Accountability Arbitrage: Ethical Tensions in AI Agent Accountability Infrastructure

A growing class of technical systems, here termed accountability infrastructure, addresses AI agents' lack of standardized recording by capturing actions in tamper-evident chains for independent verification. This paper argues that such infrastructure provides the informational substrate of accountability but not its institutional structure: it produces evidence without a court. The paper's central contribution is accountability arbitrage: a mechanism by which auditable AI agents become preferable to human workers on compliance and documentation grounds, independent of task performance. The concept is developed as a three-part typology (cost-driven, liability-driven, reputation-driven) with falsifiability conditions. The inverse dynamic, accountability avoidance, produces a selection effect where organizations monitor routine tasks but avoid auditable AI for consequential decisions, creating accountability theater. The paper examines adjacent tensions (surveillance, equity, data protection, labor) and concludes that technical deployment without concurrent institutional governance is insufficient.