An Infrastructure for Managing AI Bias in Complex Web Systems

Abstract

AI components are increasingly integrated into complex web systems, especially in the context of AI-as-a-Service. However, trustworthiness of the resulting architectures is jeopardized by AI biases in the models used, the data they were trained on, or the way they are deployed. This problem needs to be addressed by systematic approaches of managing the AI bias-related risks. We propose AIBDB, an infrastructure inspired by the Common Vulnerabilities and Exposures (CVE) system that serves as a knowledge base about known AI biases and supports software architects in automatically identifying architectural risks of AI bias. With the first proof-of-concept implementation of this web-based AI bias management infrastructure, we conducted a qualitative user study yielding feedback from 7 participants through thematic analysis of their 112 responses on the usability, usefulness and suggestions for improvement. We contribute to the development of future fair and unbiased AI-driven solutions in the web in the context of the ongoing research on Trustworthy and Human-centric AI and the increasing importance for organizations to address AI bias through recent regulations.

Ähnliche Arbeiten