Not yet human-reviewed
This source record was written automatically by the research pipeline and has not yet been reviewed by the maintainer or integrated into a topic page. Treat it as a provisional summary and verify claims against the original source.
Accountability Asymmetry and Structural Trust in Autonomous AI Systems
Citation. DeBardeleben, N. (2026). Accountability Asymmetry and Structural Trust in Autonomous AI Systems. arXiv:2608.03670.
Summary. This preprint argues that delegating operational authority to autonomous AI creates accountability asymmetry: consequences for bad decisions fall on deploying people and institutions rather than on the system selecting the action. It frames governance as an infrastructure-reliability problem and proposes engineered heterogeneity, in which the action proposer is not its sole approver or auditor. The paper is conceptual and uses scientific-computing infrastructure to illustrate how authority, reversibility, observability, and review should be designed together.
Key claims
- Alignment and organizational liability can improve deployment behavior but do not reproduce the pre-action deterrent that constrains human operators whose own futures are at stake.
- Governance should scale with an autonomous system’s operational authority, possible harm, uncertainty, and reversibility rather than treating all AI assistance as equivalent.
- Structural trust requires separating proposal, approval, execution, monitoring, and later audit so that the component initiating an action is not the only process able to validate it.
- Long-horizon audit trails are needed to detect gradual authority expansion or declining observability that may not be evident in a single action.
Evidence & limitations
- Conceptual governance paper grounded in illustrative scenarios from scientific-computing infrastructure; it does not report an empirical evaluation of the proposed controls.
- Public provenance verified against arXiv metadata and the public arXiv PDF; the public PDF exactly matches the ingested Drive artifact by SHA-256.
- Preprint; its proposed accountability and engineered-heterogeneity framework requires empirical and comparative validation across deployment settings.