakes deployments.
Key points
- A 2022 academic framework categorises AI/ML risks into data-level and model-level risk types.
- The framework targets high-stakes decision settings like healthcare and transport - directly relevant to APS use cases.
- This is a spotlight of a three-year-old paper, not new guidance; practical application requires further translation work.
Summary
The MIT AI Risk Repository has spotlighted a 2022 academic paper by Zhang et al. that systematically categorises AI/ML risks into data-level risks (bias, dataset shift, out-of-domain data, adversarial attack) and model-level risks (model bias, misspecification, prediction uncertainty). The framework emphasises high-stakes decision contexts and draws on reliability engineering concepts to support risk-aware AI development. While conceptually useful, this is a summary of existing academic work rather than new guidance, and would require adaptation to be directly applicable to APS procurement, assurance, or governance processes.
Implications for Australian agencies
- Consider APS risk and assurance teams could consider whether Zhang et al.'s data-level and model-level taxonomy usefully supplements existing agency AI risk registers or assessment frameworks.
- Monitor Teams engaging with the MIT AI Risk Repository may want to monitor the full collection of spotlighted frameworks for patterns that inform whole-of-government risk guidance.
Implications are AI-generated. Starting points, not advice.
"Towards Risk-Aware Artificial Intelligence and Machine Learning Systems: An Overview" Source: MIT AI Risk Repository – Blog Published: 19 December 2025 URL: https://airisk.mit.edu/blog/towards-risk-aware-artificial-intelligence-and-machine-learning-systems-an-overview The MIT AI Risk Repository has spotlighted a 2022 academic paper by Zhang et al. that systematically categorises AI/ML risks into data-level risks (bias, dataset shift, out-of-domain data, adversarial attack) and model-level risks (model bias, misspecification, prediction uncertainty). The framework emphasises high-stakes decision contexts and draws on reliability engineering concepts to support risk-aware AI development. While conceptually useful, this is a summary of existing academic work rather than new guidance, and would require adaptation to be directly applicable to APS procurement, assurance, or governance processes. Implications for Australian agencies: - [Consider] APS risk and assurance teams could consider whether Zhang et al.'s data-level and model-level taxonomy usefully supplements existing agency AI risk registers or assessment frameworks. - [Monitor] Teams engaging with the MIT AI Risk Repository may want to monitor the full collection of spotlighted frameworks for patterns that inform whole-of-government risk guidance. Retrieved from SIMS, 18 May 2026.