Accountability Mechanisms in AI Projects: An Empirical Analysis of Governance, Performance Control, Delivery, and Misconduct Management

Abstract

Artificial intelligence (AI) projects create significant opportunities but also raise complex accountability challenges due to system autonomy, opacity, and multi-stakeholder involvement. While accountability is widely recognized in AI governance, limited research has examined how it is operationalized within project environments. This study addresses this gap by identifying and empirically examining accountability mechanisms in AI projects. Drawing on project governance and AI governance literature, accountability is conceptualized as a system of interrelated mechanisms embedded within project organizations. Survey data from AI project practitioners are used to examine how accountability mechanisms cluster and how their use is influenced by system autonomy, project experience, and organizational role distribution. The results identify 18 accountability mechanisms, operationalized through 23 measurement items, and structured into four accountability dimensions: directing the project, controlling performance, managing delivery, and treating misconduct. A key contribution is the identification of treating misconduct as a distinct accountability dimension that is not represented in major project governance standards and extends accountability beyond project execution to the investigation, remediation, and management of harmful outcomes. The findings further show that system autonomy selectively increases the use of delivery and misconduct-related mechanisms, while foundational governance mechanisms remain stable. Experience is positively associated with all accountability dimensions, indicating that accountability mechanisms become more prevalent as organizations accumulate AI project experience. The study contributes an empirically grounded framework of accountability mechanisms in AI projects and demonstrates that effective accountability requires mechanisms not only for governance and control but also for addressing the downstream consequences of AI system behavior.