TASRA: A Taxonomy and Analysis of Societal-Scale Risks from AI
A structured accountability-based taxonomy of societal-scale AI risks offers APS risk and governance practitioners a reusable classification lens.
Key points
- TASRA organises AI risks into six types based on accountability, intent, and scale of impact.
- The taxonomy's accountability-based framing aligns with APS responsible AI obligations around oversight and harm prevention.
- This is a 2023 academic preprint summarised in 2024 - useful for conceptual framing rather than immediate policy action.
Summary
Critch and Russell's TASRA framework classifies societal-scale AI risks into six categories - diffusion of responsibility, unexpectedly large impacts, unexpectedly harmful impacts, willful indifference, criminal weaponization, and state weaponization - using a decision tree based on who is accountable, whether actors are unified, and whether harm is deliberate. The MIT AI Risk Repository has summarised it as a reference tool. For APS practitioners, the taxonomy's emphasis on accountability structures and diffuse responsibility is directly relevant to whole-of-government AI governance frameworks, procurement accountability, and risk assessment design.
Implications for Australian agencies
- Consider Risk and governance teams could consider whether TASRA's six-category taxonomy usefully supplements existing agency AI risk registers or departmental AI risk assessment templates.
- Monitor Policy teams developing advice on AI accountability may want to monitor how TASRA's accountability framing is adopted in international governance standards or regulatory guidance.
Implications are AI-generated. Starting points, not advice.
"TASRA: A Taxonomy and Analysis of Societal-Scale Risks from AI" Source: MIT AI Risk Repository – Blog Published: 28 August 2024 URL: https://airisk.mit.edu/blog/tasra-a-taxonomy-and-analysis-of-societal-scale-risks-from-ai Critch and Russell's TASRA framework classifies societal-scale AI risks into six categories - diffusion of responsibility, unexpectedly large impacts, unexpectedly harmful impacts, willful indifference, criminal weaponization, and state weaponization - using a decision tree based on who is accountable, whether actors are unified, and whether harm is deliberate. The MIT AI Risk Repository has summarised it as a reference tool. For APS practitioners, the taxonomy's emphasis on accountability structures and diffuse responsibility is directly relevant to whole-of-government AI governance frameworks, procurement accountability, and risk assessment design. Implications for Australian agencies: - [Consider] Risk and governance teams could consider whether TASRA's six-category taxonomy usefully supplements existing agency AI risk registers or departmental AI risk assessment templates. - [Monitor] Policy teams developing advice on AI accountability may want to monitor how TASRA's accountability framing is adopted in international governance standards or regulatory guidance. Retrieved from SIMS, 18 May 2026.