What’s Covered?
This document presents a 29-criteria framework designed to capture critical information about AI-related incidents across diverse settings. The goal: make incident data more consistent, comparable, and useful across jurisdictions, sectors, and regulatory models.
The report defines an “AI incident” broadly—covering both materialised harms and near misses—and argues that current patchwork practices in incident reporting are insufficient for managing emerging system-level risks. The framework was developed by analyzing and synthesizing insights from:
- The OECD AI System Classification Framework
- The AI Incident Database (AIID)
- The OECD Global Product Recalls Portal
- The AI Incident Monitor (AIM)
From these sources, the authors identified 88 relevant reporting elements and distilled them down to 29 core criteria, organized across eight dimensions (e.g., system context, impact type, system-user interaction, and incident traceability). Seven of these criteria are mandatory to support minimum viable reporting.
Importantly, the report doesn’t attempt to enforce how countries respond to incidents, only how they structure and share information about them. It advocates for voluntary uptake and piloting through AIM, a platform open to public submissions via oecd.ai/incidents.
💡 Why it matters?
Without a shared reporting standard, global AI governance efforts are flying blind. This framework introduces a pragmatic way to pool knowledge about real-world harms without waiting for full regulatory alignment. If adopted, it would strengthen transparency, help identify systemic risks (like discriminatory patterns across models), and create early-warning signals that could inform audits, enforcement, or model retraining. It’s also a vital step toward building international AI safety infrastructure.
What’s Missing?
The framework’s strength—its flexibility—also leaves room for ambiguity. There’s limited guidance on how criteria should be interpreted or weighed, especially in grey areas like algorithmic discrimination or user error. The framework also doesn’t yet cover enforcement mechanisms or incentives for reporting, which could hinder uptake in competitive or high-liability contexts. More clarity on data governance (e.g., protecting whistleblowers or managing incident reports involving trade secrets) will be needed to make this usable at scale.
Best For:
Policymakers, regulators, standards bodies, and researchers focused on AI risk oversight and incident response. Particularly useful for teams designing national AI registries, compliance tools under frameworks like the EU AI Act, or public transparency mechanisms. Also relevant for developers and AI safety leads looking to align internal practices with emerging norms.
Source Details:
OECD (2025). Towards a Common Reporting Framework for AI Incidents. OECD.AI Papers No. 34.
This paper was developed by the OECD Working Party on AI Governance (AIGO) and the OECD.AI Expert Group on AI Incidents, with support from GPAI and input from stakeholders across civil society, industry, and academia. Lead authors include Karine Perset, Luis Aranda, and Bénédicte Rispal. The framework will be tested through open submissions to the OECD AI Incidents Monitor (AIM), offering a live implementation path for governments and institutions interested in contributing to global AI incident transparency.
