Mapping the AI Governance Landscape: Pilot Test and Update
Reveals structural gaps in global AI governance coverage—including in public administration—that Australian agencies can use to stress-test their own frameworks.
Key points
- MIT AI Risk Repository used LLMs to classify 950+ AI governance documents against risk and mitigation taxonomies.
- Governance failure, AI security vulnerabilities, and lack of transparency are the most-covered risk areas across current frameworks.
- Public administration is the most-covered sector; AI welfare, multi-agent risks, and economic devaluation are least covered.
Summary
MIT's AI Risk Repository has piloted an LLM-assisted pipeline to classify over 950 AI governance documents from CSET's AGORA archive against its AI Risk Taxonomy and a preliminary Mitigation Taxonomy. The study found that Claude Opus 4.1 and GPT-5 achieved agreement with human consensus equal to or exceeding inter-human agreement, validating the LLM-as-classifier approach. Provisional findings show governance failure, AI security, and transparency are the most-covered risk domains, while AI welfare, multi-agent risks, and economic devaluation of human effort remain neglected. The team plans to publish reports, visualisations, and a database under Creative Commons licensing.
Implications for Australian agencies
- Monitor APS AI governance and strategy teams may want to monitor the forthcoming MIT database and visualisations as a benchmarking tool against global coverage patterns.
- Consider Agencies reviewing or developing AI governance frameworks could consider whether their own documents address the identified gap areas—AI welfare, multi-agent risks, and economic devaluation—which are underserved across the global landscape.
Implications are AI-generated. Starting points, not advice.
"Mapping the AI Governance Landscape: Pilot Test and Update" Source: MIT AI Risk Repository – Blog Published: 15 October 2025 URL: https://airisk.mit.edu/blog/mapping-the-ai-governance-landscape-pilot-test-and-update MIT's AI Risk Repository has piloted an LLM-assisted pipeline to classify over 950 AI governance documents from CSET's AGORA archive against its AI Risk Taxonomy and a preliminary Mitigation Taxonomy. The study found that Claude Opus 4.1 and GPT-5 achieved agreement with human consensus equal to or exceeding inter-human agreement, validating the LLM-as-classifier approach. Provisional findings show governance failure, AI security, and transparency are the most-covered risk domains, while AI welfare, multi-agent risks, and economic devaluation of human effort remain neglected. The team plans to publish reports, visualisations, and a database under Creative Commons licensing. Implications for Australian agencies: - [Monitor] APS AI governance and strategy teams may want to monitor the forthcoming MIT database and visualisations as a benchmarking tool against global coverage patterns. - [Consider] Agencies reviewing or developing AI governance frameworks could consider whether their own documents address the identified gap areas—AI welfare, multi-agent risks, and economic devaluation—which are underserved across the global landscape. Retrieved from SIMS, 18 May 2026.