A Taxonomy of Critical AI System Characteristics for Use in Proxy System Testing

Abstract

Safety and trust are two of the most important features in a critical system. A critical system is one that must be highly reliable in that it not only completes its mission but causes zero harm to the public. The problem is testing a critical system, especially if it employs artificial intelligence (AI). The challenge is critical AI systems (CAIS) may cause unpredictable events and conditions that cannot be modeled during critical error testing. Proxy systems (non-critical prototype) are needed to test the critical system. We present a five-dimensional CAIS taxonomy and a weighting system to map system characteristics to a testing proxy in order to determine equivalent proxy systems to build and test. Ultimately this CAIS taxonomy and weighting system is a way forward to develop a set of proxy systems to use for critical error testing.

Ähnliche Arbeiten