Guardado en:
| Autor principal: | |
|---|---|
| Formato: | Recurso digital |
| Lenguaje: | inglés |
| Publicado: |
Zenodo
2026
|
| Materias: | |
| Acceso en línea: | https://doi.org/10.5281/zenodo.20112052 |
| Etiquetas: |
Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
|
Tabla de Contenidos:
- <p>Emerging governance frameworks across major jurisdictions agree that human oversight is primary to the safe deployment of Agentic AI, particularly for high-risk systems. The EU AI Act (Article 14), the NIST AI Risk Management Framework, and the African Union's Continental AI Strategy all share a common concern: allowing AI systems to operate autonomously, without human interference to monitor outputs or pause wrong processes, creates real risks to human safety and fundamental human rights. More important is the question of accountability and liability when an AI system delivers a wrong or harmful response. Human oversight creates the opportunity for shared liability and meaningful human accountability.</p> <p>This paper does not dispute these principles; in fact, it agrees with the logic. What it questions is the assumption that the conditions enabling effective human oversight are stable and reliably present. Governance frameworks treat oversight as a design feature, something you build into a system once and can thereafter depend on. In theory, it can be permissible, but in practice, environmental conditions that influence human oversight are rarely stable. Connectivity drops, institutions are under-resourced, and operators are fatigued or undertrained. The human in the loop is only as effective as the conditions surrounding them.</p> <p>The argument is built on four pieces of evidence. First, a deployment incident I led involving an agentic voice system at a career event in an underground venue with an unstable network connection. That incident is a real-world case showing how infrastructural gaps can directly undermine oversight in ways no governance framework anticipates. Second, the paper draws on existing peer-reviewed scholarship on Article 14, automation bias, African AI governance, and the Brussels Effect, which already documents the inadequacies of current frameworks. This paper extends that body of work rather than starting from scratch. Third, it examines published evidence on global infrastructure conditions, particularly within the Global South, and how those conditions shape the real-world deployment of agentic systems. Fourth, and most significantly, the findings of Dagstuhl Seminar 25272, which interrogated where the EU AI Act falls short and identified environmental factors as one of three conditions for effective human oversight, providing the scholarly validation that sits at the centre of this paper's position.</p>