Accountability arbitrage: ethical tensions in AI agent accountability infrastructure
TL;DR
It is predicted that expected record advantage will increase agent selection near evidence thresholds, while anticipated exposure will reduce recorded deployment in high-discretion tasks, and governance implications include human–AI evidentiary parity, substantive-performance floors, review authority, and proportionate data collection.
Abstract
AI agents leave operational logs, tool traces, and platform receipts. Newer systems connect these fragments, expose later changes, and support independent verification. As reviewability improves, recordability becomes an allocation factor, not only an engineering property. This article develops that claim. Human and AI-agent workflows may perform the same task while producing records of different institutional value. A regulator, auditor, insurer, court, or internal review body may find one history easier to review. Organizations anticipate that difference when designing workflows and assigning tasks. Formal acceptance can therefore influence allocation before a common technical or legal standard forms. The strict selection effect is accountability arbitrage: an expected record advantage reverses a work choice that performance and ordinary cost would otherwise make. The same infrastructure also creates a dual effect. A durable record may help prove compliance and make error or misconduct easier to establish. Selective movement away from a recorded workflow in response to that exposure is accountability avoidance. The analysis connects operational recording to actor–forum accountability. A record preserves material for judgment; an empowered forum gives that material institutional force. Until that handoff, the record is evidence without a court. A minimal model identifies the strict arbitrage region and threshold discontinuity; a brokerage example holds performance constant and shows record value reversing the choice. We predict that expected record advantage will increase agent selection near evidence thresholds, while anticipated exposure will reduce recorded deployment in high-discretion tasks. Existing recordkeeping practice and enforcement provide the basis for these predictions. Matched studies finding neither response under the stated contrasts would count against the theory. Governance implications include human–AI evidentiary parity, substantive-performance floors, review authority, and proportionate data collection.