Štěpán Los

Purpose of this post: The purpose of this post is three-fold: 1) highlight the importance of incident sharing and share best practices from adjacent fields to AI safety 2) collect tentative and existing ideas of implementing a widely used AI incident database and 3) serve as a comprehensive list of existing AI incident databases as of June 2023. Epistemic status: I have spent around 25+ hours researching this topic and this list is by no means meant to be exhaustive. It should give the reader an idea of relevant adjacent fields where incident databases are common practice and should highlight some of the more widely used AI incident databases which exist to date. Please feel encouraged to comment any relevant ideas or databases that I have missed, I will periodically update the list if I find anything new. Motivation for AI Incident Databases Sharing incidents, near misses and best practices in AI development decreases the likelihood of future malfunctions and large-scale risk. To mitigate risks from AI systems, it is vital to understand the causes and effects of their failures. Many AI governance organizations, including FLI and CSET, recommend creating a detailed database of AI incidents to enable information-sharing between developers, government and the public. Generally, information-sharing between different stakeholders 1) enables quicker identification of security issues and 2) boosts risk-mitigation by helping companies take appropriate actions against vulnerabilities. Best practices from other fields National Transportation Safety Board (NTSB) publishes and maintains a database of aviation accidents, including detailed reports evaluating technological and environmental factors as well as potential human errors causing the incident. The reports include descriptions of the aircraft, how it was operated by the flight crew, environmental conditions, consequences of event, probable cause of accident, etc. The meticulous record-keeping and best-practice