Enacting Responsible AI: A Configurational Analysis of AI Principles in Practice
通过对欧美135位AI专家的调查和访谈，采用模糊集定性比较分析，识别出五种实现高负责任AI实践的原则组合，强调问责和安全可靠原则的关键作用及情境依赖性。
Abstract Responsible AI (RAI) entails developing, using, and governing AI in a human-centred way to ensure its trustworthiness and alignment with human values. Organizations attempt to enact RAI through guiding principles that aim to minimize threats such as bias and privacy violations and enhance outcomes such as transparency and fairness. In the empirical study we report in this paper, we share insights about how AI experts view RAI principles, including relationships among them, these principles, and the enactment of RAI principles. Employing a sequential mixed-methods approach, we analyse survey data from 135 AI experts from Europe and America using fuzzy-set Qualitative Comparative Analysis (fsQCA), complemented by follow-up interviews. Our analysis identifies five equifinal combinations of RAI principles associated with high overall RAI enactment. The findings reveal the crucial role of accountability and safety-reliability principles and highlight the contextual variability of other principles. This research contributes a nuanced understanding of RAI enactment, moving beyond individual principles to demonstrate their complex interplay and their contextual dependencies in achieving responsible AI practices.