🔔 The number of AI-related governance documents is rapidly proliferating, but what risks, mitigations, and other concepts do these documents actually cover? 

MIT AI Risk Initiative researchers Simon Mylius, Peter Slattery, Yan Zhu, Alexander Saeri, Jess Graham, Michael Noetel, and Neil Thompson teamed up with CSET’s Mina Narayanan and Adrian Thinnyun to pilot an approach to map over 950 AI governance documents to several extensible taxonomies. These taxonomies cover AI risks and actors, industry sectors targeted, and other AI-related concepts, complementing AGORA’s thematic taxonomy of risk factors, harms, governance strategies, incentives for compliance, and application areas. 

We first compared the performance of five different LLMs against six human reviewers across six documents and ultimately selected Claude Sonnet 4.5 to classify documents at scale. We then used Claude Sonnet 4.5 to map over 950 AGORA documents to several taxonomies. Our preliminary findings include:

• The most covered risk subdomains were Governance failure, AI system security vulnerabilities & attacks, and Lack of transparency or interpretability.
• The least covered risk subdomains were AI Welfare and Rights, Multi-agent risks, and Economic and cultural devaluation of human effort.
• The sectors with the most coverage were Public Administration (excluding National Security), Scientific R&D, and National Security.
• The least covered sectors were Accommodation, Food, and Other Services; Arts, Entertainment, and Recreation; and Real Estate and Rental and Leasing.

Going forward, we will work on providing reports, visualizations, and a database to help users make sense of the AI governance landscape and identify which risks and mitigations are addressed or neglected by AI governance strategies.

For more information about our methodology and findings, make sure to read the Mapping the AI Governance Landscape blog post.