
Este preprint presenta el marco de Control Supervisado Externo No Invertible (CSENI), una propuesta teórica y arquitectónica orientada a examinar las limitaciones estructurales de los enfoques de seguridad en inteligencia artificial basados exclusivamente en alineación interna. El trabajo argumenta que, bajo condiciones de opacidad del modelo, auditabilidad limitada y capacidad estratégica del sistema supervisado, la seguridad no puede descansar únicamente en restricciones embebidas dentro del mismo sustrato computacional del agente. CSENI se formula como una capa de supervisión externa basada en tres elementos centrales: desacoplamiento del controlador, fricción operativa gradual y uso de observables heterogéneos. El texto incluye un modelo de amenaza preliminar, una formalización mínima del supervisor, una agenda experimental inicial y un apéndice técnico con una instanciación mínima reproducible para agentes con herramientas. Este trabajo no pretende ofrecer una solución definitiva al problema del control de sistemas avanzados de IA. Su propósito es fundacional: proponer un lenguaje conceptual, una base formal mínima y un programa de trabajo posterior para desarrollos técnicos, regulatorios e institucionales más robustos.
This preprint introduces the Non-Invertible External Supervisory Control (NIESC/CSENI) framework, a theoretical and architectural proposal aimed at examining the structural limitations of AI safety approaches based exclusively on internal alignment. The paper argues that, under conditions of model opacity, limited auditability, and strategic capability of the supervised system, safety cannot rely solely on restrictions embedded within the same computational substrate as the agent. CSENI is formulated as an external supervisory layer centered on three elements: controller decoupling, gradual operational friction, and heterogeneous observables. The paper includes a preliminary threat model, a minimal formalization of the supervisor, an initial experimental agenda, and a technical appendix with a reproducible minimal instantiation for tool-using agents. This work does not claim to offer a definitive solution to the control problem for advanced AI systems. Its purpose is foundational: to propose a conceptual vocabulary, a minimal formal basis, and a forward-looking program for future technical, regulatory, and institutional developments.
Goodhart's law, AI containment, artificial intelligence safety, supervisory control, seguridad en inteligencia artificial, AI regulation, supervisión externa, contención de IA, external oversight, specification gaming, operational risk, fail-closed systems, AI governance
Goodhart's law, AI containment, artificial intelligence safety, supervisory control, seguridad en inteligencia artificial, AI regulation, supervisión externa, contención de IA, external oversight, specification gaming, operational risk, fail-closed systems, AI governance
| selected citations These citations are derived from selected sources. This is an alternative to the "Influence" indicator, which also reflects the overall/total impact of an article in the research community at large, based on the underlying citation network (diachronically). | 0 | |
| popularity This indicator reflects the "current" impact/attention (the "hype") of an article in the research community at large, based on the underlying citation network. | Average | |
| influence This indicator reflects the overall/total impact of an article in the research community at large, based on the underlying citation network (diachronically). | Average | |
| impulse This indicator reflects the initial momentum of an article directly after its publication, based on the underlying citation network. | Average |
