The MIT AI Risk Repository

Artificial intelligence risk research lacked a common language until MIT’s FutureTech initiative addressed the gap directly. Its AI Risk Repository consolidates existing knowledge into one structured, public resource.

Why a Unified Taxonomy Was Needed

Frameworks for classifying AI risk had multiplied without converging. MIT researchers reviewed 65 existing taxonomies and extracted 1,612 unique risks. The AI Risk Repository is the result: a living, publicly accessible database built to expand as new risks emerge.

Two Ways to Classify Risk

MIT structured the repository around two complementary taxonomies.

1. Causal Taxonomy

This taxonomy classifies risk along three dimensions:

• Entity: whether the risk originates from human decisions, AI actions, or an ambiguous source.
• Intentionality: whether the risk is deliberate or accidental.
• Timing: whether the risk arises before deployment or after.

The data shows a clear pattern. 41% of risks originate directly from AI systems, and 62% surface only after deployment. Post-deployment monitoring is not optional. It is where most risk materialises.

2. Domain Taxonomy

The second taxonomy sorts risks into seven domains: discrimination and toxicity, privacy and security, misinformation, malicious use, human-computer interaction, socioeconomic and environmental harm, and AI system safety and failures.

System safety failures, socioeconomic harm, and discrimination are the most extensively documented domains. AI welfare and rights, information ecosystem pollution, and competitive dynamics remain comparatively under-researched. That gap deserves attention when boards set research and monitoring priorities.

Who Should Use It

The repository serves several audiences directly.

• Policymakers use it to draft regulation and build oversight frameworks grounded in evidence rather than assumption.
• Industry and auditors use it to design internal risk management and compliance programmes tailored to their exposure.
• Researchers and educators use it to identify gaps in current studies and structure curricula.
• The public and advocacy groups use it to ground the AI risk debate in a shared vocabulary.

Outlook

MIT designed the repository to evolve. Contributions from researchers worldwide keep it current as new risk categories emerge.

For boards overseeing AI adoption, this repository is a starting reference, not a substitute for company-specific risk assessment. It gives structure to a debate that has, until now, lacked one.

Access the repository at airisk.mit.edu.