AI Forensics

Abstract

Artificial intelligence is now a daily topic of public discussion. Not only are intelligent systems such as autonomous vehicles taking the streets, but recommendation algorithms are shaping human behavior. While AI offers a significant increase in efficiency, it can also indirectly cause harm to humans by replacing jobs, violating privacy and even threatening autonomy. To keep track of these cases in which AI has negative implications on a human, databases of AI incidents have been created. We extend this idea of AI incidents to not only include times whereby an AI system caused a real-world harm, but also when it introduces a benefit. Prior work in adjacent fields has defined taxonomies and standard procedures for root cause analysis, digital forensics, AI risk management, and more. Despite these frameworks, there is no means to investigate an AI system to discover the root cause of an incident. We aim to evaluate the body of knowledge that leads to introducing the field of AI Forensics. AI forensics can serve as a postmortem analysis of AI incidents to discover the primary harm catalyst.