Examining the differential risk from high-level artificial intelligence and the question of control
A structured AI risk taxonomy backed by expert survey data - relevant context for agencies building or stress-testing their own AI risk frameworks.
Key points
- MIT AI Risk Repository summarises a four-class risk taxonomy covering misuse, accident, structural, and agential AI risks.
- Expert survey data identifies monopolistic race dynamics, alignment failures, and power-seeking as highest-impact risks.
- This is a 2023 academic paper summary - useful background context but not new guidance for APS practitioners.
Summary
The MIT AI Risk Repository has published a summary of a 2023 academic paper by Kilian, Ventura, and Bailey that proposes a hierarchical risk taxonomy for advanced AI systems. The framework organises risks into four classes - misuse, accident, structural, and agential - and uses expert survey data to rank impact and likelihood. Monopolistic race dynamics, AI alignment failures, and power-seeking behaviour are identified as highest-impact risks. The paper also presents exploratory modelling of how social and technological change influences risk trajectories. While not directed at government specifically, the taxonomy offers a structured lens that may be useful for APS agencies conducting AI risk assessments or reviewing risk register design.
Implications for Australian agencies
- Consider Agencies developing or reviewing AI risk frameworks could assess whether this four-class taxonomy adds value to existing APS risk categorisation approaches.
- Monitor Policy teams tracking AI safety research may want to monitor the MIT AI Risk Repository as a consolidation point for peer-reviewed AI risk frameworks.
Implications are AI-generated. Starting points, not advice.
"Examining the differential risk from high-level artificial intelligence and the question of control" Source: MIT AI Risk Repository – Blog Published: 1 December 2024 URL: https://airisk.mit.edu/blog/examining-the-differential-risk-from-high-level-artificial-intelligence-and-the-question-of-control The MIT AI Risk Repository has published a summary of a 2023 academic paper by Kilian, Ventura, and Bailey that proposes a hierarchical risk taxonomy for advanced AI systems. The framework organises risks into four classes - misuse, accident, structural, and agential - and uses expert survey data to rank impact and likelihood. Monopolistic race dynamics, AI alignment failures, and power-seeking behaviour are identified as highest-impact risks. The paper also presents exploratory modelling of how social and technological change influences risk trajectories. While not directed at government specifically, the taxonomy offers a structured lens that may be useful for APS agencies conducting AI risk assessments or reviewing risk register design. Implications for Australian agencies: - [Consider] Agencies developing or reviewing AI risk frameworks could assess whether this four-class taxonomy adds value to existing APS risk categorisation approaches. - [Monitor] Policy teams tracking AI safety research may want to monitor the MIT AI Risk Repository as a consolidation point for peer-reviewed AI risk frameworks. Retrieved from SIMS, 18 May 2026.