MIT’s AI Risk Initiative released version 3 of its AI Risk Mitigation Taxonomy, cataloging 2,206 strategies from 47 technical, policy, and governance frameworks. It matters because managing AI risk requires action across developers, deployers, regulators, organizations, and society—not just better AI systems.
The database classifies each mitigation by its responsible actor, timing, target risk, goal, and implementation lever. Developers account for 41% of strategies, deployers 33%, and governance actors 18%; 55% focus directly on AI systems. The taxonomy also reveals gaps: although experts rank competitive dynamics and power centralization among the five most severe risks by 2030, only 15 entries address them directly.
Because inconsistent descriptions make strategies hard to compare, the team introduced PRIMER, an eight-item reporting checklist covering the mitigation, risk, implementer, timing, and method. The interactive explorer and paper are available through the MIT AI Risk Initiative, and developers and deployers are invited to use PRIMER and provide feedback.
Takeaways
The taxonomy suggests that AI-risk responsibility is distributed across the ecosystem: nearly half of the listed mitigations are assigned to deployers or governance actors rather than developers.
A notable gap is the mismatch between expert concern about competitive dynamics and power centralization and the small number of mitigation records addressing those risks directly.
Article timeline
Build this article’s timeline
Extract dated events, key actors, and open questions using the same timeline analysis available in Signal.
Alexander Saeri, Kun Zhao, Michael Noetel, and MIT AI Risk Initiative
🤖 AI prefill
📰What
MIT’s AI Risk Initiative released version 3 of its AI Risk Mitigation Taxonomy, cataloging 2,206 strategies from 47 technical, policy, and governance frameworks.
🤖 AI prefill
📅When
2030
🤖 AI prefill
📍WhereWhere did this occur?
💡Why
Centers on AI Risk Mitigation Taxonomy, AI risk mitigation strategies, and PRIMER reporting checklist
🤖 AI prefill
Login to answer and earn XP for each gap you fill.
Manipulation Score: LowHow much this article uses emotional framing vs. straight reporting
Loaded Language825%
Absolutist215%
Threat/Panic1025%
Us vs Them015%
Engagement Bait2820%
Manipulation score look right?
Factual Claims (5)
lowtechnology
MIT’s AI Risk Initiative released version 3 of its AI Risk Mitigation Taxonomy, cataloging 2,206 strategies from 47 frameworks.
The MIT AI Risk Initiative just released the third version of our AI Risk Mitigation Taxonomy: 2,206 mitigation strategies for AI risks, extracted from 47 frameworks
mediumtechnology
Managing AI risk requires action across developers, deployers, regulators, organizations, and society, not only improvements to AI systems.
Reducing AI risk extends far beyond improving the technology itself.
lowtechnology
The database classifies each mitigation by responsible actor, timing, target risk, goal, and implementation lever.
We extracted and classified each by who acts, when, against which risk, what it aims to achieve, and the lever used
mediumtechnology
Developers account for 41% of strategies, deployers 33%, and governance actors 18%; 55% focus directly on AI systems.
Developers account for 41% of mitigations, deployers 33%, and governance actors such as regulators 18%.
mediumtechnology
Competitive dynamics and power centralization rank among the five most severe risks by 2030, but only 15 mitigation entries address them directly.
Our Delphi study of 272 experts rated competitive dynamics and power centralization among the five risks most likely to cause severe or catastrophic harm by 2030, but only 15 mitigation records direct
Comments
Login to view and post comments.