LLM-Powered AI Agents with Autonomous Executive Capacity and the Limits of Liability Attribution in Continental Private Law

Ramos-Zaga F.A. · Revista Juridica Austral (2026) · DOI: 10.26422/rja.2026.0701.zag

Resumen

Continental private law allocates liability on the basis of the distinction between persons and things. The emergence of large language model (LLM)-powered AI agents capable of autonomously performing actions in real-world digital environments, however, calls into question the adequacy of these traditional categories. This article examines whether classical doctrines of liability attribution remain conceptually adequate in light of the operational characteristics of such agents and evaluates the viability of a framework of limited technical subjectivity for allocating patrimonial losses arising from their conduct. Through a doctrinal and comparative analysis of continental liability regimes, the study finds that traditional forms of indirect liability face structural limitations stemming from the agent’s lack of deliberative capacity, the impracticability of continuous human oversight, and their susceptibility to external instructional manipulation. To address this functional gap, the article proposes a regime of limited technical subjectivity based on a segregated asset pool and a tiered liability structure. The findings suggest that existing liability doctrines are ill-equipped to address harms generated by autonomous AI agents and that the proposed framework provides a coherent mechanism for risk allocation while preserving the foundational architecture of private law.

Instituciones: Universidad Privada del Norte