Last Week in AI Sean McGregor on the AI Incident Database and the AI XPRIZE
Jul 6, 2021
In this engaging discussion, Sean McGregor, ML architect at Syntiant and creator of the AI Incident Database, dives into the critical role of incident documentation in improving AI safety. He shares compelling case studies from Google and Tesla, emphasizing harm classification and prevention. McGregor also sheds light on the cultural aspects of AI through the XPRIZE initiative, illustrating its global impact. He balances the vibrant dualities of AI as a tool and a potential threat while sharing personal stories from his pandemic experiences.
AI Snips
Chapters
Transcript
Episode notes
AI Incident Database Learns from Failures
- The AI Incident Database, inspired by aviation incident databases, collects AI failures to improve safety.
- Unlike aviation, AI failures are often unpredictable, requiring real-world examples to understand and address.
Examples of AI Incidents
- The database includes incidents like inappropriate content on YouTube Kids and racial bias in Google Photos.
- It also includes real-world harms, such as the Tesla autopilot fatality involving a driver watching a movie.
Broad Definition of Harm
- The AI Incident Database defines harm broadly, erring on the side of inclusion.
- A new taxonomy classifies incidents by harm type and related entities, aiding analysis.