RiskNet: A large-scale dataset of AI risk incidents from news with alignment and multi-dimensional annotations

RiskNet provides the first large-scale, systematically organized database of real-world AI harms and failures, drawing from hundreds of millions of news reports across multiple languages. This addresses a critical gap in responsible AI governance, which has relied on high-level principles without sufficient empirical data on actual incidents occurring in practice.

The dataset groups fragmented news reports into unified incident records, allowing researchers and policymakers to see the full scope of individual AI failures rather than isolated headlines. For example, multiple news stories about the same AI system causing harm are now linked together, providing a more complete picture of recurring problems.

RiskNet uses a multi-dimensional classification system to categorize AI risks, moving beyond simple categorization to capture the complexity of real-world incidents—including what went wrong, which stakeholders were affected, and what domain (healthcare, criminal justice, employment, etc.) was involved. This structured approach enables data-driven analysis of which types of AI risks are most prevalent.

The resource is designed for continuous monitoring and computational analysis, rather than being a static collection. This means researchers can use machine learning tools to automatically identify and track emerging AI risk patterns over time, supporting proactive rather than reactive governance.

RiskNet is publicly accessible through an online platform, democratizing access to incident data for researchers, journalists, policymakers, and the public—enabling broader scrutiny of AI system failures and more informed decision-making about AI deployment and regulation.