AI Policy Wiki
Dashboard

MIT AI Risk Initiative

high confidence · updated 2026-08-14

MIT-housed research initiative cataloguing AI risks and governance documents under a shared taxonomy. Operates the AI Risk Repository (publicly maintained risk database) and, as of April 29, 2026, the AI Risk Navigator — a centralizing front-end that lets researchers and policy actors pull risks, incidents, and governance documents for any domain in one place.

The MIT AI Risk Initiative is a research initiative at MIT that systematizes the study of AI risks under a shared taxonomy. It operates the AI Risk Repository, a publicly maintained database that aggregates risks identified across academic literature and policy frameworks, and, as of April 29, 2026, the AI Risk Navigator, a front-end across the initiative's risk datasets.

Activities

The AI Risk Repository aggregates risks identified across academic literature and policy frameworks into a publicly maintained database.

On April 29, 2026, the initiative launched the AI Risk Navigator, which centralizes its risk datasets under a shared taxonomy so that researchers and policy actors can pull risks, incidents, and governance documents for any domain in one place (Source: airisk.mit.edu). The Navigator is positioned as an interface to the initiative's existing data products, aggregating the AI risk research literature by domain.

The Navigator is one of several cross-source AI-risk catalogs that researchers and policymakers reference by name, alongside the OECD's AIM (AI Incidents Monitor) and the AI Incident Database (AIID); relative to those, the initiative emphasizes taxonomy and governance-document coverage. It serves a different function from the FLI AI Safety Index Winter 2025 scoring framework: where the FLI index scores labs against safety practices, the Navigator catalogs the risks themselves and tracks governance responses across domains.

The Navigator launch on April 29, 2026 coincided with a separate U.K. AISI report exploring whether AI models could sabotage AI safety research, covering the question of how the AI safety field studies itself.

Expert elicitation on 24 risk domains

A three-round Delphi study conducted with the University of Queensland School of Psychology between September and November 2025 asked 272 experts from 37 countries to rate the 24 subdomains of the initiative's taxonomy on harm severity and probability, actor and sector vulnerability, and responsibility for mitigation (Prioritization of Risks from Artificial Intelligence: A Delphi Study of 272 International Experts). The paper is dated June 2026; it reached wider attention through an MIT Sloan press release of August 13, 2026 (Source: mitsloan.mit.edu).

Under a business-as-usual scenario the panel assigned at least a 10% probability of catastrophic outcomes over 2025–2030 to 18 of the 24 risks, with catastrophic anchored at more than one million deaths, more than USD $100B in financial loss, or civilization-scale intangible harms. The five highest mean severity ratings went to dangerous capabilities (3.49/5, 21.5% catastrophic probability), competitive dynamics (3.49, 16.6%), weapons and cyberattacks (3.49, 21.0%), power centralization (3.47, 18.0%), and false information (3.44, 12.8%). A separate list applies to the mitigations scenario: severity fell for all 24 risks, and five remained above 10% catastrophic probability — dangerous capabilities and weapons and cyberattacks at 12% each, environmental harm at 12%, inequality and unemployment at 11%, and power centralization at 11% — while all 24 stayed above 5%.

The study's structural finding is a separation between vulnerability and responsibility. AI users and affected stakeholders drew median vulnerability ratings of 4–5 across nearly all risks but median responsibility ratings of only 2–3, while general-purpose AI developers and governance actors drew responsibility medians of 4–5. Infrastructure providers were rated least vulnerable at a median of 2. Information and national security were rated the most vulnerable sectors, with finance and insurance close behind. The paper states that its panelists are risk specialists rather than calibrated forecasters, that self-nominating respondents may be more concerned than the wider expert population, and that the panel was 68% male and 79% North American or European.

Distinct from MIT NANDA

The MIT AI Risk Initiative is structurally distinct from MIT NANDA, the Networked Agents and Decentralized AI initiative responsible for *The GenAI Divide* (2025). Both are MIT-housed and share academic personnel. NANDA focuses on agentic-AI economic adoption; the AI Risk Initiative focuses on AI risk taxonomy and governance.

Relationships

Sources